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.1007135 | Innate immune sensor LGP2 is cleaved by the Leader protease of foot-and-mouth disease virus | The RNA helicase LGP2 (Laboratory of Genetics and Physiology 2) is a non-signaling member of the retinoic acid-inducible gene-I (RIG-I)-like receptors (RLRs), whose pivotal role on innate immune responses against RNA viruses is being increasingly uncovered. LGP2 is known to work in synergy with melanoma differentiation-associated gene 5 (MDA5) to promote the antiviral response induced by picornavirus infection. Here, we describe the activity of the foot-and-mouth disease virus (FMDV) Leader protease (Lpro) targeting LGP2 for cleavage. When LGP2 and Lpro were co-expressed, cleavage products were observed in an Lpro dose-dependent manner while co-expression with a catalytically inactive Lpro mutant had no effect on LGP2 levels or pattern. We further show that Lpro localizes and immunoprecipitates with LGP2 in transfected cells supporting their interaction within the cytoplasm. Evidence of LGP2 proteolysis was also detected during FMDV infection. Moreover, the inhibitory effect of LGP2 overexpression on FMDV growth observed was reverted when Lpro was co-expressed, concomitant with lower levels of IFN-β mRNA and antiviral activity in those cells. The Lpro target site in LGP2 was identified as an RGRAR sequence in a conserved helicase motif whose replacement to EGEAE abrogated LGP2 cleavage by Lpro. Taken together, these data suggest that LGP2 cleavage by the Leader protease of aphthoviruses may represent a novel antagonistic mechanism for immune evasion.
| Foot-and-mouth disease virus (FMDV) is the causative agent of a devastating disease affecting livestock worldwide. FMDV is considered an extremely successful pathogen able to replicate and spread rapidly among its hosts. The induction of type-I interferon (IFN) response is a crucial event in mammalian cells against infections, and viruses have evolved a variety of mechanisms to avoid it. LGP2 (Laboratory of Genetics and Physiology 2) is a cellular protein involved in sensing viral RNA during infection and plays a relevant role on regulation of the signaling pathways leading to IFN induction. Here we show that LGP2 is specifically targeted for cleavage by the FMDV-encoded Leader protease (Lpro) and a correlation with a decrease in the IFN levels induced in infected cells. Our data unveil a new viral mechanism of immune evasion based on direct cleavage of LGP2 by the Leader proteases of aphthoviruses. To our knowledge this is the first report of the proteolytic cleavage of LGP2 by a virally-encoded protease. Our findings will contribute to a better understanding of the virus-host interplay involved in pathogenesis and to the development of more efficient strategies to combat infectious diseases.
| Antiviral response against RNA viruses greatly relies on detection of infection by cytoplasmic sensors. Among the different pattern-recognition receptors (PRRs) involved in antiviral immunity, the retinoic acid-inducible gene-I (RIG-I)-like receptors (RLRs), recognize non-self RNA species derived from viral infection triggering the downstream signaling cascade leading to type-I interferon (IFN) secretion and host antiviral response [1,2]. RLRs are ubiquitous cytosolic RNA helicases including RIG-I, melanoma differentiation-associated gene 5 (MDA5) and LGP2 (Laboratory of Genetics and Physiology 2). All three RLRs share a DExD/H box RNA helicase domain and a C-terminal domain (CTD). The helicase domain and the CTD bind to viral RNA, CTD being essential for the specific recognition of RNA substrate features. The helicase domain generally functions to coordinate RNA binding, ATP hydrolysis, and conformational rearrangements upon RNA recognition [2,3]. The RLRs share the ability to detect molecular signatures of virus infection, but differ in both their RNA recognition specificity and signaling properties. RIG-I senses primarily 5´-triphosphate blunt-end dsRNA, while MDA5 is activated by long dsRNA, consequently responding to different but overlapping sets of viruses [4,5]. LGP2 has the highest RNA binding affinity of the RLRs, and has the ability to recognize diverse dsRNAs, regardless of the presence of 5´-triphosphate or RNA length [6]. RIG-I and MDA5 contain N-terminal tandem caspase activation and recruitment domains (CARDs) which upon recognition of viral RNA, interact with the CARD of the mitochondrial activator of virus signaling (MAVS) protein, the essential adaptor molecule for RLR signaling. LGP2 lacks the N-terminal CARDs and then independent signaling activity. However, LGP2 is known to be widely involved in viral RNA recognition and regulation during innate immune responses, remaining the most enigmatic member of the RLR family [7]. Both negative and positive regulatory roles have been reported for LGP2 in antiviral immunity. An enhancing effect on MDA5-mediated signaling was found when LGP2 was present at low cellular concentrations. According to a model based on a concentration dependent biphasic switch, at early stages of infection low levels of LGP2 would enhance MDA5-mediated antiviral signaling, but as infection progresses and LGP2 production is induced by IFN, LGP2 would act as a negative feedback regulator inhibiting MDA5 signaling [7–9]. Single molecule RNA binding experiments and biochemical analysis revealed that ATP hydrolysis activity is required to enable LGP2 to efficiently engage diverse dsRNA species, and for enhancement of MDA5 signaling [8]. An RNA- and virus-independent inhibitory role for LGP2 in antiviral signaling has also been reported, likely involving CARD-independent interaction with MAVS by competition with an essential kinase for binding and interfering with downstream signaling [10].
Foot-and-mouth disease virus (FMDV) is the etiologic agent of a highly infectious vesicular disease affecting swine, cattle and other domestic and wild cloven-hoofed animals worldwide [11,12]. FMDV is included in the Aphthovirus genus of the Picornaviridae family. Picornaviruses are small non-enveloped viruses and their capsids enclose a single-stranded RNA genome coding for a polyprotein which is subsequently cleaved by viral proteases to yield the different viral proteins. MDA5 is involved in recognizing the dsRNA synthesized during picornavirus replication [13] and experimental evidence using lentivirus-driven RNA interference supports that FMDV is sensed by MDA5 in porcine epithelial cells [14].
During co-evolution with their hosts viruses have acquired strategies to actively counteract host antiviral responses [15–17] and the balance between innate response and viral antagonism may determine the outcome of disease and pathogenesis. A repertoire of mechanisms aimed at confronting the host IFN response has been described for FMDV, most of them involving the proteolytic activity of the two virally encoded Leader and 3C proteases [18,19].
The FMDV Leader protease (Lpro) is the first protein encoded in the ORF, a papain-like cysteine protease which is present as two different forms, Lab and Lb, generated by translation initiation at two in-frame AUG codons separated 84 nt on the viral RNA [20] and subsequent intramolecular self-processing. Both forms of Lpro are active but Lbpro is more efficiently translated and abundant in infected cells [21]. FMDV Lpro impairs cap-dependent translation through cleavage of initiation factor eIF4G, leading to a translational host shut-off [22,23] and plays an important role in viral pathogenesis. Several cellular proteins have been identified as Lpro targets [24] and Lpro activity is known to disrupt signaling pathways involved in host defenses, like degradation of the p65 subunit of NF-κB and suppression of IFN-β and inflammatory chemokines by reduction of IRF-3/7 expression [25,26]. The deubiquitinase activity of Lpro is also known to cleave ubiquitin moieties from critical signaling proteins of the type-I IFN signaling pathway, such as RIG-I, TBK1, TRAF3, and TRAF6 [27].
A few reports have identified LGP2 as a potential target for viral antagonism. The paramyxovirus V protein binds to the helicase domains of both MDA5 and LGP2 disrupting their enzymatic activity [28]. A recent work describes the interaction between the Nonstructural Protein 3 (NS3) encoded by the hepatitis C virus (HCV) and the helicase domain of LGP2 by quantitative micro-spectroscopic imaging (Q-MSI) [29]. Overexpression of LGP2 has been shown to reduce FMDV growth and interaction of LGP2 with non-structural protein 2B has been detected by immunoprecipitation experiments [30].
Here, we show that FMDV Lpro targets LGP2 helicase for cleavage, resulting in lower levels of IFN-β and antiviral activity in co-transfected cells. No evidence of proteolysis could be detected with a catalytically inactive version of Lpro. The Lpro target sequence in LGP2 was identified as an RGRAR motif which is part of the conserved helicase motif VI of LGP2. Direct interaction between both proteins was evidenced by immunoprecipitation and co-localization assays. LGP2 processing was also detected during FMDV infection, suggesting that LGP2 cleavage by the Leader protease may be a mechanism developed by aphthoviruses to counteract the host immune response. This is the first report of LGP2 proteolytic cleavage exerted by a viral protease and unveils a novel role for the FMDV leader protease on immune evasion.
The FMDV-encoded Leader protease is an important virulence factor involved in IFN antagonism. Given that LGP2 is an innate immunity effector with synergistic effect on MDA5-induced antiviral response, we sought to determine whether FMDV is targeting LGP2 by a mechanism involving the activity of Lpro. First, we studied the effect of the co-expression on HEK293 cells of (Myc-DDK-tagged)-human LGP2 together with either the wildtype catalytically active form of Lbpro (LbWT) or LbC51A, an inactive form of the protease carrying a mutation in the active site [31,32]. The levels and integrity of LGP2 were analyzed 24 h later by immunoblot using antibodies against the N- or C-terminal regions of human LGP2, and compared to those observed after co-transfecting with the empty vector (EV) (Fig 1A). Expression of LbWT induced a drastic decrease in the full-length LGP2 levels. Interestingly, two LGP2-derived products of approximately 49 KDa and 27 KDa were specifically detected with the antibodies against the N- and C-terminal regions of LGP2, respectively. In contrast, when LGP2 and LbC51A were expressed together, no decrease in the helicase levels or additional bands were observed, suggesting that the LGP2 fragments detected may result from specific proteolytic cleavage by Lbpro.
Cleavage of eIF4G, a known cellular target protein for Lbpro was analyzed in HEK293 cells expressing the protease as a control of its catalytic activity in the experimental conditions used (Fig 1B). The 110 KDa cleavage fragment of eIF4G [33] was readily detected in cells expressing LbWT, like in cells infected with FMDV, in contrast to those expressing LbC51A (Fig 1B). The impact on the cap-dependent expression of a control DDK-tagged protein during co-expression with Lbpro due to eIF4G cleavage is shown for comparison (Fig 1C). In sum, these results suggest that LGP2 is a target for the proteolytic activity of the Leader protease.
Next, we further characterized the Lbpro-LGP2 interaction. When lysates from cells co-expressing LGP2 and Lbpro were analyzed at different times after transfection, a progressive degradation of full-length LGP2, concomitant with detection of the N- and C-terminal products, could be observed (Fig 2A). The 27 KDa C-terminal fragment was also clearly detected with the anti-DDK antibody, consistent with the C-terminal location of the DDK tag in the LGP2 fusion protein transiently expressed. A dose-dependent effect of Lbpro on LGP2 was also observed, with accumulation of the N- and C-terminal products as higher concentrations of the protease were co-expressed with a fixed amount of the helicase (Fig 2B). With the highest amount of Lbpro assayed, the level of detection of all LGP2-derived bands 24 h after transfection decreased drastically, likely due to the extensive processing of the protein by Lb and subsequent degradation of the resulting products. The integrity of eIF4G in the lysates corresponding to the time course and Lbpro dose experiments was also analyzed to monitor the activityof Lbpro in each case (S1A and S1B Fig).
To address whether the caspase or the proteasome pathways were involved in LGP2 cleavage, the caspase inhibitor zVAD-FMK or the proteasome inhibitor MG132 was added to the transfection medium during LGP2 expression assays. As shown in Fig 2C, induction of apoptosis or proteasome did not result in LGP2 cleavage. In contrast, eIF4G analysis revealed the presence of the expected caspase-dependent fragments [34] (S1C Fig). Additionally, the 49 KDa N-terminal and 27 KDa C-terminal fragments were generated when LGP2 and Lbpro were co-expressed in the presence of the inhibitors. These results suggest that LGP2 cleavage was specifically attributable to the protease activity of the Lb protein and not a result of activation of cellular apoptosis and proteasome.
Given that pigs are among the most relevant natural host species for FMDV, we next assessed the ability of Lbpro to process porcine LGP2. For this purpose, we expressed the porcine helicase fused to an N-terminal DDK tag, as we failed in our attempts to detect the endogenous LGP2. Amino acid sequence alignment showed an 82% identity between human and porcine LGP2 proteins. Similarly to human LGP2, we found evidence of porcine LGP2 cleavage when it was co-expressed with the catalytically active form of Lbpro (Fig 2D and 2E). At 24 h after transfection, the full length porcine LGP2 was hardly detectable and no effect could be observed co-expressing the inactive form of the protease LbC51A (Fig 2D). Consistently, an N-terminal fragment of similar size to that generated from the human protein was readily detected with the antibody against the N-terminal region of LGP2 or the anti-tag antibody. The N-terminal LGP2 cleavage product generated by the activity of Lbpro accumulated over time after transfection (Fig 2E). The N-terminal fragment of porcine LGP2 showed a slightly slower migration than that derived from the human helicase (50 KDa approximately), consistent with the presence of the N-terminal DDK tag (Fig 2F).
Interestingly, no C-terminal products were found when using the specific antibody against the LGP2 C-terminal region in lysates from cells co-expressing the porcine helicase and LbWT (Fig 2D and 2E). To further address that issue and rule out any interference with the cross-reactivity of the antibody, raised against the C-terminal region of the human LGP2, we made a new construct for the expression of porcine LGP2 fused to a C-terminal Myc tag. Using an anti-Myc antibody we were unable to detect any LGP2-derived product around 27 KDa when porcine LGP2 and Lbpro were co-expressed, unlike that generated from human LGP2, and only a very faint product of approximately 18 KDa could be detected (Fig 2G). The cleavage pattern of porcine LGP2 was further confirmed in porcine SK6 cells (Fig 2H). The difference in the LGP2 C-terminal patterns observed between the human and porcine helicases induced by Lbpro might be the result of a more efficient degradation of the cleavage product generated from the porcine protein. Indeed, smaller degradation products can also be observed as the N-terminal fragments from both human and porcine LGP2 accumulate in transfected cells (Fig 2C, 2F and 2G). Additionally, secondary cleavage sites for Lbpro might be present in the C-terminal region of porcine LGP2.
Taken together, these results show that the FMDV Lbpro specifically cleaves human as well as porcine LGP2 when both helicase and protease are co-expressed in either human HEK293 or porcine SK6 cells.
Having established that LGP2 is a target for the catalytic activity of Lbpro, we sought to determine whether Lbpro interacts with LGP2. First, we carried out co-immunoprecipitation (coIP) assays in porcine SK6 cells co-expressing both proteins (Fig 3A). As expected, both full-length LGP2 and its N-terminal fragment were efficiently pulled down by the anti-tag antibody. Two different concentrations were used for SDS-PAGE analysis in order to separate the approximately 49 KDa LGP2 fragment from the 50 KDa IgG heavy chain band. We found that both LbWT and the inactive LbC51A mutant co-immunoprecipitated with LGP2, while no Lb was detected when co-expressed with the control tagged vector. According to the intensity of the LbWT and LbC51A bands in the corresponding IP fractions, it seems that both Lb forms are able to bind to full-length LGP2. As the amount of intact LGP2 24 h after co-transfection with LbWT is scarce, the low amount of protease immunoprecipitated with LGP2 detected as a faint band (Fig 3A). In contrast, LbC51A is accumulated in transfected cells and its interaction with LGP2 was readily detected. Additionally, the interaction between Lb and LGP2 might be abolished after cleavage contributing to a better detection of the LbC51A-LGP2 interaction. We also found by confocal microscopy that both LbWT as well as LbC51A co-localized with LGP2 when transiently co-expressed in BHK21 cells (Fig 3B). Taken together, these results suggest that the FMDV Lpro and LGP2 physically interact in vivo.
Having shown that Lbpro interacts with and cleaves LGP2, we hypothesized that the helicase cleavage event could play a role on viral host immune evasion. First, we determined whether FMDV infection induced LGP2 cleavage. For that purpose, human or porcine LGP2 was expressed in swine SK6 cells that were then infected with FMDV and lysed at different times after infection. Viruses in the supernatants collected from transfected and infected cells were titrated at the corresponding time points (Fig 4A and 4B). In these assays, two serologically and genetically divergent FMDV isolates were used: type-O O1BFS and type-C CS8. When SK6 cells expressing human LGP2 were infected with FMDV, the N- and C-terminal LGP2 cleavage products were clearly detected at 2 or 4 h post-infection (hpi) for CS8 or O1BFS isolates, respectively and up to 8 hpi (Fig 4A). When porcine LGP2 was expressed, FMDV infection with O1BFS or CS8 isolates generated the LGP2 N-terminal product of about 50 KDa which could be detected between 4–8 hpi using antibodies against either the N-terminal region of LGP2 or the DDK-tag (Fig 4B). Similarly to assays using ectopically expressed Lbpro, no C-terminal products derived from porcine LGP2 could be detected in human or swine cells. In all cases, detection of the human or porcine LGP2 cleavage products correlated with accumulation of the FMDV-encoded Lpro which could be often detected as a doublet, including Lb and a slower migrating form corresponding to Lab (Fig 4A and 4B). Also, LGP2 cleavage products detection coincided with times of higher viral titers (Fig 4A and 4B). These results show that LGP2 cleavage occurs during FMDV infection and the cleavage patterns observed are equivalent to those found in the above experiments using transiently expressed Lbpro (Fig 2). The impact of FMDV infection on eIF4G as result of Lpro activity on a known cellular target was monitored over time in SK6 cells transfected with human or porcine LGP2 (S2A and S2B Fig, respectively). While eIF4G cleavage was complete at 8 hpi (S2 Fig), full-length LGP2 was still abundant in cells (Fig 4A and 4B), likely due to a lower affinity for LGP2 together with the excess of overexpressed protein within cells at the time of infection.
Next, the pattern of LGP2 was analyzed at different times after infection with another aphthovirus, equine rhinitis A virus (ERAV). A drastic decrease in full-length LGP2 was observed after 24 h of infection and we were able to detect the 27 KDa C-terminal cleavage product at 48 hpi (S3 Fig). In contrast, during infection with different swine viruses causing a clinical disease similar to FMDV, like swine vesicular disease virus (SVDV)—a picornavirus -, or vesicular stomatitis virus (VSV)—a member of the Rhabdoviridae family, no cleavage products were detected and the levels of full-length LGP2 were maintained with no obvious decrease associated with infection (S4A and S4B Fig). Similar results were found when infection by two other distantly related picornaviruses—Aichivirus (Aiv) or encephalomyocarditis virus (EMCV)—was analyzed for LGP2 cleavage (S4C and S4D Fig). Again, no decrease in full-length LGP2 levels or any cleavage products could be detected. Interestingly, with the exception of ERAV, none of the picornaviruses analyzed express an active Leader protease. Altogether, these results suggest that LGP2 cleavage is not a general event during the course of infection by picornaviruses or vesicular swine viruses, but a specific mechanism occurring during FMDV infection and likely shared among aphthoviruses.
It has been shown that LGP2 overexpression negatively affects FMDV replication in cultured cells [30]. To determine whether LGP2 cleavage by Lpro is involved in IFN antagonism operating during FMDV infection, the effect of LGP2 and Lpro co-expression on the resulting viral titers, IFN-β mRNA levels and antiviral activity induced was analyzed in swine SK6 cells (Fig 5). First, the viral titers after 8 h of infection in cells co-expressing porcine LGP2 and either LbWT or inactive mutant LbC51A were compared (Fig 5A). As expected, expression of LGP2 induced a significant reduction in viral titers. Interestingly, co-expression of LbWT restored the viral titers recovered from control cells and no significant differences were found between cells co-expressing LGP2 and LbWT and those transfected with the EV alone or with LbWT or LbC51A independently. However, co-expression of LGP2 and LbC51A failed to have a stimulatory effect on FMDV replication and viral titers were equivalent to those obtained when LGP2 alone was expressed. When the integrity of LGP2 was analyzed, the N-terminal cleavage fragment could be detected in cells co-transfected with EV or LbC51A (Fig 5A), consistent with the cleavage pattern observed during infection (Fig 4). In contrast, no full-length LGP2 could be detected in cells co-expressing LbWT, suggesting an additive cleavage effect of overexpressed LbWT and FMDV-encoded Lpro on LGP2, though a putative contribution of the 3Cpro activity has not been analyzed and cannot be ruled out. In all cases, complete cleavage of eIF4G was observed, as expected at 8 h after infection (Fig 5A).
Since the Lpro activity is able to subvert the inhibitory effect mediated by LGP2 expression, ultimately promoting viral growth, we sought to address whether the mechanism behind that effect was related to type-I IFN induction. For that, the mRNA levels of porcine IFN-β were analyzed by RT-qPCR in SK6 cells transfected and infected as above (Fig 5B). Consistent with the different viral titers recovered from cells co-expressing porcine LGP2 and either LbWT or LbC51A, the IFN-β mRNA levels in cells co-transfected with LGP2 and LbWT were significantly lower than those measured in cells co-expressing LGP2 and either inactive LbC51A or an empty vector (100- and 128-fold lower, respectively) (Fig 5B). We also observed that lower levels of IFN-β mRNA were induced when LbC51A and LGP2 were co-expressed compared to cells co-expressing LGP2 and an EV. This could be due to interference by the inactive LbC51A, that is still able to bind LGP2, with the antiviral response triggered by the helicase. Additionally, a residual protease activity of the LbC51A, expressed at high levels in transfected cells, might contribute to the IFN reduction observed. In any case, this difference only seemed to induce a small insignificant reduction in viral titers.
Altogether, we concluded that, when LGP2 is transiently overexpressed, co-expression of catalytically active Lpro was associated with lower levels of IFN-β induction and higher levels of FMDV replication, in correlation also with complete cleavage of the helicase. Next, we aimed to determine whether the differences in IFN-β mRNA induction and FMDV titers described above correlated with different levels of antiviral activity present in the supernatants of the corresponding cells. Then, we carried out an IFN bioassay based on VSV infection inhibition in cells pre-treated with the supernatants from SK6 cells transfected and infected as above. The antiviral activity in each case reflects the protein levels of type-I IFN effectively secreted after mRNA induction and biologically active against viral infection. As shown in Fig 5C, antiviral activity was detected and measured in supernatants from cells co-transfected with LGP2 and either the EV or LbC51A but not with LbWT. This is consistent with the low levels of IFN-β mRNA measured by RT-qPCR and the higher FMDV titers recovered from those cells (Fig 5A and 5B). No antiviral activity could be detected in mock-transfected/infected or non-infected control cells either. The antiviral activity found in cells expressing LGP2 alone or together with inactive LbC51A was completely neutralized by incubation of the supernatants with a monoclonal antibody against porcine IFN-α, while no inhibitory effect on VSV infection was observed in untreated cells, proving the specificity of the neutralizations observed and further confirming that the antiviral activity subverted by Lbpro was indeed type-I IFN specific.
Though several cellular proteins are known targets for Lpro cleavage, only a few substrate sequences which, however, do not share a unique amino acid sequence motif, have been identified experimentally. Given that the estimated molecular weights of the N- and C-terminal fragments generated after Lbpro cleavage of human LGP2 seem to add up to the size of the full-length protein (77 KDa), suggesting a single cleavage site, and the similar size observed for the N-terminal fragment cleaved from the porcine protein (about 50 KDa), the amino acid sequence around the putative cleavage region was analyzed for similarities with previously reported Lpro target sequences. A stretch of positively charged R amino acids (RGRAR) resembling the (R)(R/K)(L/A)(R) target motif defined for Gemin5 and Daxx (death-domain associated protein) proteins [35] was identified in the conserved helicase motif VI of both human and porcine LGP2 sequences (Fig 6A). To determine the relevance of the arginine residues in the candidate target sequence, a triple substitution mutant to negatively charged glutamic acid (RGRAR/EGEAE) was generated in the human protein (Fig 6A). Unlike LGP2WT, when the LGP2 triple mutant (LGP2MT) was co-transfected with Lbpro no cleavage products could be detected, though the expected decrease in the cap-dependent expression of the tagged-polypeptide, concomitant with Lpro expression was evident (Fig 6B). These results suggest that residues 69–73 in human LGP2 (corresponding to amino acids 72–76 in the porcine protein) are a target motif for Lpro proteolytic activity.
In this study, we identified the innate immune sensor LGP2 as a target for the FMDV Leader protease. The IFN system is a powerful component of the antiviral response and viruses have evolved sophisticated strategies to evade the host innate immune response by interfering with the different events involved in PRR activation and signaling [15,16]. FMDV is no exception, and viral proteases have been found to counteract the innate responses induced in cells during the course of infection [17,18]. Lpro is known to prevent the host antiviral response by several mechanisms including cleavage of initiation factor eIF4G - and then prevention of the synthesis of IFN and other cytokines immediately after infection-, degradation of NF-κB, and deubiquitination of immune signaling molecules [18,24]. Though the contribution of LGP2 to innate immune activation is still not fully understood, recent work unveils a relevant regulatory role on RLR signaling through CARD-independent interactions. The dsRNA generated during picornavirus replication is sensed by MDA5, and LGP2 is believed to promote the viral RNA-MDA5 interaction leading to efficient antiviral signaling [36]. Indeed, RNA derived from EMCV infection with strong MDA5-stimulatory activity was immunoprecipitated with LGP2 [37]. It is also known that either LGP2 or MDA5 deficiency results in higher susceptibility to picornavirus infections [38,39]. A recent report shows that, in the absence of infection or viral proteins, LGP2 functioned as a biphasic master activator of numerous innate immunity genes, sequentially induced in a cascade fashion leading to production of IFN. In turn, LGP2 was subject to negative control by cellular translation regulators [40].
Here, we first provide evidence that LGP2 is cleaved by the FMDV Lbpro resulting in a drastic decrease of full-length LGP2 and accumulation of specific cleavage products in an Lb-dose-dependent manner. Catalytically inactive mutant LbC51A had no effect on LGP2. We found that LGP2 cleavage was specifically attributable to Lbpro catalytic activity and not a result of activation of cellular apoptosis or proteasome. When the patterns of human and porcine LGP2 after co-expression with Lbpro were compared, a similar N-terminal fragment of about 50 KDa was observed in both cases, while an additional C-terminal fragment of 27 KDa was only detected for human LGP2.
Though several cellular proteins are known targets for Lpro cleavage, only a few substrate sequences have been identified experimentally and there is no consensus for Lpro target sequence. Together with the L/VP4 junction of the viral polyprotein, cleavage sites in eIF4GI, eIF4GII, Gemin5 and Daxx have been determined. The analysis of the amino acid sequence around the putative cleavage region revealed a conserved motif in human and porcine LGP2 proteins resembling the Lpro target site (R)(R/K)(L/A)(R) reported in Gemin5 and Daxx [35]. Replacement of all three positively charged R residues by E (RGAR/EGEAE) in human LGP2 protein completely abolished cleavage by Lbpro, defining the RGRAR sequence in the conserved helicase motif VI as the FMDV Leader protease target site in LGP2. This result further increases the number of host factors cleaved by this protease, opening new avenues for the identification of novel targets. Cleavage at that position would excise a 27 KDa C-terminal fragment, in agreement with our results, involving removal of part of the helicase domain and the complete CTD region. An intact CTD is required for RNA specific recognition, and LGP2 mutants in the CTD, helicase domain or both are known to be RNA binding-deficient and have poor enhancing activity towards MDA5 [7]. Thus, cleavage by Lpro would likely abolish the antiviral function of LGP2.
Additionally, we found that both active LbWT and inactive LbC51A were able to interact with LGP2, though LbC51A co-immunoprecipitated with the helicase more efficiently than LbWT. Besides the higher levels of expression achieved with the inactive version of the protease, these results suggest that Lbpro forms a transient interaction with LGP2 that is abolished after LGP2 cleavage. This is also consistent with the low levels of full-length LGP2 and high levels of the N-terminal fragment present in the lysates when LbWT was expressed.
Given that LGP2 cleavage by Lpro was found using the ectopically expressed protease, we analyzed the fate of overexpressed LGP2 during FMDV infection in order to address the biological relevance of the cleavage event. Interestingly, the human and porcine helicase patterns observed during FMDV infection with two different viral isolates overlapped with those observed under LbWT overexpression. Consistently, the corresponding N- and C-terminal LGP2 cleavage products accumulated in cells around 4 h after infection, coinciding with accumulation of viral Lpro, and thus supporting that LGP2 is processed by Lpro during FMDV infection.
Evidence of LGP2 cleavage was also found during infection with the equine aphthovirus ERAV, which shares with FMDV, unlike most picornaviruses, the presence of a proteolytically active Leader protein at the N-terminus of the polyprotein [41]. In contrast, no evidence of LGP2 cleavage or noticeable decrease was observed during infection with a set of different viruses, including unrelated picornaviruses AiV and EMCV, and other RNA viruses causing swine vesicular disease similar to FMD like SVDV—picornavirus—and VSV—rhabdovirus, suggesting that no evident specific mechanisms targeting LGP2 integrity were exerted by these other viruses. In agreement with these data, Feng et al. could not find any sign of LGP2 cleavage in HeLa cells during infection with Coxsackievirus B3, a subtype of Enterovirus B, another picornavirus [42]. Interestingly, Gemin5 cleavage by Lpro was only observed in FMDV-infected cells but not during infection with picornaviruses belonging to different genera like SVDV or EMCV [35].
A previous study showed a detrimental effect of LGP2 expression on FMDV replication, as well as a decrease in the helicase levels when different viral proteins—including Lpro, 3C and 2B - were expressed [30]. The authors concluded that 2B interaction with LGP2 was responsible for this effect and linked it to regulation of the inflammatory response in infected cells by an unknown mechanism. In this study, we showed that LGP2 cleavage by Lpro is involved in IFN antagonism operating during FMDV infection. Expression of catalytically active Lpro was able to subvert the IFN-β induction and inhibitory effect on viral growth mediated by LGP2, circumventing the type-I IFN specific antiviral activity induced in porcine cells that had been transfected and later infected with FMDV.
A potential role on LGP2 antagonism has been suggested for the binding of paramyxovirus V and HCV NS3proteins to the helicase domain of LGP2. LGP2-V protein interaction disrupts the ATP hydrolysis activity of LGP2 [28]. The HCV NS3 protein has a protease domain at its N-terminus linked to a C-terminal helicase domain. LGP2-NS3 interaction might contribute to localize NS3 to mitochondria for MAVS cleavage [29]. To our knowledge, this is the first report of an LGP2 cleavage event involving a virally encoded protein, with implications in the resulting type-I IFN response of the host against viral infection. FMDV is highly sensitive to IFN, and IFN-based strategies have proved to be efficient biotherapeutic approaches against the virus [43–45]. Cleavage of LGP2 by the Leader protease unveils a new antagonistic mechanism evolved by FMDV, and likely other aphthoviruses, directed to suppress the host IFN response. LGP2 cleavage expands the list of cellular proteins involved in immune response targeted by the FMDV Lpro, highlighting its relevance as a crucial proteolytic virulence factor. The suppressor effect exerted by Lpro on the LGP2-dependent type-I IFN induction may be the result of blockade at several points of the pleiotropic activation of innate responses orchestrated by LGP2, included its synergistic effect on MDA5 signaling. Our findings encourage further studies focused on the role of LGP2 on the antiviral response against FMDV and the putative involvement of MDA5 in the LGP2-Lpro interplay. A better understanding of the mechanisms employed by viruses to circumvent the host antiviral signaling will contribute to development of new therapeutic strategies to fight viral infections, including antiviral approaches and novel vaccines. This is of particular relevance for FMDV, given the rapid spread of the virus and the devastating economic consequences associated with FMD outbreaks.
HEK293, Vero and BHK21 cells (all three from ATTC) and SK6 and IBRS-2 cells (both obtained from Centro de Investigación en Sanidad Animal, CISA-INIA, Madrid, Spain) were all cultured in Dulbecco’s modified Eagle’s medium (DMEM; GIBCO) supplemented with 10% fetal bovine serum at 37 oC with 5% CO2. These cell lines were used for propagation and infection assays of the corresponding viruses.
FMDV type-C CS8 and type-O O1BFS isolates were propagated in swine SK6 or IBRS-2 cells. ERAV and Aichivirus were propagated in Vero cells. SVDV was propagated in SK6 cells. VSV was propagated in IBRS-2 cells.
Plasmids encoding the WT or C51A mutant Lb protease were generated by PCR amplification of the corresponding regions from an FMDV O1K full-length cDNA clone [46] and insertion into the BamHI and XbaI sites of pcDNA3.1(+) (Invitrogen). Plasmids encoding the sequence of porcine LGP2 with a C-terminal Myc tag and/or an N-terminal DDK tag were generated by gene synthesis (NZYTech) and cloning into the NheI and XbaI sites of pcDNA3.1(+) (Invitrogen). Plasmid encoding (C-terminal Myc-DDK-tagged)-human LGP2 was from Origen. Plasmid pcDNA3/Flag-METTL3 encoding human methyltransferase-like 3 was from Addgene (# 53739).
For transfection, 2 μg of LGP2-encoding plasmids and 1 μg of plasmids encoding FMDV proteases were used using Lipofectamine 2000 (Invitrogen) following the manufacturer's recommendations. The total amount of transfected DNA was balanced to 3 μg with empty vector. In some experiments, the transfection medium was supplemented with 20 μM Puromycin (apoptosis inducer, Sigma-Aldrich), 20 μM zVAD-FMK (broad caspase inhibitor, Promega) or 10 μM MG132 (proteasome inhibitor, Cayman Chemical).
Mouse monoclonal anti-LGP2 (E-1, raised against a peptide mapping at the C-terminus of human LGP2) and goat polyclonal anti-LGP2 (N-14, raised against a peptide mapping near the N-terminus of human LGP2) were purchased from Santa Cruz Biotechnology Inc. Mouse monoclonal anti-FLAG (M2) was purchased from Sigma-Aldrich. Rabbit polyclonal anti-PARP and mouse monoclonal anti-cleaved PARP (Asp214) (19F4) were purchased from Cell Signaling Technology. Mouse monoclonal anti-Pig IFN-Alpha (K9) was purchased from PBL Assay Science. Rabbit polyclonal anti-FMDV Leader protease was raised against the Lab/Lb fusion protein expressed by pE16 plasmid [47] and kindly provided by Ewald Beck. Rabbit polyclonal anti-βII-tubulin [48] was achieved from Sobrino F Lab. Mouse monoclonal anti-G3BP (clone 23) was purchased from BD Biosciences. Goat polyclonal anti-eIF4G (D-20) antibody was purchased from Santa Cruz Biotechnology Inc. Goat anti-mouse, goat anti-rabbit and rabbit anti-goat IgG (H + L) secondary antibodies HRP conjugate were purchased from Thermo Scientific.
Cells were transfected with indicated plasmids with Lipofectamine 2000 (Invitrogen) according to the manufacture’s protocol and 24 h later, cells were infected with the corresponding viruses. At different times after infection, supernatants were harvested, serially diluted and viral titers were determined by plaque assay on fresh monolayers. After 1 h of infection, cells were washed twice and overlaid with 0.5% agar. After 24 h, cells were fixed with 10% formalin and stained with crystal violet. Viral titers were expressed as plaque forming unit (pfu)/ml. The mean values and standard deviations were calculated from triplicate determinations.
Cells were harvested after transfection or infection, and lysed with PBS containing 1% NP-40, 1 mM DTT and protease inhibitor cocktail (Complete, Roche). Whole cell lysates were incubated at room temperature for 5 min and cleared by centrifugation at 9.300 x g for 5 min at 4°C. Protein concentrations were determined based on the Bradford method using the Bio-Rad protein assay kit. Equal amounts of proteins (20–50 μg) were separated by 6–12% SDS-PAGE and electrophoretically transferred onto a nitrocellulose membrane (GE Healthcare). After blocking with 3% non-fat milk in 0.05% Tween20 PBS, the membranes were incubated with the primary antibodies, followed by horseradish peroxidase-conjugated goat anti-rabbit, anti-mouse or rabbit anti-goat IgG. Membrane bound antibodies were detected by enhanced chemiluminescent luminol substrate (Western Lightning Plus Chemiluminescent Substrate kit, Perkin Elmer) and visualized by exposure to X-ray films.
1x106 SK6 cells were co-transfected with 2 μg of DDK-poLGP2 or DDK-vector plasmids together with 1 μg of plasmids encoding LbWT or LbC51A. Cells were harvested 24 h later and lysed with 100 μl of lysis buffer (50 mMTris-HCl [pH, 7.5], 150 mMNaCl, 0.5% NP-40, and protease inhibitor cocktail). The supernatants were collected by centrifugation at 10,000 x g for 5 min at 4°C and precleared with 25 μl of protein G-Agarose (Roche) for 1 h at 4°C with rotation. Proteins were immunoprecipitated by addition of monoclonal anti DDK and incubation for 4 h at 4°C and then, addition of protein A agarose and incubation at 4°C for 16 h. Immunoprecipitated complexes were washed three times with 400 μl of wash buffer 1 (0.1% NP-40, 50 mM Tris pH 7.5, 150 mM NaCl) and once with 400 μl of wash buffer 2 (50 mM Tris pH 7.5, 150 mM NaCl). Then, beads were collected via centrifugation at 10,000 x g for 2 min at 4°C, boiled at 100°C for 3 min in SDS protein-loading buffer and analyzed via WB.
BHK-21 cells were transfected with indicated plasmids with Lipofectamine 2000 (Invitrogen) according to the manufacture’s protocol. After 20 h, cells were washed three times with room temperature PBS, then fixed with 4% paraformaldehyde solution in PBS for 20 min, washed again with PBS three times and permeabilized with 0.05% Tween in PBS for 15 min. After wash three times with PBS, cells were blocked with 5% BSA in PBS for 1 h. Cells then were incubated with specific primary antibodies overnight at 4°C, followed by incubation for 1h with goat anti-mouse Alexa fluor 488 and goat anti-rabbit Alexa fluor 647 as secondary antibody. Nuclei were stained with 4',6-diamidino-2-phenylindole (DAPI) at 1 μg/ml. Images were acquired with a Zeiss LSM 880 Meta confocal microscope (Carl Zeiss Microimaging, Thornwood, NY) with a Plan Apochromatic ×40/1.4 oil objective lens. Image processing and analysis of intensity of fluorescence by histograms were carried out using Fiji/ImageJ software.
Total RNA was isolated from SK6 cells using TriReagent (Sigma), quantified by spectrometry and DNase-treated with Turbo DNA-free kit (Ambion). 500 ng of RNA was used for RT with 20U of SuperScript III RT (Invitrogen) at 55 oC for 30min. Quantitative PCR was carried out using aliquots of the RT reactions (1/10) and LightCycler FastStart DNA master SYBR green I (Roche). All reactions were conducted in triplicate. Data were analyzed using the ΔΔCT method. IFN-β gene expression was normalized to that of the GAPDH and was expressed as the fold increase above the level of mock-transfected cells. Primers for amplification of IFN-β and GAPDH have been previously described [49].
The antiviral activity in supernatants from transfected and/or infected SK-6 cells was determined by a VSV infection inhibition assay on IBRS-2 cells (IFN bioassay) as described [49]. Briefly, after transfection, SK-6 cells were incubated at 37°C and 24 h later infected with FMDV CS8 isolate at an MOI of 5. Supernatants were collected 7 h later and infectious particles were inactivated by acidic treatment (pH = 2–3) with 10M HCl for 16–20 h at 4°C and then, neutralized (pH = 7) with 10 M NaOH. Dilutions of the treated supernatants (up to 1/15) were added on fresh IBRS2 monolayers and incubated for 24 h at 37°C. Then, cells were washed and infected with VSV (50–60 pfu per 1x106 cells) and plaques were counted 24 h after infection. Antiviral activity was defined as the reciprocal of the highest dilution resulting in a 50% reduction in the number of plaques relative to untreated cells. In order to block the antiviral activity exerted by IFN-α in SK6 cells supernatants on VSV infection, some samples were incubated for 1 h at 37°C with 1μg of a monoclonal antibody against swine IFN-α (clone K9) from PBL Assay Science.
The unpaired Student’s t test for independent samples was used to compare data using IBM SPSS Statistical (v.24) software; a p value of < 0.05 was considered statistically significant and a p value of > 0.05 was considered statistically non-significant. In all graphs, three asterisks indicate a p value of <0.001, two asterisks indicate a p value of <0.01, one asterisk indicates a p value of <0.05, and ns indicates not significant (p > 0.05). The number of replicates used in experiments is specified in the corresponding figure legends.
|
10.1371/journal.ppat.1005133 | Mitochondrial Activity and Cyr1 Are Key Regulators of Ras1 Activation of C. albicans Virulence Pathways | Candida albicans is both a major fungal pathogen and a member of the commensal human microflora. The morphological switch from yeast to hyphal growth is associated with disease and many environmental factors are known to influence the yeast-to-hyphae switch. The Ras1-Cyr1-PKA pathway is a major regulator of C. albicans morphogenesis as well as biofilm formation and white-opaque switching. Previous studies have shown that hyphal growth is strongly repressed by mitochondrial inhibitors. Here, we show that mitochondrial inhibitors strongly decreased Ras1 GTP-binding and activity in C. albicans and similar effects were observed in other Candida species. Consistent with there being a connection between respiratory activity and GTP-Ras1 binding, mutants lacking complex I or complex IV grew as yeast in hypha-inducing conditions, had lower levels of GTP-Ras1, and Ras1 GTP-binding was unaffected by respiratory inhibitors. Mitochondria-perturbing agents decreased intracellular ATP concentrations and metabolomics analyses of cells grown with different respiratory inhibitors found consistent perturbation of pyruvate metabolism and the TCA cycle, changes in redox state, increased catabolism of lipids, and decreased sterol content which suggested increased AMP kinase activity. Biochemical and genetic experiments provide strong evidence for a model in which the activation of Ras1 is controlled by ATP levels in an AMP kinase independent manner. The Ras1 GTPase activating protein, Ira2, but not the Ras1 guanine nucleotide exchange factor, Cdc25, was required for the reduction of Ras1-GTP in response to inhibitor-mediated reduction of ATP levels. Furthermore, Cyr1, a well-characterized Ras1 effector, participated in the control of Ras1-GTP binding in response to decreased mitochondrial activity suggesting a revised model for Ras1 and Cyr1 signaling in which Cyr1 and Ras1 influence each other and, together with Ira2, seem to form a master-regulatory complex necessary to integrate different environmental and intracellular signals, including metabolic status, to decide the fate of cellular morphology.
| Candida albicans is a successful fungal commensal and pathogen of humans. It is a polymorphic organism and the ability to switch from yeast to hyphal growth is associated with the commensal-to-pathogen switch. Previous research identified the Ras1-cAMP-protein kinase A pathway as a key regulator of hyphal growth. Here, we report that mitochondrial activity plays a key role in Ras1 activation, as respiratory inhibition decreased Ras1 activity and Ras1-dependent filamentation. We found that intracellular ATP modulates Ras1 activity through a pathway involving the GTPase activating protein Ira2 and the adenylate cyclase Cyr1. Based on our data the canonical Ras1 signaling model in C. albicans needs to be restructured in such a way that Cyr1 is no longer placed downstream of Ras1 but rather in a major signaling node with Ras1 and Ira2. Our studies suggest that the energy status of the cell is the most important signal involved in the decision of C. albicans to undergo the yeast-to-hyphae switch or express genes associated with the hyphal morphology as low intracellular ATP or associated cues override several hypha-inducing signals. Future studies will show if this knowledge can be used to develop therapies that would favor benign host-Candida interactions by promoting low Ras1 activity.
| Candida albicans, one of the most common human fungal pathogens, is an important cause of morbidity and mortality in immunocompromised individuals, particularly in patients with AIDS or those undergoing cancer chemotherapy or transplantation procedures [1]. The prolonged use of antifungal agents in such compromised populations can lead to an increase in C. albicans resistance to many currently used therapies [2]. For this reason, there is an immediate need for new treatment options that can prevent or control diseases caused by C. albicans.
In addition to being a pathogen, C. albicans is also a member of the commensal microflora of most individuals and the transition from commensal to pathogen is associated with the morphological switch from yeast to hyphal growth [3–5]. Environmental factors like 37°C, 5% CO2, N-acetylglucosamine, pH, and serum, induce the yeast-to-hyphae switch [6]. However, most of these signals are always present in vivo, thus we still do not understand what governs the switch from benign colonization to symptomatic infection. During host colonization C. albicans lives amidst other microbes and, both, clinical data, that suggest a link between antibiotic usage and increased risk of fungal infections, and laboratory studies indicate that C. albicans interacts with bacteria in biologically important ways [7–14]. Further studies on bacterial-fungal interaction have led to the identification of new ways by which microbes modulate C. albicans growth. For example, 3-oxo-C12-homoserine lactone, produced by the Gram-negative bacterium Pseudomonas aeruginosa, inhibits hyphal growth by directly inhibiting the fungal Ras1-cAMP-protein kinase A (PKA) signaling pathway, a key regulator of the yeast-to-hyphae switch in C. albicans, by blocking cAMP synthesis [15,16].
The Ras1-cAMP-PKA signaling pathway is critical for C. albicans virulence in animal models [17–19]. Ras1 is a small GTPase that exists in the cell in an inactive (GDP-bound) form and an active (GTP-bound) form whose switch is regulated by the guanine nucleotide exchange factor (GEF) Cdc25 and GTPase-activating protein (GAP) Ira2 [20]. In its GTP-bound form, Ras1 directly interacts with the adenylate cyclase Cyr1 and stimulates cAMP production [18,21,22]. The cAMP signal subsequently derepresses two PKA isoforms which promote several cellular processes [23,24]. In current models of virulence, activation of the Ras1-cAMP-PKA pathway by host-associated stimuli induces the transition from yeast-to-hyphae growth and the expression of hypha-specific virulence factors. Hyphal growth increases tissue adherence and penetration, as well as the formation of adherent biofilms on medical devices [6,25,26]. This pathway also controls genes involved in glycolysis, stress resistance, cell wall composition, and mating [6,27].
More recently, an additional class of molecules that repress hyphal growth of C. albicans, phenazines, have been identified [28,29]. While phenazines are best known as small-molecule toxins with antibiotic properties toward bacterial and eukaryotic species at high concentrations [30], recent studies have found that some phenazines (phenazine-1-carboxylic acid (PCA), phenazine methosulfate (PMS), and pyocyanin (PYO)) inhibit hyphal growth, intercellular adherence and biofilm development of C. albicans at low sub-lethal (micromolar) concentrations that are more than 100-fold below the concentrations which affect fungal survival [29]. Phenazines were found to inhibit C. albicans respiration [29], which is consistent with other published data that phenazines can impact mitochondrial activity [31–33]. Subsequent analysis indicated that the decreased ability of C. albicans to develop wrinkled colonies (consisting of hyphal and yeast cells) or robust biofilms on plastic was due to inhibition in electron transport chain activity [29]. Indeed it was shown that during the filamentation process C. albicans activated the TCA cycle, inhibited the pentose phosphate pathway, and increased mitochondrial respiration [34]. This suggests that hyphal growth in C. albicans depends on functional respiration to cover the metabolic needs of the cell which is inhibited by phenazines. Early studies with mammalian mitochondria showed that phenazines uncouple oxidative phosphorylation by shunting electrons from endogenous pathways [35–37], and this is most likely how respiration is inhibited in C. albicans. The inhibition of C. albicans filamentation by phenazines occurs despite the presence of robust fermentation pathways capable of supporting rapid growth in the absence of mitochondrial activity, and suggests communication between filamentation inducing pathways and metabolic state [29]. Indeed, in eukaryotes, it is becoming increasingly apparent that signaling pathways that sense and respond to extracellular cues often also incorporate input from the mitochondria themselves or from mitochondrially-derived molecules (like ATP and reactive oxygen species) [38,39].
In this report we show, that respiratory inhibition via genetic or biochemical manipulation decreases Ras1 activity and inhibits Ras1-dependent filamentation in C. albicans. Ras1 activation is also decreased by mitochondrial inhibition in the pathogenic Candida species, Candida parapsilosis and Candida tropicalis. Furthermore, utilizing a NRG1 overexpression strain and an efg1/efg1 null mutant we show that decreased Ras1 signaling in the presence of respiratory inhibitors is independent of morphological change. Filamentation was not repressed by MB in strains lacking Tup1, a hyphal growth repressor, or in a strain overexpressing Ume6, a transcription factor involved in the induction of hyphal growth indicating that the effects of MB on GTP-Ras1 can be circumvented with activation of downstream parts of the pathway. Analysis of overall metabolic changes due to respiratory inhibition shows perturbation of carbon metabolism, evidence for changes in redox state and increased AMP kinase activity (increased β-oxidation, decreased sterol levels). Subsequent analysis showed that intracellular ATP modulates Ras1 activity independent of AMP kinase. Furthermore, while the GEF Cdc25 is dispensable for decreased Ras1 signaling due to respiratory inhibition, the GAP Ira2 is necessary. In addition, the adenylate cyclase Cyr1 is essential for this signaling cascade, showing for the first time that Cyr1 affects Ras1 activation state and with that it is not just a downstream effector of Ras1. Rather Cyr1, Ira2, and Ras1 seem to form a regulatory complex that combines a multitude of signals to decide if the yeast-to-hyphae switch should take place.
Low micromolar concentrations of the bacterially-produced toxin, pyocyanin (PYO), or its thioanalogue, methylene blue (MB), perturb mitochondrial activity [40,41] and repress C. albicans filamentation (Fig 1A) [29]. Exposure to 1.5 μM MB, a compound used therapeutically in humans [41], caused growth solely in the yeast form as indicated by a smooth colony morphology and cellular yeast morphology as determined by microscopy. While under control conditions, C. albicans grew as a mix of yeast and hyphae in wrinkled colonies. (Fig 1A and S1A Fig panels 1 to 4). Furthermore, MB led to decreased expression of hypha specific genes and increased levels of yeast-specific transcripts (Fig 1B). Because the colony phenotype, cellular morphology, and expression profile of cells grown with MB were similar to those of the afilamentous ras1/ras1 mutant (Fig 1A and 1B and S1A Fig panels 5 to 8) we sought to test the hypothesis that MB inhibits the yeast-to-hyphae switch by inhibition of Ras1 signaling, a pathway critical for C. albicans filamentation and virulence in animal models (Fig 1C) [17,18]. Examination of Ras1 protein levels in cells grown on solid medium with and without MB found that MB led to reductions in levels of active GTP-bound Ras1 (GTP-Ras1) without affecting total Ras1 levels (Fig 1D).
When MB was added to C. albicans cultures in liquid YNBAGNP medium, we observed that MB decreased clumping and increased the percentage of cells in the yeast morphology at concentrations of 3 and 6 μM, but not at 1.5 μM, a concentration that completely inhibited filamentation on solid medium (S2A Fig and Fig 1A). Analysis of the fraction of Ras1 in the GTP-bound state found that 3 and 6 μM MB also decreased the fraction of Ras1 in its active form (S2A Fig). To test if these concentrations of MB have an impact on C. albicans growth we used a strain in which the hyphal gene repressor Nrg1 is overexpressed (NRG1-OE). Nrg1 acts downstream of Ras1 and overexpression of this repressor prevents filamentation in the presence of hypha-inducing signals; the use of this strain makes it possible to measure growth via OD600 measurements in the presence of filamentation inducing signals [3,42]. The different concentrations of MB had no or only minimal impact on C. albicans growth excluding this as the reason for a decrease of GTP-Ras1 levels (S2B Fig). Because filamentation is completely inhibited by 1.5 μM MB on solid media, but filamentation is only partially suppressed even at 6 μM in liquid medium, we used colony-grown cells in subsequent assays.
We tested if MB affected GTP-binding of Ras1 in yeast growth conditions, and found that MB did not impact GTP-Ras1 levels (Fig 1D and S2C Fig). This suggests that MB selectively inhibits the increase in GTP-Ras1 that occurs in the presence of filamentation-inducing signals which include 37°C, buffering at pH 7, and the amino acids and N-acetylglucosamine in YNBAGNP medium. To determine whether the decrease of GTP-Ras1 by MB is specific to YNBAGNP, we tested another common filament-inducing condition (YPD + 5% serum at 37°C). Consistent with our findings on medium with N-acetylglucosamine, amino acids, 37°C, and neutral pH, as the hyphal growth inducers, MB led to lower GTP-Ras1 levels when grown on medium with serum and repressed filamentation (S2D Fig and S1B Fig). In liquid YPD + 5% serum, the effects of MB on morphology and GTP-Ras1 levels were modest suggesting that different media create different physiological states in C. albicans (S2E Fig). All further experiments were conducted using YNBAGNP medium as it is a defined stimulus that mimics a number of aspects of the host (pH 7, amino acids, 0.2% glucose).
To test whether a link between MB and Ras1 activation state can also be observed in other Candida species, we examined GTP-Ras1 levels of two other pathogenic Candida species, Candida parapsilosis and Candida tropicalis. These two fungal pathogens also had lower GTP-Ras1 levels on YNBAGNP with MB, indicating that Ras1 activation is also impacted by MB in other Candida species (Fig 2 and S1A Fig panels 9 to 12). Under these conditions, these fungi grow as yeast in the absence and presence of MB.
In mammalian cells, MB decreases oxidative phosphorylation potential by oxidizing NAD(P)H-dependent dehydrogenase (complex I) and directly reducing cytochrome C thereby bypassing proton transfer by complex I and complex III (Fig 3A) [43]. Inhibition of mitochondrial activity with PYO and/or the complex III inhibitor Antimycin A (AA) reduced GTP-Ras1 levels; both PYO and AA also repressed hyphal growth as previously reported (Fig 3B) [29,44]. Mutants lacking complex I (ndh51/ndh51) or complex IV (cox4/cox4) did not filament and had low levels of GTP-Ras1 under control conditions (Fig 3C and S1A Fig panels 13 to 16 for cellular morphology) [45]. GTP-Ras1 levels were not further reduced by MB, and in fact levels increased in cells exposed to MB (Fig 3C). Null mutants lacking complex II (sdh1/sdh1) or both alternative oxidases (aox1-A/aox1-A aox1-B/aox1-B), which do not participate in the formation of the proton gradient, still filamented and had high GTP-Ras1 levels (S3A and S3B Fig). MB also caused a decrease in GTP-Ras1 in the complex II and alternative oxidase mutants comparable to wild type (S3A and S3B Fig).
All three complexes important for high GTP-Ras1 levels under hypha-inducing conditions (complex I, III, and IV) pump protons into the intermembrane space for use in ATP synthesis (Fig 3A). MB reduces ATP synthesis in C. albicans as relative ATP levels were 2.3-fold lower upon growth of the wild type (WT) with MB (Fig 3D). Furthermore, while intracellular ATP in ndh51/ndh51 and cox4/cox4 mutants in control conditions were significantly lower than in the WT (CAF2) (Fig 3D), ATP levels were not reduced by MB (Fig 3C).
To determine if filamentation was required for elevated GTP-Ras1 and higher ATP, we examined the effects of MB on mutant strains that are unable to undergo the yeast-to-hyphae switch. It is well known that the transcription factor Efg1 is an essential regulator for morphogenesis in C. albicans [46]. In hyphae inducing conditions Efg1 is activated through the Ras1–cAMP–PKA pathway and induces the expression of many hyphae specific genes that are essential for the yeast-to-hyphae transition [47,48]. In addition, we tested the NRG1 overexpression strain (NRG1-OE). Because both Efg1 and Nrg1 act downstream of Ras1 we hypothesized that MB effects on Ras1 should be unaffected in these strains. As expected the efg1/efg1 mutant formed a smooth colony consisting of yeast cells in the presence and absence of MB (S4 Fig and S1A Fig panels 17 and 18). The NRG1-OE strain showed a weakly wrinkled colony morphology under control conditions consisting of mainly yeast cells with some elongated yeast cells and short pseudohyphae, while with MB a completely smooth colony consisting of only yeast cells was observed (Fig 4A). Subsequent western blot analysis and intracellular ATP measurements showed that both strains had less GTP-Ras1 and less ATP with MB, as the WT (Fig 4A and 4B and S4 Fig). The reduction of GTP-Ras1 in WT ranged from 26.5% to 94.5% with an average reduction of 63.6% over all experiments done; even in assays with only a 26.5% reduction in the ratio of GTP-Ras1/ total Ras1 relative to control, filamentation was repressed.
Furthermore, we tested the effects of MB on mutant strains that are constitutively filamentous due to the alteration of downstream transcriptional regulators of hyphal growth to determine if the effects of MB were upstream in the hyphal growth pathway and if hyphal growth could be reactivated in the presence of MB. Loss of the hyphal gene repressor Tup1 or overexpression of the transcription factor Ume6 have been previously shown to result in constitutive filamentation [3,49,50]. Both strains are able to filament in the presence of MB, while GTP-Ras1 levels are decreased (Fig 4C). Filamentation and wrinkled colony formation of these strains is not as strong as under control conditions. However, overall this shows that the effects of MB on ATP and GTP-Ras1 are upstream events in the control of C. albicans morphology and that low Ras1 signaling inhibits filamentation in the WT.
While MB, PYO, and AA all modulate mitochondrial activity and reduced relative GTP-Ras1 levels (Fig 3B), these compounds are not equivalent. For example, MB did not impact growth rates, while PYO and AA were inhibitory perhaps in part due to increased ROS formation. In addition, while AA and PYO led to acidification of the medium due to increased fermentation to acetate, MB did not [29,44]. Thus, we sought to determine the strongest common signals in order to gain insight into factors that control Ras1 GTP-binding. Metabolomics analysis of the WT SC5314 grown under control conditions or with MB, PYO, or AA revealed that some compounds were only differentially regulated in one condition (Fig 5A, S5 Fig, and S1 Table). For instance, only MB-grown cells had significantly high relative levels of glycerol, an alternative fermentation product, and this is consistent with the observation that the medium pH was not altered even in the presence of this fermentation-inducing mitochondrial inhibitor (S5 Fig). A large group of metabolites showed a similar pattern in the presence of all three compounds (Fig 5A); these signatures included increased lipid catabolism and decreased lipid biosynthesis (higher levels of acetyl-CoA, increased lysophospholipids, and decreased fatty acids (palmitate, oleate, and stearate)) as well as low ergosterol and related compounds (Fig 5B and S5 Fig).
The decrease of the fatty acid palmitate is particularly interesting because it is needed for one of two lipid modifications that tether Ras1 in C. albicans to the plasma membrane [51]. Loss of palmitoylation re-localizes Ras1 largely to endomembranes and changes in Ras1 localization negatively affect Ras1 activation [51]. To test if the changes in GTP-Ras1 levels seen with MB or the other inhibitors were due to re-localization of Ras1 away from the membrane, we determined GTP-Ras1 levels in two strains carrying truncated Ras1 proteins that are no longer associated with the plasma membrane. The first strain is a ras1/ras1 mutant reconstituted with a ras1 allele missing the last 67 amino acids (ras1Δ67) [52] and the second is a ras1/ras1 mutant reconstituted with a ras1 allele only including the conserved N-terminal region of RAS1 (ras1 N-term). Interestingly, the ras1Δ67 strain showed lower levels of GTP-Ras1 in control conditions compared to the ras1 N-term or wild type strain (S6 Fig) suggesting a possible GTP-binding inhibitory domain or function activated by Ras1 cleavage [52]. However, both truncated Ras1 variants showed a wild type reduction of GTP-Ras1 levels with MB (S6 Fig). Thus, while Ras1 localization is controlled by its C-terminal lipid modifications, changes in these that might occur in the presence of MB were not responsible for altered GTP-Ras1 levels.
Interestingly, the metabolomics pattern strongly resembled the response of mammalian cells to MB [53]. Furthermore, in mammalian cells the same metabolic shift due to respiratory inhibition is mediated by AMP kinase (AMPK), an energy sensor that responds to relative ATP:AMP/ADP levels. Lipids are a rich source of ATP, and AMPK induces a lipid catabolic state when ATP levels are low [54,55]. These signatures suggest that the common signal in response to PYO, MB and AA is likely low intracellular ATP. GTP-Ras1 levels were not controlled by AMPK itself as a mutant lacking the γ-subunit of AMPK (snf4/snf4) which is essential for AMPK activity in Saccharomyces cerevisiae [56], still shows a reduction in GTP-Ras1 and intracellular ATP upon growth with MB (Fig 5C and 5D and see S1A Fig panels 19 to 22 for cellular morphology). Under control conditions the snf4/snf4 mutant has very low levels of intracellular ATP and is unable to filament (Fig 5C and S1A Fig panels 29).
Across eukaryotes, it has been shown that diverse cellular processes from proteome function to neurotransmitter responses are directly regulated by ATP levels. Thus, phenazine-mediated repression of C. albicans filamentation may occur through effects on ATP levels, as ATP is the precursor to cAMP, a second messenger that is a key positive regulator of hyphal growth (Fig 1C) [6]. PYO reduces levels of both cAMP and its precursor ATP in human epithelial cells due to its effects on respiration and oxidative phosphorylation [57]. To determine more directly if decreased ATP levels were impacting Ras1 signaling, we examined the effects of inhibitors of the proton gradient (dinitrophenol (DNP)) and the ATP synthase (oligomycin) which each caused a significant decrease of intracellular ATP (Fig 6A). For both, relative levels of GTP-Ras1 were decreased and filamentation was repressed (Fig 6B) strongly suggesting that ATP levels were the connecting signal between mitochondrial activity and Ras1 signaling. To further test this hypothesis, we measured GTP-Ras1 levels in the ssn3/ssn3 mutant that had been previously shown to have increased intracellular ATP due to increased oxidative metabolism, without increased growth [44], and found increased GTP-Ras1 levels compared to the reconstituted strain (Fig 6C). In summary, our data show that GTP-Ras1 levels correlate with intracellular levels of ATP.
C. albicans Ras1 GTP-binding has been genetically shown to be controlled by a GEF, Cdc25, and a GAP, Ira2 [58,59]. The cdc25/cdc25 mutant had low levels of GTP-Ras1, and was unable to filament, but MB caused a further reduction in GTP-Ras1 comparable to WT (Fig 7A and see S1A Fig panels 23 and 24 for cellular morphology). In contrast, loss of Ira2 resulted in a hyperfilamentous phenotype and strongly increased levels of GTP-Ras1 which were unaffected by addition of MB (Fig 7B and 7C), while ATP levels were decreased comparable to WT (Fig 7D), showing that the decrease of GTP-Ras1 by MB is Ira2 dependent.
In S. cerevisiae, Ira2 activity is negatively regulated through direct interactions with Tfs1 and positively regulated through protein stabilization by Gpb1/2 [60,61]. While C. albicans tfs1/tfs1 mutants displayed phenotypes consistent with increased Ira2 activity (decreased filamentation and less GTP-Ras1), the reduction of GTP-Ras1 levels upon growth with MB was similar to that of the WT (S7A Fig). Deletion of the C. albicans Gpb1 homolog resulted in increased GTP-Ras1 levels under control conditions that were decreased with MB comparable to wild type (S7B Fig). Together the Gpb1 and Tfs1 data suggest that new inputs into Ira2 may link ATP levels to GTP-Ras1.
We suspected this link may be the adenylate cyclase Cyr1, which is known to be activated by Ras1, and integrates diverse signals. The cyr1/cyr1 mutant, like a ras1/ras1 strain, forms smooth colonies consisting only of yeast (Fig 8 and see S1A Fig panels 25 to 28 for cellular morphology). Surprisingly, the cyr1/cyr1 strain had a higher proportion of GTP-Ras1, and this increase was complemented by addition of the native CYR1 gene. Furthermore, in the absence of Cyr1, Ras1 GTP-binding was not decreased by MB or AA but rather increased (Fig 8A and 8B). The cAMP signal itself appeared to be important, as a strain expressing only a catalytically-inactive Cyr1 (cyr1/cyr1 +cyr11334) also had higher basal GTP-Ras1 levels that were increased and not decreased by MB (Fig 8C and see S1A Fig panels 29 to 32 for cellular morphology). However, neither subunit of PKA, the only known cAMP sensor, was required for the control of GTP-Ras1 levels (S7C Fig).
In summary, our data suggest that low ATP causes Cyr1-mediated activation of Ira2 activity to reduce GTP-Ras1 levels. Thus, it appears that Ras1 and Cyr1 participate in a regulatory circuit that integrates multiple signals before triggering the expression of virulence related attributes.
In this study we identified a previously unknown link between total intracellular ATP levels and Ras1 signaling in C. albicans by characterizing the mechanism by which MB inhibits the C. albicans yeast-to-hypha switch (Fig 9). Interestingly, a recent study in the yeast S. cerevisiae found that dysfunctional mitochondria decrease cAMP-PKA signaling, adhesion production, and filamentous growth further emphasizing that the link between respiratory activity and Ras1-cAMP-PKA signaling is conserved beyond the Candida genus [62]. The same study also showed that the filamentous-growth-specific MAPK pathway is not involved in this signaling as this pathway retained functionality in respiratory-deficient S. cerevisiae yeast cells [62]. Furthermore, while it is not known whether Ras1 signaling is important for filamentation or virulence in C. tropicalis and C. parapsilosis, when grown on YNBAGNP media with and without MB, both Candida species had decreased Ras1 activation state with MB indicating that the link between respiratory activity and Ras1 signaling is conserved across Candida species. However, whether this decrease in Ras1 activation impacts filamentation and virulence of these fungal pathogens needs to be determined in future studies.
In liquid conditions, in which C. albicans hyphal growth is fast, increased MB concentrations were necessary to see a decrease in GTP-Ras1 levels and filamentation. This requirement for higher levels of MB may be due to higher or altered respiratory activity, or differences in ATP homeostasis under well-mixed planktonic conditions. Interestingly, in our assays, GTP-bound Ras1 was lower and filamentation was inhibited by MB on both solid YPD + 5% serum medium and solid and liquid YNBAGNP. However, in liquid YPD + 5% serum conditions, the effect of MB on GTP-Ras1 levels was minor and no impact on morphology at concentrations that were not inhibitory. A recent publication by O’Meara and colleagues reported a global analysis of C. albicans morphology which showed that the role of different pathways in filamentation varied depending on the medium condition [63], and we speculate that the effects of MB on ATP pools, Ras1, and filamentation also varies in different media in ways that we cannot yet understand. Together, this variability shows the importance of understanding the interactions between nutrient sources and growth substrate and signaling inputs and outputs.
The observation that MB had no impact on Ras1 GTP-binding under yeast growth conditions in C. albicans which could indicate that intracellular ATP pools serve as a “check point” in Ras1 signaling under hypha-inducing conditions. The differential effects of MB may be related to temperature (yeast are grown at 30°C while hyphae are grown at 37°C) though MB did modulate ATP levels at 30°C. Indeed, isolated mitochondria from C. albicans were shown to be more active at 37°C compared to 30°C [34]. At lower temperatures, mammalian mitochondria have a reduced respiratory rate and hyperpolarization of the mitochondrial membrane, potentially due to decreased ATPase activity [64–67], which results in the accumulation of reduced flavins and cytochromes (Fig 3A). This state may render cells less susceptible to the action of MB. Increased respiration upon growth at 37°C most likely result in an increase in intracellular ATP concentrations which may promote or permit the hyphal growth program. MB inhibits the accumulation of ATP (Fig 3D) causing cells to stay in the yeast morphology in conditions that would normally induce filamentation.
Mutants defective in complex I and complex IV never establish a high intracellular ATP state, and are unable to undergo the yeast-to-hypha switch. Interestingly, in the presence of MB the ndh51/ndh51 and cox4/cox4 mutants showed an increase of GTP-Ras1 levels instead of a decrease (Fig 3C). This is consistent with observations made in mammalian cells where it has been shown that MB can restore some electron flow to dysfunctional mitochondria [43]. There was only a small but measurable increase of ATP levels in these mutants with MB which could result in a small increase in GTP-Ras1 (Fig 3D) indicating some increased electron flow might occur in the presence of MB in these C. albicans mitochondria similar to mammalian cells.
ATP production via respiration by the mitochondria is the main source of the chemical energy that fuels many different processes and pathways in the cell. Consequently, inhibition of respiration and ATP production will have a major impact on many aspects of cellular metabolism as shown by the metabolomics analysis in this study (Fig 3A, S1 Table, and S5 Fig). However, the strongly reduced ATP levels observed in the presence of MB did not block filamentous growth in the constitutively filamentous tup1/tup1 mutant or UME6-OE strain suggesting that MB is not inhibiting other parallel signaling or metabolic pathways important for induction or maintenance of filamentation (Fig 4C). Filamentation and wrinkled colony formation of these strains is not as robust as under control conditions, possibly due to reduction in other Ras1-controlled pathways or other effects of low ATP levels on growth dynamics that could be Ras1 independent, but hyphal growth is clearly evident in the strains even when MB is present. Furthermore, the NRG1-OE strain showed that filamentation is not required for higher ATP and elevated GTP-Ras1 under filamentation inducing conditions or the effects of MB on Ras1 signaling (Fig 4A and 4B). Under the conditions tested, the NRG1-OE strain formed some wrinkles, however, it does not form true hyphae. Microscopy of the cells showed mainly budding yeast with some elongated yeast cells and short pseudohyphae (S1A Fig panel 21). Interestingly, the occurrence of elongated yeast cells and pseudohyphae is inhibited by MB. Previous publications with the NRG1-OE strain looked at YPD +serum, which in our hands is not an as strong an inducer of filamentation and wrinkle formation as YNBAGNP on plates and the overexpression level of the NRG1-OE strain might just not be enough to overcome this stronger induction completely (Fig 1A and S1B Fig) [3,42]. Wrinkle formation of C. albicans colonies by MB and other respiratory inhibitors may support the model in which wrinkles promote usage of and demand for oxygen and thus the regulation of wrinkle production is downregulated upon respiratory inhibition [68]. In the bacterium Pseudomonas aeruginosa wrinkled colony morphology has been shown to be a redox-driven adaptation that maximizes oxygen accessibility and increased oxygen is able to inhibit wrinkle formation [69].
It is very interesting that the response of C. albicans to respiratory inhibition is similar to what has been seen in mammalian cells [53]. We observed a metabolome profile typical for AMP kinase activation, which has also been shown in mammalian cells exposed to MB [70]. This activation is not surprising as AMP kinase is a known energy sensor that measures ratios of ATP to ADP/AMP [55] that is activated when cellular energy status is low in eukaryotes. To increase ATP availability, the cells increase the catabolism of energy stores, such as fatty acids, and the biosynthesis of “costly” fatty acids and ergosterol is decreased (S1 Table, S5 Fig) [54,55,71]. Our data indicate that AMPK is needed to sustain levels of ATP, as intracellular ATP levels were very low in the snf4/snf4 strain (Fig 5D). We suspect this is the reason why GTP-Ras1 levels are so low and why this strain does not form filaments (Fig 5C). In agreement, a recent study showed that loss of only the kinase activity due to a point mutation in Snf1, an essential protein in C. albicans, inhibited the yeast-to-hyphal switch indicating how important AMPK activity is for energy homeostasis and filamentation [72]. Even though GTP-Ras1 and ATP levels are low in the snf4/snf4 strain, they are still responsive to MB showing that AMPK is not necessary for this signaling pathway. Overall, we believe that by understanding how C. albicans metabolism changes in different environments, we can use this fungus as an important probe for conditions within the host in states of health and disease.
In many cells, ATP concentrations control diverse processes such as autophagy [39], the retrograde response pathway [73], activation of neutrophils [74], and neurotransmitter responses [75]. ATP can act as an essential co-factor or can be recognized directly by binding to a receptor which triggers signaling. We do not yet know how ATP levels impact the Ras1 signaling cascade. One candidate ATP sensor is adenylate cyclase, Cyr1, itself. Cyr1 catalyzes the conversion of ATP to cAMP and thus has an ATP binding site that could function as a sensor of ATP levels. In S. cerevisiae, studies indicate that Cyr1 acts as a scaffold protein for Ras2 (homolog to Ras1 in C. albicans) interactions with Ira2 [76] and one could imagine that this scaffold activity may be regulated by ATP concentration. We know that Cyr1 signaling is required for the effects of MB in C. albicans as a catalytically inactive Cyr1, which can still serve some structural roles, was also insensitive to the Ras1-inhibiting effects of MB (Fig 8C). Indeed the cyr1/cyr1 mutant and the strain expressing catalytically inactive Cyr1 showed an increase of GTP-Ras1 with MB. Together with previously published data showing that Ras1 signaling is important for mitochondrial activity in S. cerevisiae, this might indicate that the mitochondria are not functioning normally in these strains and that like in mammalian cells and the cox4/cox4 and ndh51/ndh51 strains MB is able restore some electron flow to these dysfunctional mitochondria [43,77].
In S. cerevisiae, Ira2 has been shown to interact with the protein kinase A regulatory subunit, and Cyr1 and Ira2 have both been found at the plasma membrane and on mitochondria [78] providing further support for the potential for interactions between Cyr1 and Ira2, probably in ways that respond to mitochondrial activity [79] (Fig 9). These reports, with the data presented here, suggest that Cyr1 and Ras1 form a master regulatory circuit. Furthermore, the canonical Ras1 signaling pathway model has to be restructured from a pathway in which Cyr1 is just a factor downstream of Ras1 that is activated by GTP-bound Ras1 (Fig 1C) to a new model in which Cyr1 and Ras1 influence each other and together with Ira2 form a master-regulatory network necessary to coordinate the response to different environmental and intracellular signals in order to decide the fate of the cell (Fig 9).
These data reveal important aspects of the regulatory cascade that controls the C. albicans switch to a state more capable of causing host damage (Fig 9). Our findings indicate that the energy status of the cell is one of the most important signals involved in the decision of C. albicans to undergo the yeast-to-hyphae switch as it is able to override an array of filamentation inducing signals (Fig 9) Thus, host or host microbiome factors that impact energy levels will likely modulate C. albicans Ras1 signaling. Previous studies showed that Cyr1 can be directly activated by bicarbonate and muramyl dipeptides (MDPs) (Fig 9), though MDPs are only weak activators of filamentation in the absence of Ras1 [80,81] showing that Ras1 input is required for strong MDP-induced filamentation. Indeed, clinical data have linked the use of antibacterials to increased risk of C. albicans infections in multiple distinct body sites with very different bacterial community compositions [7–14]. In addition, numerous studies have shown that many bacteria inhibit C. albicans filamentation [13,14]. Future studies will focus on establishing whether repression in Ras1 activation, through modulation of ATP by competition with other microbes, contributes to the control of Candida behavior in a healthy mucosal microbial community. Furthermore, the future will show if known therapies or strategies can be used to favor benign host-Candida interactions by promoting low Ras1 activity.
All C. albicans strains were streaked from-80°C onto YPD (1% yeast extract, 2% peptone, 2% glucose) plates every 8–10 days and maintained at room temperature. All strains used in this study can be found in S2 Table. Overnight cultures were grown in 5 ml of YPD, supplemented with uridine as necessary, and washed in distilled water (dH2O) prior to use. C. albicans cells were mostly grown under filament-inducing conditions which included 37°C on YNBAGNP (1.5% agar, 0.67% yeast nitrogen base medium with ammonium sulfate (RPI Corp), 10 mM dextrose, 5 mM N-acetylglucosamine (GlcNAc), and 2% [wt/vol] casamino acids (BD Bacto), 25 mM potassium phosphate buffer). Cells were also grown on YPD + 5% fetal bovine serum when indicated. For yeast growth conditions C. albicans cells were grown at 30°C on YNBGP (1.5% agar, 0.67% yeast nitrogen base medium with ammonium sulfate (RPI Corp), 10 mM dextrose, 25 mM potassium phosphate buffer). For filamentation-inducing liquid growth conditions media were prepared as described above without the addition of agar and cells were incubated in the roller drum for 12 hours at 37°C.
Stock solutions were prepared of: methylene blue (MB) (Fisher Scientific)- 3 mM in dH2O; pyocyanin (PYO) (Cayman Chemicals)- 30 mM in 100% ethanol (EtOH); Antimycin A (AA) (Sigma)- 10 mM in 100% EtOH; oligomycin (Sigma) – 8 mg/ml in 100% EtOH; Dinitrophenol (DNP) (Sigma) – 100 mM in DMSO; menadione (Sigma) – 50 mM in 100% EtOH. All experiments were conducted in the dark to avoid light-induced ROS production.
The deletion mutant strains were constructed in the BWP17 strain background using a previously described method [58,82]. Briefly, gene-disruption cassettes for transformation were amplified using ~75 bp primers and the plasmids, pRS-ARG4 or pGEM-HIS1 [82] which contain ARG4 and HIS1 for PCR-directed integration. The forward primer was designed to have homology to 50 bp sequence upstream of the gene of interest start codon while the reverse primer had homology to the 50 bp sequence following the stop-codon. Both the primers were flanked by a 20 bp sequence homologous to the plasmids, as mentioned above. Sequential transformations of these gene-disruption cassettes into C. albicans BWP17 strain yielded the deletion strain. Plasmid pSM2 and pSMTC were used to complement the cyr1/cyr1 strain only with URA3 or with CYR1-URA3 at the URA3 locus [80]. The ira2/ira2 strain was reconstituted with URA3 using the pClp10 plasmid at the RP10 locus [83]. Strain ras1/ras1 + ras1 N-term was generated by transforming DH482 with PacI linearized pAP13+ras1 N-term and integration at the endogenous RAS1 locus was confirmed by PCR. To construct pAP13+ras1 N-term a PCR product encoding the first 161 residues of Ras1 was amplified from pAP14 [51] with primers RAS1XhoIF [51] and Ras1delta129BamHI-R, digested with XhoI and BamHI and ligated into similarly digested pAP13. All plasmids and primers used in this study can be found in S3 and S4 Tables.
For wrinkled colony formation, 10 μl from overnight cultures re-suspended in dH2O at an optical density (OD) of 8.0 were spotted onto YNBAGNP unless otherwise specified. The medium was supplemented with methylene blue (MB) from a 3 mM stock solution to a final concentration of 1.5 μM. 5 μM MB was used for the metabolomics experiment. Pyocyanin (PYO), Antimycin A (AA), and oligomycin (olig.) were added to the medium to a final concentration of 20 μM, 2.5 μM, and 7.5 μg/ml, respectively, or an equivalent volume of 100% ethanol (vehicle). Dinitrophenol (DNP) was added to the medium to a final concentration of 2 mM and menadione was added to the medium to a final concentration of 0.125 mM or an equivalent volume of vehicle solution. Cells were incubated at 37°C for 25 h.
Colonies were imaged after 24 h with a digital camera. Unless otherwise noted, all spot assays were completed as at least three independent replicates and a representative data set is shown. Cell morphology in colonies was assessed using a ZeissAxiovert inverted microscope equipped with a 100x long working distance objective and Axiovision software. To image the morphology of cells within the colony, the cells were resuspended in water, then applied to an agarose-coated slide to immobilize cells of different morphologies. The images shown were representative of the make up of the entire colony.
For western blot analysis spot colonies were scraped from agar plates after 24 h incubation at the conditions indicated, washed into a collection tube with dH2O and, after centrifugation, immediately snap-frozen in an ethanol/dry ice bath. Lysate preparation was conducted as previously published, with some modifications [51]. Whole-cell lysates were prepared by resuspending cells in Lysis/Binding/Wash Buffer (Active Ras Pull-Down and Detection Kit, Pierce) with protease inhibitors (Halt Protease Inhibitor Single-Use Cocktail, Pierce) and disrupting cells with glass beads in a Bio-Spec bead beater with six rounds of 50 seconds disruptions at 4°C and 1 minute rests on ice. Protein concentrations were determined by Bradford assay (BioRad).
Active or GTP-bound Ras1 was isolated utilizing the Active Ras Pull-Down and Detection Kit (Pierce) following the manufacturer’s instructions. In general, 200 μg of total protein were used for the pull-down unless otherwise specified. Due to the strong increase of GTP-Ras1 levels in ira2/ira2 strain only 100 μg of total protein were used (indicated in the figure). 12.5 μl of the pull-down samples containing active Ras1, and, for the input control, a total of 10 μg total protein diluted in SDS loading buffer were separated by SDS-PAGE, transferred to polyvinylidene difluoride (PVDF) with the Trans-Blot Turbo Transfer system (BioRad), and detected with monoclonal anti-Ras clone 10 (1.5 μg/ml; Millipore), followed by secondary detection with goat anti-mouse (Pierce) and enhanced chemiluminescent visualization (Pierce). As a control protein Pma1 was detected as described previously [52]. Densitometry analysis of Ras1 levels on Western blots was conducted with ImageJ [84].
Nanostring nCounter (Nanostring Technologies) analysis was used to quantify C. albicans gene expression. After 24 h spot colonies were harvested and fungal RNA was isolated using MasterPure Yeast RNA Purification Kit (Epicentre). Each Nanostring reaction mixture contained 80 ng fungal RNA, hybridization buffer, reporter and capture probes. Overnight hybridization of RNA with probes at 65°C preceded sample preparation using Nanostring prep station. Targets were counted on the nCounter using 255 fields of view per sample [85]. Raw counts for hyphal and yeast specific transcripts (HWP1, ECE1, HGC1, HYR1, ALS3, YWP1 and ALS4) were normalized within each sample to the geometric mean of two C. albicans housekeeping genes (ACT1, PMA1) and scaled to WT control conditions; the numerical average was taken from three biological replicates. Heat maps were developed using Z-scoring of Nanostring counts of selected yeast- and hyphal-specific genes using the “heatmap.2” function in the “gplots” package [86] in R (R Foundation for Statistical Computing, Vienna, Austria).
Spot assays were completed as previously described on YNBGNP agar plates and incubated at 37°C for 24 h. Cells were harvested, by scraping colonies from the surface of the agar using a coverslip, and then snap-frozen in an ethanol/dry ice bath. A total of 5 biological replicates were submitted to Metabolon for metabolite profiling, by GC/MS and LC/MS, of SC5314 wild type treated with vehicle (EtOH), 5 μM MB, 20 μM PYO, or 2.5 μM AA. All metabolites with mean values that had significant differences (p≤0.05) between treated and untreated samples were clustered into the 2 groups “UP” (upregulated ≥1.00-fold) or “DOWN” (downregulated <1.00-fold). VennMaster (http://sysbio.uni-ulm.de/?Software:VennMaster) [87] was used to determine the overlap of biochemicals that were either “UP” or “DOWN” in any of the treated samples. To visualize the result of this overlap analysis the tool eulerAPE (http://www.eulerdiagrams.org/eulerAPE) [88] was used to generate the Euler diagrams.
Spot assays were completed as previously described on YNBGNP agar plates with and without 1.5 μM MB or 7.5 μg/ml oligomycin and incubated at 37°C for 24 h. After harvesting by scraping colonies from the surface of the agar using a coverslip, the cells were disrupted with glass beads and 1x PBS in a Bio-Spec bead beater with 3 rounds of 60 seconds disruptions at 4°C and 1 minute rests on ice in between. A standard curve was prepared using Adenosine 5’-triphosphate disodium salt hydrate (Sigma). ATP levels were measured using the CellTiter-Glo Luminescent Cell Viability Assay (Promega) following the manufacturer’s instructions. The luminescent signal, which is proportional to ATP levels, was measured using a Tecan Infinite 200 Pro equipped with Magellan software (Tecan). All data were normalized to the protein concentration of each sample, which was determined using a Bradford Assay (BioRad). Three independent biological replicates, each including three technical replicates, were conducted and a representative data set is presented.
RAS1: C2_10210C_A; CYR1: C7_03070C_A; TPK1: C1_10220C_A; TPK2: C2_07210C_A; CDC25: C3_03890W; IRA2: C1_12450C_A; TFS1: C5_00930C_A; GPB1: C4_02150C_A; NDH51: C2_04550C_A; SDH1: C1_05260C_A; AOX1-A: C1_09160W_A; AOX1-B: C1_09150W_A; COX4: C2_01620W_A; SNF4: C6_03920W_A; NRG1: C7_04230W_A; UME6: C1_06280C_A; EFG1: CR_07890W_A; TUP1: C1_00060W_A; SSN3: C2_04260W_A
|
10.1371/journal.ppat.1007352 | Niche adaptation and viral transmission of human papillomaviruses from archaic hominins to modern humans | Recent discoveries on the origins of modern humans from multiple archaic hominin populations and the diversity of human papillomaviruses (HPVs) suggest a complex scenario of virus-host evolution. To evaluate the origin of HPV pathogenesis, we estimated the phylogeny, timing, and dispersal of HPV16 variants using a Bayesian Markov Chain Monte Carlo framework. To increase precision, we identified and characterized non-human primate papillomaviruses from New and Old World monkeys to set molecular clock models. We demonstrate specific host niche adaptation of primate papillomaviruses with subsequent coevolution with their primate hosts for at least 40 million years. Analyses of 212 HPV16 complete genomes and 3582 partial sequences estimated ancient divergence of HPV16 variants (between A and BCD lineages) from their most recent common ancestors around half a million years ago, roughly coinciding with the timing of the split between archaic Neanderthals and modern Homo sapiens, and nearly three times longer than divergence times of modern Homo sapiens. HPV16 A lineage variants were significantly underrepresented in present African populations, whereas the A sublineages were highly prevalent in European (A1-3) and Asian (A4) populations, indicative of viral sexual transmission from Neanderthals to modern non-African humans through multiple interbreeding events in the past 80 thousand years. Remarkably, the human leukocyte antigen B*07:02 and C*07:02 alleles associated with increased risk in cervix cancer represent introgressed regions from Neanderthals in present-day Eurasians. The archaic hominin-host-switch model was also supported by other HPV variants. Niche adaptation and virus-host codivergence appear to influence the pathogenesis of papillomaviruses.
| Epidemiologic studies have demonstrated that persistent infection of select oncogenic human papillomaviruses (HPVs) is the main cause of cervix precancer and cancer. Nevertheless, our knowledge of the underlying evolutionary mechanisms driving the divergence and emergence of viral oncogenicity in specific types of HPVs is incomplete. To better understand the molecular evolution of oncogenic HPVs, we isolated viruses from non-human primates, evaluated papillomavirus molecular clock models, and estimated the divergence times of HPV16 and other HPV type variants from their most recent common ancestors. Primate PV-host tissue tropisms indicated niche adaptation of viruses to host ecosystems as the first stage of the evolution of oncogenic HPVs. The data also provided evidence of ancient codivergence of HPV variants with archaic hominins and recent viral transmission from Neanderthals to modern non-African humans through sexual intercourse. Understanding the evolution of papillomaviruses should provide important biological insights and suggest mechanisms underlying HPV-induced cervical cancer, since niche adaptation rather than oncogenicity drives viral fitness.
| Papillomaviruses (PVs) are ubiquitous, non-enveloped, small double-stranded circular DNA viruses that cause proliferation of epithelial cells in a wide range of vertebrate host species, from reptiles to mammals [1, 2]. Currently, over 200 PVs infecting primate hosts (human and non-human) have been characterized and shown to group predominantly within 3 highly divergent genera—Alphapapillomavirus, Betapapillomavirus, and Gammapapillomavirus [3]. All oncogenic PVs associated with the development of cervical carcinoma, including human PV (HPV) types 16, 18, 31, 33, 35, 39, 45, 51, 52, 56, 58, and 59 and Macaca fascicularis PV type 3 (MfPV3), share a common ancestor within the Alphapapillomavirus [4–7]. Among these oncogenic types, which are sexually transmitted primarily through intercourse [8, 9], HPV16 is globally the most prevalent HPV type detected, suggesting an increased fitness [10–12]. Moreover, HPV16 is also the most common HPV type detected in cervical cancer, which is the fourth most common cancer among women worldwide [13]. Nevertheless, most exposures to HPV types are transient, and many PVs appear to be more commensal than pathogenic [14].
Strict coevolution of a host and its pathogen is more likely if the pathogen is transmitted vertically and there is little or no cross-species acquisition. Persistent infection by pathogens generally indicates that they are well adapted to their host and that extinction will be rare so long as the host survives. Hence, in scenarios of coevolution, the evolutionary history of a pathogen should mirror that of its host, both in divergence times and phylogenic history (Fahrenholz’s rule) [15, 16]. These criteria have been shown to hold for feline PVs within the genus Lambdapapillomavirus isolated from oral lesions [17]. On the other hand, horizontal transmission of pathogens through host switching without restricted species specificity will produce a very different evolutionary history between host and pathogen. In hosts harboring many different types of PVs (e.g., bovines, humans, and macaques), the selection pressure exerted by PVs on their hosts appears negligible in comparison with what the hosts exert on the PV pathogens. Within human populations, for example, the ancient dispersal of HPV variants (e.g., HPV16 and HPV58) challenges a simple evolutionary pattern of viruses migrating with modern Homo sapiens [18], and instead indicates codivergence of viruses with archaic hominins and transmission to modern humans [19, 20]. The genetic heterogeneity of PVs implies a complex evolutionary history with many interacting factors, including but not limited to virus-host codivergence, tissue tropism, lineage sorting, transmission, recombination, and natural selection [21, 22]. Understanding the capacity for, and history of, viral adaptation to host ecological environments is essential for understanding the genetic basis of HPV carcinogenicity [23]. However, the origin and evolution of oncogenic PVs remains poorly understood.
In this report, we estimate the divergence times of HPV16 and other oncogenic HPV types using a well-established Bayesian molecular clock model with newly characterized primate PV genomes that validate the divergence times of primate HPVs within niche-specific clades. Our analyses of the evolutionary dynamics of primate PVs, including specific focus on HPV16 variants, provide novel insights into the complex phylodynamic interactions between viruses and hosts and their pathologic outcomes.
In an effort to study the diversity of non-human primate PVs (NHP-PVs) to better understand the evolution of oncogenic HPVs, we screened cervicovaginal specimens from 10 adult female squirrel monkeys (Saimiri sciureus), and the paired oral, perianal, and genital samples from 8 adult rhesus monkeys (Macaca mulatta) (4 females and 4 males). Three novel Saimiri sciureus PV types (SscPV1, 2 and 3) and three novel Macaca mulatta PV types (MmPV2, 3 and 4) were isolated and characterized and had genomes ranging in size from 7424 bp to 8051 bp (S1 Table). All genomes contained five early genes (E6, E7, E1, E2, and E4), two late genes (L2 and L1), and an upstream regulatory region (URR) between L1 and E6 genes. Phylogenetic trees based on the nucleotide sequence alignment of the concatenated four open reading frames (ORFs) (E1, E2, L2, and L1) (Fig 1 and S1 Fig) or individual genes, e. g., E1 or L1 ORFs (S2 Fig, S3 Fig and S4 Fig) support a monophyletic clade grouping SscPV1/2/3 and howler monkey Alouatta guariba PV 1 (AgPV1, KP861980) [24] within the genus Dyoomikronpapillomavirus. MmPV2 and MmPV3 cluster into the genus Alphapapillomavirus, with the closest HPVs being HPV54 (within the species Alpha-13) and HPV117 (within the species Alpha-2), respectively. MmPV4 shares <70% of L1 ORF similarity with members of the species Gamma-10 (e.g., HPV121 and HPV130) and may represent a novel species within the genus Gammapapillomavirus.
We focused on HPV16 because it is the most prevalent and potent carcinogen among the oncogenic HPVs [5]. To interrogate HPV16 evolution using a molecular clock, we utilized HPVs and NHP-PVs characterized in our labs and by others where the host species separation times have been well established [25, 26]. This step is essential in order to validate a vertical mutation rate model suitable for HPV variants. This model estimates the mutation rate for infectious PVs over long periods of time and might differ from horizontal mutation rates not measured in this study.
Papillomaviruses have been identified in a wide range of NHP species, including Old World monkeys and apes (e.g., macaque, chimpanzee) and New World monkeys (e.g., squirrel monkey, brown howler) [24, 27–33]. Using a maximal likelihood algorithm and a nucleotide sequence alignment of the concatenated E1-E2-L2-L1 ORFs for 141 PV types representing each species or unique host (S2 Table), we found that the majority of primate PVs phylogenetically clustered into Alphapapillomavirus, Betapapillomavirus, or Gammapapillomavirus genera, corresponding predominantly to the anatomical sites where the viruses were originally isolated (e.g., mucosal or cutaneous epithelium), which was independent of the host species (Fig 1, S1 Fig and S2 Table). For example, MmPV1 is a rhesus macaque PV type (within the species Alpha-12) isolated from cervicovaginal cells that shares a most recent common ancestor (MRCA) with oncogenic mucosal HPV16 (within the species Alpha-9) but is distantly related to MmPV4 (within the genus Gammapapillomavirus), which was also isolated from a rhesus macaque. Since topological incongruence has been noted in the phylogenies of HPVs when trees are constructed with either late or early regions of the viral genomes [22, 34], we also examined the topologies of such trees. Although there was some incongruence, the majority of the primate PVs maintained their topological positions (see S2 Fig, S3 Fig and S4 Fig).
Fahrenholz’s proposal for strict codivergence of host and parasites states that the “parasite phylogeny mirrors that of its host,” indicating that specific pathogens isolated from an individual host species should be monophyletic to the exclusion of viruses from other host species (reviewed in de Vienne et al.) [35] (Fig 2). In the case of primate PVs, however, viruses infecting a given host species do not always cluster together, implying an ancient viral divergence model in which viral ancestors may have first split into separated viral clades corresponding to niche adaptation to specific host ecosystems (i.e., tissue tropism). Following host ancestor speciation, distinct but homophyletic viruses were transmitted to similar ecosystems (e.g., mucosal or cutaneous sites) between closely related host animals, resulting in the radiation observed in the extant primate PV tree where viruses sort by tissue tropism and not host species. This prediction was evaluated with a permutational multivariate analysis of variance (PERMANOVA) test [36] using primate PV nucleotide sequence pairwise distances, which revealed that tissue tropism (here defined by different genera) contributed to more of the variability of viral divergence (accounting for 26% of the total variance, p<0.001) than that of the host (6%, p<0.001) (Table 1).
To estimate the divergence times of primate PVs from their MRCAs, we used a Bayesian statistical framework employing previously established PV evolution rates [17]. Infectious PVs have been shown to have a slow mutation rate based on the observations that these double-stranded DNA viruses use the host cell DNA replication machinery, characterized by high fidelity, proofreading capacity, and post-replication repair mechanisms [37]. Since primate PVs, taken together, do not follow strict viral-host codivergence, each genus was evaluated separately to estimate divergence times. A combination of relaxed lognormal molecular clock and coalescent constant population models provided the best performance using the phylogenetic tree as shown in Fig 3A. The Alphapapillomavirus–Dyoomikronpapillomavirus split from a MRCA around 39.9 million years ago (mya) (95% highest posterior density (HPD), 36.4–43.7 mya) (Fig 3A, S2 Fig and Table 2) is consistent with the time frame of the split between New World and Old World primate ancestors [26].
Similar virus-host codivergence events were observed between Old World monkey PVs and their closest HPV relatives, and were estimated to approximately 14–31 mya (Fig 3, S5 Fig, S6 Fig and S7 Fig). For example, the species Alpha-12 (PVs mainly isolated from genital lesions of macaques) split from a MRCA with the species Alpha-9 (represented by oncogenic genital HPV16) around 27 mya coincided with the time span of the speciation between macaques and apes/humans that occurred approximately 25 mya [38, 39]. An enigmatic observation in these data is the clustering of macaque PVs (e.g., MfPV3) and baboon PV (Papio hamadryas PV 1, PhPV1) within the species Alpha-12 group, suggesting either a recent viral transmission between macaque and baboon monkeys, or a more complex phylogeny of the sub-family Cercopithecinae. The majority of distinct human PV types arose during the end of the Miocene and/or the beginning of the Pliocene epoch coincident with the divergence of humans and chimpanzees occurring around 6–8 mya (Fig 3) [40].
The divergence times and tree topologies support a model of intrahost divergence of primate PVs in which ancient viruses diverged and adapted to specific host ecosystems (e.g., tissue tropism or different types of epithelial cells) within an ancestral host animal lineage (e.g., the MRCA of primate animals) (Fig 4). Following periods of host speciation, continuing intrahost viral divergence events occurred as distinct but phylogenetically related viral types were transmitted to similar host ecosystems by the closely related host animals. This pattern of ancient viral divergence coupled to niche adaptation may explain, for example, the differences in the prevalence of HPV16 and HPV18 between squamous cell carcinomas and adenocarcinomas of the cervix [41]. This difference might represent the emergence of further viral adaptation to different ecological niches within the cervix, one dominated by stratified squamous epithelium the other by columnar epithelium, respectively [42]. The fact that we do not observe similar or parallel diversity of NHP-PVs compared to HPVs (broken lines in right panel of Fig 4B) could be due, in part, to reduced sampling effort, limited population size of NHPs, bottlenecks of viral transmission, and/or restricted host migration.
Next, we constructed a phylogenetic tree of HPV16 variants based on 212 complete genomes to classify variant lineages and sublineages (S3 Table). The tree topology shows two deeply separated clades corresponding to the previously classified Eurasian and African lineages (S8 Fig), with a mean nucleotide sequence difference of 1.72% ± 0.09% (S4 Table). The African lineage variants were more than twice as diverse (intragroup mean difference of 0.77% ± 0.04%) as the Eurasian variants (0.32% ± 0.02%). Since geographic nomenclature systems suffer from sampling biases and preconceived notions about virus ancestry, we utilized an agnostic alphanumeric nomenclature based on HPV16 phylogeny and complete genome nucleotide differences to assign HPV16 variants into four lineages designated A, B, C, and D. Each lineage could be divided into four sublineages (A1-4, B1-4, C1-4, and D1-4), based on previously described criteria (S9 Fig) [43]. The previously named Asian (As) and North American 1 (NA1) variants are designated sublineages A4 and D1, respectively [44]. The maximum pairwise difference between the most diverse isolates, from sublineages A1 and D3, was 2.23%.
Based on single-nucleotide polymorphism (SNP) patterns and phylogenetic tree topologies, we assigned 3256 HPV16 partial sequences from 22 countries/studies into variant lineages and sublineages using maximum likelihood methods (Table 3). As shown in the summarized charts of HPV16 phylogeography (Fig 5A), isolates from Asians and Caucasians (Australians/Europeans, and North Americans) were predominantly represented by A variants, with abundances of 92% and 83%, respectively. The majority of A4 variants (352/357, 99%) were from Asian individuals. Within the African population, 90% of HPV16 infections were B and C lineages. HPV16 variants in South/Central Americans were equally assigned as A1-3 (50%) and D (48%). Using a weighted UniFrac algorithm, variants were well clustered into groups (African, Eurasian, and South/Central American) corresponding to the geographic origin of the isolates (Fig 5B). Globally, A1-3 sublineages were the most widespread; whereas, the D lineages were detectable at low prevalences in many populations outside of South/Central Americans, such as in Caucasian (11%), African (7%), and Asian (6%) individuals (Fig 5C). In contrast, A4 and B/C lineages were rarely found outside of Asian and African populations, respectively.
The molecular clock models used to estimate the divergence times of primate PVs support a scenario of virus-host codivergence after the virus has adapted to a specific host ecosystem. Using a similar Bayesian Markov chain Monte Carlo (MCMC) framework, we initially applied six combinations of clock models to estimate the divergence of HPV16 variants from their MRCA, without any prior assumption of virus-host codivergence (Table 4, no calibration). Interestingly, a combination of the relaxed lognormal molecular clock and coalescent Bayesian skyline models indicated that HPV16 A and BCD had divided around 618.5 thousand years ago (kya) (95% HPD: 331.5–996.1). This estimation is within the time span of the separation between Homo sapiens and archaic hominins (e.g., Neanderthal/Denisova) but around two-five times longer than the estimated modern Homo sapiens divergence time (ca. 150–200 kya) [45] indicative of an ancient divergence of HPV16 variants prior to the emergence of modern human ancestors. Based on the geographic distribution of HPV16 variants above, we then used an archaic hominin-host-switch (HHS) scenario to calibrate the divergence time between HPV16 A and non-A variants (500 kya, 95% HPD: 400–600), and a modern-out-of-Africa (MOA) scenario between BC and D variants (90 kya, 95% HPD: 60–120). When time calibrations were introduced into the phylogenetic tree, the HHS scenario showed the strongest support for time inference and estimated an initial divergence of HPV16 variants at approximately 489 kya (95% HPD: 394–581), predating the out-of-Africa migration of modern humans (ca. 60–120 kya) (Fig 6 and S10 Fig) [46, 47]. In addition, the demographic model of the Bayesian skyline plot for the population function through time showed a recent exponential expansion of the effective population size of present-day HPV16 occurring in the last 25 kya, lagging behind the growth of modern human populations (starting from the last 40–50 kya) (see the top panel of Fig 6). This plot most likely reflects the concurring increase and mobility of modern human populations and present-day virus populations in the last epoch.
We observed a similar divergence timeframe for other HPV variants, splitting from their MRCAs approximately 300–600 kya and showing a strong correlation between evolution times and genomic diversities (Fig 7, Table 5). In all cases, the deep separation between HPV16 variant lineages A and BCD (and the deepest lineage separations of other HPV variants) suggests an ancient virus-host codivergence, coinciding with the split between archaic Neanderthal/Denisova and modern human ancestors from their MRCA (Fig 8). Neanderthals spread out over Eurasia with at least two populations splitting approximately 77–114 kya from each other based on analysis of archaic genomes from Vindija, Mezmaiskaya (Caucasus), and Denisova (Siberia) [48]. This time period corresponds to the diversion of HPV16 A sublineages and in particular the split of A4 from A1/2/3 and the emergence of HPV16 A4 in Asia, likely representing independent transmission of A4 from archaic hominins to modern humans in the east.
In this work, we used a Bayesian MCMC framework to estimate the divergence times of primate PVs and propose an early ancient intrahost viral divergence model (i.e., niche adaptation) followed by viral-host coevolution. This form of viral evolution has been documented for polyomaviruses [49], herpesviruses [50], and some retrovirus genera [51]. With the assumption of host niche adaptation as a fundamental process, the estimation of primate PV divergence times within niche-specific clades mirrors that of the primate host evolutionary history (Fig 4). It is clear that the evolutionary history of these well adapted, slowly evolving PVs may be significantly more complex than previously appreciated [37]. The implication of host niche adaptation of primate PVs preceding virus-host codivergence suggests a critical role for viral genetic heterogeneity and natural selection. The origin of viral genetic determinants of cervical niche adaptation further supports the hypothesis that a group of well-evolved viral genotypes also contain the determinants for cervical cancer, since this phenotype cannot exert selective pressure, as it does not support the production of infectious virus. It may also explain why a large set of cervicovaginal macaque PVs (within the species Alpha-12) associated with cervical neoplasia shares a common origin with the high-risk clade of human PVs (e.g., Alpha-9) (Fig 3A) [6, 27]. Our findings provide a framework for studying the past evolution of primate PVs infecting the genital tract niche and support a molecular clock based on phylogeny, since the generation time of PVs can only be extrapolated from empiric data based on coevolution models [17, 52].
We used this well-supported molecular clock model to estimate the divergence times of HPV16 variants. HPV16 is the most common oncogenic HPV type and shows diversity in persistence and carcinogenicity [53–55], suggesting further biological differences between variant lineages. We observed specific geographic/ethnic dispersals of HPV16 variants, such as A4 predominance in Asian populations and BC predominance in African populations. The estimated divergence times between HPV16 A and BCD variants largely predated that of the out-of-Africa migration of modern human populations, consistent with a previously reported archaic hominin-host-switch scenario [19, 20]. One interpretation of the data implies that the present-day Eurasian HPV16 A variants were probably the products of multiple interactions between Neanderthals/Denisovans and modern Homo sapiens established during sexual contact after a long period of separation (e.g., 400–600 kya). This notion of viral sexual transmission between groups is reflected in the recent genetic admixture (e.g., 80 kya) between groups [48, 56–59], with evidence of 2–4% of nuclear DNA in Eurasians that can be traced to Neanderthals [48, 58]. This assumption is likely ubiquitous in a number of Alpha-HPV variants (Fig 7, Table 5), although their pathogenesis, evolution, and epidemiology warrant further study.
Recent evidence indicates that Neanderthals spread out over the Eurasian continent and also admixed with ancestors of the present-day East Asian population [60, 61]. Since HPV16 A4 lineage is exclusively found in East Asians (approximately 40% of HPV16) and presents a higher risk of cervix cancers in Asian populations [62, 63], we speculate that a subset of Neanderthals heading east into Asia over more than 100 thousand years of existence in Eurasia could have interbred with East Asian modern humans and transmitted the HPV16 A4 sublineage and introgressed specific gene alleles that provided a selective advantage to the HPV variants coevolving with them [59, 64]. Overall, HPV16 BCD variants have higher genomic diversity than A isolates (see S4 Table), which may imply a potential population bottleneck of horizontal transmission reducing the diversity of current day A lineage isolates. In contrast, BCD variants have accumulated more genetic mutations, consistent with the observations that African populations and their pathogens have deeper origins reflected in greater diversity [65]. This idea supports one theory that both HPV16 BCD and modern humans arose in Africa (Fig 8). Following a relatively recent out-of-Africa migration, the modern humans acquired the A variant from sex with archaic hominins and possibly carried D variants into Eurasia under conditions of a small population size. The ancestors of East Asian people crossed the Bering Strait and were early populators of the Americas (based on historical records and genetic relatedness) [66]. Surprisingly, the D lineage is phylogenetically rooted in the African clade, but we did not find a major reservoir of the D lineage in the present-day African populations. This interesting observation suggests either an advantage of niche colonization and expansion of HPV16 D variants in Native Americans or a bottleneck of HPV16 variants present in people populating the Americans. Alternatively, the lack of A4 and the high proportion of D lineages in the Americans could be the result of an early colonization of the Americas by an unknown group from Africa. More data is needed to sort out the evolutionary history of the HPV16 D lineage and might provide clues to new features of the populating of the Americas.
Sexual interactions between archaic hominins and modern human ancestors likely occurred over multiple time- and space-scales. For example, viral transmission might have also occurred from modern humans to Neanderthals/Denisovans, based on the evidence of ancient gene flow from early modern humans into Eastern Neanderthals [57]. Since PVs usually establish infections at the basal layer of epithelial cells, it will be impossible to detect viruses from fossil bones of archaic hominins and document the presence of HPVs in archaic hominin populations [20]. The evolutionary histories and origins of modern H. sapiens are undergoing dramatic revisions with the introduction of advanced sequencing techniques and methods to analyze genomic samples from archaic hominin specimens [67–69]. Since the reproductive success per copulation between H. sapiens and archaic hominins is predicted to have lower viability than that of modern human reproductive events, high levels of sexual interaction were likely present facilitating HPV transmission, in addition to genetic introgression observed in modern non-African populations [70]. For example, the human leukocyte antigen (HLA) B*07:02 and C*07:02 alleles associated with increased risk in cervix cancers appear to be introgressed regions in present-day Eurasians and Melanesians from Neanderthals or Denisovans [71–73]. This also suggests that adaptive introgression of modern humans from archaic hominins influences the pathogenic outcome of these infections by as yet unknown mechanisms [70, 74]. However, it can be speculated that introgressed genes providing some selective advantage to hybrid human-archaic hominin offsprings could also make them more susceptible to HPV variants adapted to archaic hominins over hundreds of thousands of years of coevolution. The introgressed genes are most likely related to immunity against infections, whatever the pathogens might be and HPV was along for the ride, since HPV is not known to affect reproductive fitness of the host.
This study has its strengths and limitations. We expand the current understanding of HPV16 evolution beyond the recent description of HPV transmission between archaic and modern humans that used existing data [20] in important ways. We have expanded the understanding of HPV16 in the context of human and non-human primate PV evolution by characterizing additional New World and Old World monkey PVs and using the known divergence times of specific primate species to establish a valid molecular clock. This approach was used to establish the times of Neanderthal divergences [48]. We demonstrate that niche adaption had to proceed viral-host coevolution, and suggest that subsequent niche adaptation might underlie the difference in prevalence of HPV16 and HPV18 in cervical squamous and glandular lesions. We have identified and characterized additional HPV16 variants enabling us to establish the HPV16 variant taxonomy that includes subvariants that have unique biological characteristics [53]. Moreover, we propose that evolution of HPV16 A in Neanderthals over time led to allopatric emergence of the HPV16 A4 lineage as Neanderthals moved east and interbred with modern humans in Asia. We have also expanded the number of HPV16 isolates from around the world to establish the global distribution of HPV16 variants. Lastly, we provide new interpretations and questions on the HPV16 D lineage that is part of the African clade, but is highly prevalent in South/Central America. Nevertheless, there are also limitations to the current study and interpretations. The understanding of human evolution is constantly being challenged with new data and it is possible the models of human evolution used in this study will change [75]. We have not sampled every population and it is possible that additional HPV16 isolate data could change our interpretations. The data obtained on the geographic locations of the HPV partial sequences could be incorrect resulting in underestimating the true associations between variants and historic origins. Lastly, it is possible that very low population sizes of humans migrating out of Africa carried HPV16 A lineage variants leaving no traces in Africa, but expanding throughout Eurasia. This unlikely possibility would influence the interpretations of both our work and that of previous studies analyzing the evolution of HPV16 [20].
In conclusion, the biology and natural life cycle of oncogenic HPVs that results in infectious viral particles (i.e., vegetative virus life cycle) is highly adapted to the differentiation program of epithelial cells [76]. Poorly differentiated precancerous and cancerous cells in the cervix do not support the HPV vegetative life cycle, and thus viral-associated transformation does not contribute to the fitness of HPVs. Viral phenotypes that serve to adapt to a specific ecological niche, evade host immune mechanisms, and support persistent viral production, however, should contribute to viral fitness. Therefore, further investigations of viral-host interactions and the underlying mechanisms of viral oncogenicity, should continue to focus on features of viral evolution and niche adaptation that contribute to fitness, since the oncogenic outcome of HPV infections appear to be “collateral damage” affecting host morbidity and mortality. The current data provides a framework to unravel the mysteries of oncogenic HPV genomes as we expand our understanding of viral-host evolution.
The studies providing human cellular samples have been approved by the Institution Review Board of the Albert Einstein College of Medicine, Bronx, NY, and the Joint Chinese University of Hong Kong-New Territories East Cluster Clinical Research Ethics Committee. All human subjects were older than 18 years of age and samples were anonymized without individual identifying information. Written informed consent was obtained from each participant.
The animal use protocol was reviewed and approved by the Institutional Animal Care and Use Committee (IACUC) of Albert Einstein College of Medicine (protocol number 20060908). All procedures involving animals were conducted in compliance with applicable state and federal laws, guidelines established by the Animal Care and Use Committees of the respective institutions, and standards of the U.S. Department of Health and Human Services, including the National Institutes of Health Guide for the Care and Use of Laboratory Animals. The programs for animal care and welfare at Albert Einstein College of Medicine has been fully accredited by the Association for Assessment and Accreditation of Laboratory Animal Care (AAALAC). The Animal Welfare Assurance (A3312-01) is on file with the Office for Laboratory Animal Welfare.
The Saimiri sciureus PV DNA was isolated from exfoliated cervical cells of two adult female squirrel monkeys screened using polymerase chain reaction (PCR)-based MY09/11 and FAP59/64 primer systems [77, 78]. Sequences from the PCR products were compared with a PV database maintained in the Burk lab using a Blastn search and shown to have < 90% similarities to previously characterized PV types. The whole genomes were PCR-amplified as two overlapping fragments using degenerate primer sets designed on available L1 gene sequences and consensus E1 alignments, and subsequently Sanger sequenced using primer walking in the Einstein Sequencing Facility, New York [33]. Geneious R9.1.7 was used to assemble segmented sequences into the complete genome sequences and identify ORFs [79].
The Macaca mulatta PV DNA was purified from exfoliated cervical cells of one adult female rhesus monkey and swabs from the penis surface of one adult male rhesus monkey. The viral DNA was initially detected using multiplexed next-generation sequencing (NGS) assays targeting two small fragments (136 bp and 83 bp, respectively) within the L1 ORF [80, 81]. Sequences of a Blastn search against a PV database showed < 90% similarities to characterized PV types. The total DNA underwent a metagenomic sequencing on an Illumina HiSeq4000 at Weill Cornell Medicine Genomics Resources Core Facility, New York, using paired-end 100 bp reads. The short reads were filtered for host genome contamination and assembled de novo using Megahit v1.0.6 to build contigs with long length [82]. The whole genomes of novel Macaca mulatta PVs were validated using type-specific PCR in three overlapping fragments and Sanger sequencing using a primer walking strategy.
The complete genome sequences of SscPV1/2/3 and MmPV2/3/4 have been submitted to NCBI/GenBank database, with access numbers of JF304765 to JF304767 and MG837557 to MG837559, respectively.
In our previous work, we sequenced the complete genomes of 78 HPV16 isolates (see HPV16 list in S3 Table) [83, 84]. In the current study, 122 cervicovaginal samples containing HPV16 DNA were randomly chosen from the Kaiser Permanente Northern California (KPNC)-NCI HPV Persistence and Progression (PaP) cohort study [85] and a population-based HPV prevalence survey coordinated by the International Agency for Research on Cancer (IARC) [63]. The complete genomes were characterized using nested overlapping PCR and Sanger sequencing as previously reported [86]. The PaP study samples were also sequenced using Ion PGM platform [87]. In addition, 12 HPV16 complete genomes sequenced by others were included in this study [88–92].
To evaluate the phylogenetic relationships of PVs, the concatenated nucleotide sequences of four open reading frames (ORFs) of the E1, E2, L2, and L1 genes of 141 PV types representing 132 species and unique hosts were used (see PV list in S2 Table, column labelled “Selected type” marked yes). Because all known PVs contain these four core ORFs, the concatenated partitions provide a comprehensive evaluation of the evolutionary history of Papillomaviridae. In addition, the highly conserved E1 early gene and L1 late gene were used to characterize phylogenetic incongruence. The nucleotide sequences of each coding region were aligned based on the corresponding amino acid sequences previously aligned using MUSCLE v3.8.31 [93] in Geneious R9.1.7. For HPV16 lineage/sublineage classification and phylogenetic analyses, all 212 complete genome nucleotide sequences (see HPV16 list in S3 Table) were linearized at the ATG of the E1 ORF and aligned using MAFFT v7.221 [94].
Maximum likelihood (ML) trees were constructed using RAxML MPI v8.2.3 [95] and PhyML MPI v3.1 [96] with optimized parameters based on the aligned complete genome nucleotide sequences. Data were bootstrap resampled 1,000 times in RAxML and PhyML. MrBayes v3.1.2 [97] with 10,000,000 cycles for the Markov chain Monte Carlo (MCMC) algorithm was used to generate Bayesian trees. A 10% discarded burn-in was set to eliminate iterations at the beginning of the MCMC run. The average standard deviation of split frequencies was checked to confirm the independent analyses approach stationarity when the convergence diagnostic approached <0.001 as runs converge. For Bayesian tree construction, the computer program ModelTest v3.7 [98] was used to identify the best evolutionary model; the identified General Time Reversible (GTR) model was set for among-site rate variation and allowed substitution rates of aligned sequences to be different. The CIPRES Science Gateway [99] was accessed to facilitate RAxML and MrBayes high-performance computation.
Permutational multivariate analysis of variance was performed using the adonis function in R’s package ‘vegan’ and the pairwise distance based on 220 primate papillomavirus E1-E2-L2-L1 nucleotide sequences (S2 Table).
A dataset of 3256 partial sequences spanning variable genes/regions of HPV16 was obtained from GenBank that included the geographic source of the sequences mainly from indigenous ethnicities and/or local communities including 22 countries/regions throughout the world. These included, in Africa: Burkina Faso [100], Nigeria [101], Rwanda [102], Uganda [103], and Zambia [104]; in Asia: China [105–107], India [108, 109], Japan [110], Korea [111], and Thailand [112, 113]; in Europe: Germany [114], Italy [115–118], Netherland [119, 120], Portugal [121], Russian [122], Spain [123], and United Kingdom [124]; in North America: Canada (GenBank, see details in Table 3), Costa Rica [9]; in South/Central America: Bazile [125–127] and Mexico [128–131]; and Australia [132] (see Table 3). We used a maximum phylogenetic likelihood algorithm in pplacer v1.1.alpha17 [133] to place partial sequences on a reference tree inferred from an alignment composed of the 212 HPV16 variant complete genomes described in this study. A cutoff value of maximum likelihood ≥ 0.8 was set as confident assignment of HPV16 isolates into lineages and sublineages. The abundance of each lineage from the same country was combined and normalized using a percentage. According to the geographic patterns of HPV16 variants [44], four ethnical groups, namely African, Asian, Caucasian, and South/Central American, were summarized; for each HPV16 (sub)lineage, its frequency in each group was calculated based on the summary of individual percent abundance divided by the summary of total percent abundance. We used a weighted UniFrac method in R’s package ‘GUniFrac’ [134] to calculate the pairwise distances between geographic locations, based on which a principle component analysis (PCoA) was performed to visualize the clustering of geographic groups of HPV16 variants using the betadisper function in R’s package ‘vegan’.
We used a Bayesian Markov Chain Monte Carlo (MCMC) method implemented by BEAST v2.4.5 [135] and the previously published PV evolutionary rates [17] to estimate the divergence times of primate PVs from their most recent common ancestors (MRCAs). Times were calculated separately for Alphapapillomavirus (n = 85), Betapapillomavirus (n = 54), and Gammapapillomavirus (n = 81) (S2 Table), given that primate PVs, taken together, do not follow strict virus-host codivergence. Three tree priors were estimated using the following demographic models: (1) coalescent constant population, (2) Yule model, and (3) coalescent Bayesian skyline, with assumptions that (1) the PV genome has a strict mutation rate or (2) there is an uncorrelated lognormal distribution (UCLD) molecular clock model of rate variation among branches, resulting in six combinations of models. In addition, we chose the GTR sequence revolution model with the gamma-distributed rate heterogeneity among sites and a proportion of invariant sites (GTR + G + I) determined by the best-fit model approach of Modeltest v3.7 [98]. The concatenated nucleotide sequence partitions of six ORFs (E6, E7, E1, E2, L2, and L1) with variable rates of substitution over time were used: 2.39 × 10−8 (95% confidence interval 1.70–3.26 × 10−8) substitutions per site per year for the E6 gene, 1.44 × 10−8 (0.97–2.00 × 10−8) for the E7 gene, 1.76 × 10−8 (95% CI: 1.20–2.31 × 10−8) for the E1 gene, 2.11 × 10−8 (95% CI: 1.52–2.81 × 10−8) for the E2 gene, 2.13 × 10−8 (95% CI: 1.46–2.76 × 10−8) for the L2 gene, and 1.84 × 10−8 (95% CI: 1.27–2.35 × 10−8) for the L1 gene, as previously described [17]. In order to calibrate the divergence times, we introduced three time points inside and at the root of the Alphapapillomavirus tree, with assumptions of codivergence histories between primate PVs and their hosts: (1) the node between HPV13 and chimpanzee PpPV1 (Pan paniscus PV 1) at 7 mya (95% CI, 6–8 mya) matching the split between hominin and chimpanzee ancestors; (2) the node between the species Alpha-12 (represented by Macaca mulatta PV 1) and Alpha-9/11 (represented by HPV16) at 28 mya (25–31 mya) matching the speciation between hominin and macaque ancestors; and (3) the node between Alphapapillomavirus and Dyoomikronpapillomavirus (represented by Saimiri sciureus PV 1) at 49 mya (41–58 mya) matching the divergence between Old World and New World monkey ancestors [26]. For Betapapillomavirus and Gammapapillomavirus trees, the calibration time point(s) was set between macaque PVs and their closet HPV relatives.
To estimate divergence times of HPV16 complete genome variants, a Hominin-host-switch (HHS) model assuming there was an ancestral viral transmission between archaic and modern human populations [20] was applied by setting two evolutionary time points to calibrate the HPV16 variant phylogenetic tree: (1) the archaic divergence of modern humans and Neanderthals/Denisovans around 500 thousand years ago (kya) (95% CI, 400–600 kya) [136] matching the split between HPV16 Eurasian (A) and African variants (BCD), and (2) the modern human out-of-Africa migration at 90 kya (95% CI, 60–120 kya) [45, 137], locating the era when HPV16 D variants diverged from their most recent common ancestor (MRCA). A HPV16 variant substitution rate was used for validation of a uniform prior rate: 1.84 x 10−8 (95% CI, 1.43–2.21 x 10−8) [20], with combinations of three tree priors and two clock models as described above. Due to the lack of geographic/ethnic dispersal information of other HPV type variants, we estimated the youngest divergence events splitting from their MRCA using complete genome alignments and HPV16 variant substitution rate without time point calibration.
To compare the population dynamics of HPV16 variants and the modern human host, Bayesian skyline plots were created using BEAST. A total of 311 globally sampled present-day human mitochondrial DNA (mtDNA) sequences, excluding the 1120 bp non-coding D-loop (that evolves at a higher rate) to give an alignment of 15,471 bp in length [138], were analyzed using a strict clock model and a coalescent Bayesian skyline, with an estimated rate of 2.47 x 10−8 (95% CI, 2.16–3.16 x 10−8) substitutions per site per year [139], as these sequences have been shown to evolve in a roughly clock-like manner [140, 141]. Two evolutionary time points were used to calibrate the modern human mtDNA tree: (1) the age of the MRCA between the maximum distanced modern humans, estimated to be 171,500 ± 50,000 years ago, and (2) the age of the MRCA of the youngest clade that contains both African and non-African lineages, approximately 52,000 ± 27,500 years ago [140].
The MCMC analysis was run for 100,000,000 steps, with subsampling every 10,000 generations. A discarded burn-in of the first 10% steps was set to refine trees and log-files for further analysis. Effective sample sizes (ESS) of all parameters are >300 (Alphapapillomavirus tree) and >2000 (HPV variant trees of each type), indicating that all Bayesian chains were well sampled and have converged. Best model estimates were selected using a posterior simulation-based analogue of Akaike's Information Criterion for MCMC samples (AICM) [142], as implemented in Tracer v.1.6. The lower AICM values indicated a better model fit. A consensus tree was inferred using TreeAnnotater v.2.4.5 and visualized using scripts developed in-house in R. The linear model (lm) function in R was used to estimate the correlation between sequence diversity and divergence time of HPV types and variants.
|
10.1371/journal.pbio.2005752 | Disruption of CTCF-YY1–dependent looping of the human papillomavirus genome activates differentiation-induced viral oncogene transcription | The complex life cycle of oncogenic human papillomavirus (HPV) initiates in undifferentiated basal epithelial keratinocytes where expression of the E6 and E7 oncogenes is restricted. Upon epithelial differentiation, E6/E7 transcription is increased through unknown mechanisms to drive cellular proliferation required to support virus replication. We report that the chromatin-organising CCCTC-binding factor (CTCF) promotes the formation of a chromatin loop in the HPV genome that epigenetically represses viral enhancer activity controlling E6/E7 expression. CTCF-dependent looping is dependent on the expression of the CTCF-associated Yin Yang 1 (YY1) transcription factor and polycomb repressor complex (PRC) recruitment, resulting in trimethylation of histone H3 at lysine 27. We show that viral oncogene up-regulation during cellular differentiation results from YY1 down-regulation, disruption of viral genome looping, and a loss of epigenetic repression of viral enhancer activity. Our data therefore reveal a key role for CTCF-YY1–dependent looping in the HPV life cycle and identify a regulatory mechanism that could be disrupted in HPV carcinogenesis.
| Oncogenic human papillomavirus (HPV) infection causes cancers of the anogenital and oropharyngeal tracts. HPV infects undifferentiated basal cells of the epithelium at these body sites and expresses low levels of viral early genes, required for replication of the viral genome. In normal epithelia, cellular migration away from the basal layer induces cell cycle exit and differentiation. However, in an HPV-infected cell, differentiation induces increased transcription of the viral early genes to prevent cell cycle exit, supporting amplification of the viral DNA. In this study, we show that the HPV genome recruits the cellular transcriptional regulators CTCF and YY1, which coordinate an epigenetically repressed chromatin loop between the YY1-bound viral transcriptional enhancer and CTCF-bound early gene region to attenuate early gene expression in undifferentiated cells. As the cells differentiate, YY1 protein expression and recruitment to the viral genome is dramatically reduced. This results in a loss of chromatin loop formation, epigenetic de-repression of the viral genome, and enhanced viral early gene expression. The coordination of viral gene expression with cellular differentiation is vital for persistence of infection and completion of the virus life cycle, and disruption of HPV transcriptional control is also a key step in the development of cancer.
| Human papillomaviruses (HPVs) are a family of small, double-stranded DNA viruses that infect epithelia at specific anatomical sites. Infection with any of the 12 mucosal oncogenic HPV types is a risk factor for the development of epithelial cancers such as cancer of the uterine cervix and oropharynx [1]. The majority of these epithelial cancers are caused by infection with the HPV16 and 18 viral subtypes.
The HPV life cycle is dependent on the differentiation of infected keratinocytes. Infection is established in the undifferentiated basal cells of epithelia, allowing the virus access to the cellular DNA replication machinery required to replicate viral episomes. To maintain the cell in a proliferative state, the viral E6 and E7 oncoproteins work synergistically to delay differentiation and prevent cell cycle exit. These essential viral proteins are encoded by transcripts that initiate from a short promoter situated immediately upstream of the early transcription start site, termed P105 in HPV18 [2]. The activity of P105 is controlled by enhancer and silencer sequences upstream of the promoter in the 850 basepair (bp) viral long control region (LCR). P105 contains a canonical TATA box, essential for the recruitment of the general transcription factor II D (TFIID) and the initiation of RNA polymerase II (RNA Pol II)-dependent transcription [3]. Proximal to the TATA box is a keratinocyte-specific 3′ enhancer, which recruits cellular transcription activators such as Sp1 and AP-1 (Fos/Jun) [4–6]. Situated within the 3′ enhancer is a silencer region that contains an array of Yin Yang 1 (YY1) binding sites. YY1 recruitment within this region has strong repressive effects on early gene transcription by the exclusion of AP-1 binding [7,8]. It has also been shown in the related HPV31 that the transcription elongation factor TEF-1 and YY1 work cooperatively to activate a second 5′ distal enhancer within the viral LCR [9], and YY1 binding at this site increases as cells differentiate [10]. However, YY1 binding to the 5′ distal enhancer has minimal effects on transcription in HPV16 [7].
The episomal papillomavirus genome associates with histones to form nucleosomes that are subject to epigenetic modification through the specific recruitment of cellular transcription factors that regulate viral transcription [11]. In HPV31, levels of acetylated histone H3 and H4 within the LCR increased upon cellular differentiation, particularly in the keratinocyte-specific enhancer, and correlated with increased transcription [10].
It is clear that epigenetic regulation of HPV transcription plays an important role in the HPV life cycle and in enhanced viral oncogene expression during disease progression [10,12]; however, the mechanisms involved in this regulation have not been determined. We have previously shown that the chromatin-organising transcriptional insulator protein CCCTC-binding factor (CTCF) associates with oncogenic HPV16 and 18 in the E2 open reading frame (ORF), approximately 3,000 nucleotides downstream of the viral LCR [13]. CTCF is a ubiquitously expressed host-cell chromatin-binding protein that associates with tens of thousands of sites within the human genome [14]. Depending on the context of the binding site, CTCF can function as an epigenetic insulator or coordinate long-range interactions between gene promoters and distant enhancers [15,16]. Notably, mutation of the CTCF binding site within the E2 ORF of HPV18 resulted in increased production of E6/E7 encoding transcripts, leading to hyperproliferation of viral genome–containing keratinocytes in organotypic raft culture [13]. CTCF also binds to the DNA genomes of much larger herpesviruses such as Epstein-Barr virus (EBV), Kaposi sarcoma–associated herpesvirus (KSHV), and herpes simplex virus (HSV-1), and CTCF recruitment in these viruses is important in the regulation of epigenetic silencing of latency-associated genes [17–22]. This regulation is in part brought about by the ability of CTCF to coordinate long-range chromosomal interactions within the viral episomes [19,23]. However, in other contexts, CTCF functions to insulate epigenetic boundaries in these large DNA viruses [18,24].
The mechanism by which CTCF regulates HPV oncogene expression is not known. In this study, we identify CTCF- and YY1-dependent loop formation in the HPV18 genome as the mechanism through which viral oncogene expression is restricted in the early stages of infection in undifferentiated keratinocytes. We show that down-regulation of YY1 following differentiation results in loss of loop formation and reversal of epigenetic silencing and facilitates increased oncogene expression and completion of the viral life cycle.
To analyse the function of CTCF in the life cycle of HPV18, primary human keratinocytes, the natural host cell of HPV, were transfected with religated HPV18 genomes, and replicating episomes were established that were stably maintained at approximately 50 copies per cell. We previously showed that mutation of the CTCF binding site within the E2 ORF of HPV18 (HPV18 ΔCTCF) results in a marked reduction of CTCF binding at this site with no effect on the establishment of HPV18 episomes in primary keratinocytes [13]. However, the long-term persistence of herpesvirus saimiri (HVS) has been shown to be dependent on CTCF [25]. Therefore, we serially passaged HPV18 wild-type (WT)- and ΔCTCF-genome–containing human foreskin keratinocytes (HFKs) and performed Southern blot analysis to examine HPV18 episome copy-number variation over time. The genome copy number at all passages analysed (9–11) was similar between HPV18 WT and ΔCTCF genomes, demonstrating that CTCF binding within the E2 ORF does not play a role in the persistence of HPV18 episomes (Fig 1A). It is important to note that all of the experiments included in this study were performed on cells between passages 9 and 11 to ensure consistent episomal copy number between HPV18 WT and ΔCTCF cultures since viral episomes can integrate into the host genome in long-term culture [26].
To determine whether HPV18 genome establishment alters CTCF protein expression, we quantified CTCF protein in isogenic primary HFKs. We observed a >2.5-fold increase in CTCF protein expression following establishment of HPV18 episomes (Fig 1B). This was consistent in two independent donors and is in agreement with a previous study that demonstrated an increase in CTCF protein expression in HPV31-positive neoplastic cervical keratinocytes compared to HFKs [27]. Interestingly, the HPV18-induced increase in CTCF protein is post-transcriptional since quantitative RNA-Sequencing (RNA-Seq) and quantitative reverse transcriptase-PCR (qRT-PCR) analysis of CTCF transcripts did not show any significant differences in CTCF transcript levels following establishment of HPV18 episomes (Fig 1C and 1D and S1 Table).
To determine whether abrogation of CTCF binding at the E2 ORF affects CTCF recruitment elsewhere in the viral episome, we performed chromatin immunoprecipitation followed by quantitative PCR (ChIP-qPCR) to specifically amplify CTCF-bound regions throughout the HPV18 genome (Fig 1E). CTCF binding was enriched at the previously identified E2 ORF binding site in cells containing HPV18 WT genomes. In addition, CTCF-enriched regions were identified within the viral LCR, close to the late promoter, and within the L2 ORF. Interestingly, abrogation of CTCF binding at the E2 ORF by mutation resulted in an almost complete loss of CTCF recruitment to all regions of the viral genome, suggesting that CTCF binding at the E2 ORF influences recruitment to regulatory regions that do not contain CTCF binding sites. This phenomenon was consistent in both keratinocyte donors tested.
We previously concluded that CTCF recruitment is important in the regulation of HPV18 oncogene expression in differentiated epithelia [13]. Consistent with these results, we found that in undifferentiated cells, transcripts originating from the early promoter were increased in abundance in quantitative RNA-Seq experiments (Fig 1F and S2 Table), which was confirmed by qRT-PCR (Fig 1G). Importantly, our RNA-Seq analysis showed that this increase in early transcripts is specific to E6/E7 encoding spliced transcripts and not to alternatively spliced E2 encoding mRNA species (Fig 1F and S2 Table), which is in agreement with our previous observation that E2 protein expression is not altered in HPV18 ΔCTCF genomes compared to WT [13]. E6 and E7 protein translated from the polycistronic message increased 11.3- and 1.9-fold, respectively, when the CTCF site was mutated (Fig 1H). To exclude the possibility that abrogation of CTCF binding by mutation of the E2–CTCF binding site results in increased E6/E7 transcription by inadvertently affecting the binding of other factors involved in an alternative regulatory network, CTCF protein levels were depleted by doxycycline-induced expression of two independent CTCF-specific shRNA molecules in HPV18 WT-genome–containing cells (Fig 1I). qRT-PCR analysis of E6/E7 encoding transcript levels demonstrated that partial depletion of CTCF protein resulted in a significant increase in E6/E7 encoding transcripts (Fig 1J). This increase in E6/E7 transcripts was not observed following induction of a nontargeting shRNA control (Fig 1J).
Our data show that recruitment of CTCF within the E2 ORF represses HPV18 early gene expression, and we hypothesised that this was due to repression of early promoter activity. Regulatory genomic elements are depleted of nucleosomes, and the remaining nucleosomes are enriched in active chromatin marks (e.g., acetylated lysine residues in histone H3 and H4) [28]. Formaldehyde-assisted isolation of regulatory elements (FAIRE) can be used to identify open and nucleosome-depleted enhancer regions of DNA [29]. To gain mechanistic insight into the control of HPV early promoter activity by distal CTCF binding, the chromatin accessibility of HPV18 episomes was analysed by FAIRE. We consistently observed a higher FAIRE-to-input amplification ratio, indicative of open chromatin at the HPV18 WT viral enhancer and early promoter (Fig 2A). Notably, there was a significant enrichment of open chromatin at the early promoter of HPV18 ΔCTCF genomes (Fig 2A; p < 0.001). This increased chromatin accessibility was consistent between independent donor lines and suggests a mechanism by which CTCF binding at the distal E2 binding site influences nucleosome occupancy within the viral LCR.
Interestingly, we observed that immediately downstream of the open chromatin area in the HPV18 WT genome is a region of closed chromatin in the E6 and E7 ORFs and the late promoter, P811. These findings are in agreement with previous DNase I footprinting experiments that demonstrated dynamic nucleosome binding in the viral enhancer and tightly held nucleosomes at the viral late promoter [11]. FAIRE analysis of the HPV18 episome also revealed an area of open chromatin within the E1 ORF, the specific function of which remains unknown.
We next investigated whether the increased accessibility of chromatin within the viral LCR following disruption of CTCF binding was associated with any change in active and repressive epigenetic marks. Using ChIP-qPCR, we analysed levels of the active-promoter–associated H3K4Me3 mark and the polycomb repressor complex (PRC)-associated repressive H3K27Me3 mark across the HPV18 genome. These experiments revealed an enrichment of H3K4Me3 in HPV18 ΔCTCF genomes compared to WT, particularly within the viral enhancer and immediately downstream of the early promoter, indicative of active transcription (Fig 2B). In contrast, enrichment of the repressive H3K27Me3 mark was detected in the enhancer and early gene region of HPV18 WT genomes. This finding was surprising, given that expression of E6/E7 has been shown to cause a global reduction in cellular H3K27Me3 [30]. The enrichment of H3K27Me3 was markedly decreased to almost undetectable levels on HPV18 ΔCTCF genomes (Fig 2C). This epigenetic switching of viral genomes unable to bind CTCF at the E2 ORF is consistent with increased transcriptional activity of the viral early promoter in ΔCTCF genomes and explains the observed alterations in chromatin accessibility identified by FAIRE (Fig 2A).
Since H3K27Me3 has been shown to inhibit recruitment of the general transcription machinery, we examined RNA Pol II recruitment. Indeed, enrichment of RNA Pol II was observed at the early promoter and within the early gene region in HPV18 ΔCTCF compared to WT, consistent with increased transcription levels (Fig 2D).
Numerous host-cell transcriptional regulators have been shown to specifically bind to the HPV LCR and regulate transcription of viral early genes. To determine whether CTCF influences recruitment of specific cellular regulators of HPV18 transcription, we analysed enrichment of transcription factors that regulate early gene transcription using ChIP-PCR. Our analysis revealed that mutation of the CTCF binding site resulted in significantly reduced binding of the YY1 transcription factor at both the 5′ and 3′ enhancers in the viral LCR and at the early and late promoter regions (Fig 3A). Since YY1 functions in the sequence-specific recruitment of PRCs PRC1 and PRC2, and we detected enrichment of the PRC2-associated H3K27Me3 repressive mark in WT HPV18 genomes, we examined PRC1 and PRC2 binding in the HPV18 genome in WT- and ΔCTCF-genome–containing cells. We found that the PRC2 subunit embryonic ectoderm development (EED) was significantly depleted at the viral enhancer and early promoter in HPV18 ΔCTCF genomes compared to WT, consistent with the observed loss of H3K27Me3 (Fig 3B). Reinforcement of repressed chromatin is achieved via recruitment of PRC1 to the H3K27Me3 mark and ubiquitylation of K119 on histone H2A by the PRC1 E3 ubiquitin ligase, ring finger protein 1B (Ring1B) [31]. Our data demonstrated that Ring1B was associated with the early promoter in HPV18 WT genomes, but binding was dramatically reduced in ΔCTCF genomes (Fig 3C). Loss of Ring1B was coincident with an almost complete loss of histone 2A lysine 119 ubiquitinylation (H2AK119Ub) (Fig 3D). These data indicate that in addition to reduced PRC2 recruitment, PRC1 recruitment to the viral LCR is significantly reduced in HPV18 ΔCTCF episomes. Our data are consistent with a model in which the abrogation of CTCF binding in the E2 ORF results in a loss of YY1 binding to the viral LCR, causing reduced PRC1 and PRC2 recruitment. This leads to reduced H3K27Me3 and H2AK119Ub deposition, de-repression of the HPV18 early promoter, and the up-regulation of viral oncogene expression.
Studies have shown that CTCF and YY1 are able to directly interact and that the assembly of this protein complex induces chromatin loop formation between distant loci [32,33]. We therefore hypothesised that the repressive effects of CTCF binding in the E2 ORF is mediated through loop formation between the YY1-bound viral enhancer and the downstream CTCF-bound E2 ORF. To test this hypothesis, we used chromosome conformation capture (3C), a method that can be used to directly measure inter- and intramolecular interactions between specific distal loci. HPV18 WT and ΔCTCF genomes were cross-linked in situ, and chromatin was extracted and digested with the NlaIII restriction enzyme, which restricts the viral DNA at multiple sites (Fig 4A). Digestion efficiency was determined for each sample by qPCR analysis of amplicons that are sensitive to digestion compared with a PCR amplicon that is insensitive to digestion. Samples were only processed further if the digestion efficiency was above 90%. Proximity ligation at low dilution was then carried out to ligate restriction fragments containing DNA loci that are physically associated through interactions between chromatin-bound factors. We designed unidirectional PCR primers to amplify a 346 bp amplicon across the ligation junction that would be formed if the restriction fragments containing the CTCF site in the E2 ORF and the LCR were ligated together as a result of the formation of a chromatin loop (Fig 4A and 4B).
3C analysis of HPV18 WT genomes consistently detected a PCR product of the correct size formed by ligation of the E2 ORF to the LCR. In addition, the 346 bp PCR products were excised from the gel and sequenced to confirm ligation between the YY1-bound viral LCR and CTCF-bound E2 ORF (S1 Fig). Given the small size of the HPV genome, we controlled for nonspecific interactions by carrying out PCR using primers designed to amplify ligation products between the CTCF-bound E2 ORF and the L2 ORF. No interactions were detected between these regions of the genome, although the E2–L2 primers efficiently amplified synthesised DNA molecules containing this ligation junction (Fig 4B, middle panel). We also performed a PCR reaction with primers that anneal within the E1 ORF that are insensitive to NlaIII digestion to ensure equal amplification of digested input chromatin in all 3C experiments (Fig 4B, lower panel). Notably, we found that looping between the YY1-bound LCR and the E2 ORF was significantly reduced in HPV18 ΔCTCF genomes in both donor lines tested (Fig 4C). To confirm that the reduction in E2 ORF–LCR interactions in HPV18 ΔCTCF was due to abrogation of CTCF binding, CTCF protein levels were depleted in HPV18-genome–containing keratinocytes using three independent shRNA lentiviral vectors following induction of shRNA expression with doxycycline (Fig 4D). While E2–LCR interactions were consistently detected in control shRNA-expressing cells, this interaction was significantly reduced following partial depletion of CTCF with all three independent shRNA molecules (Fig 4E and 4F). These data therefore demonstrate that CTCF directs the formation of a chromatin loop between the viral LCR and E2 ORF.
We next assessed the role of YY1 in the formation of this chromatin loop by shRNA-mediated depletion of YY1. HPV18-genome–containing cells were transduced with lentivirus expressing doxycycline-inducible YY1-specific shRNA (Fig 4G). 3C analysis demonstrated that depletion of YY1 resulted in a consistent and significant reduction in E2–LCR interactions (Fig 4H and 4I). Together, our data demonstrate that E2–LCR loop formation in the HPV18 genome requires both CTCF and YY1.
To confirm that YY1-CTCF–dependent chromatin loop formation within the HPV18 episome regulates chromatin topology, we used FAIRE to assess the chromatin structure within the viral LCR and flanking regions following shRNA-mediated depletion of CTCF (Fig 5A) and YY1 (Fig 5B). These experiments revealed a significant increase in chromatin accessibility within the LCR following depletion of either CTCF or YY1, consistent with the increase in chromatin accessibility in HPV18 ΔCTCF genomes that are unable to bind CTCF.
Our data show that CTCF and YY1 contribute to chromatin loop formation within the HPV18 genome, resulting in epigenetic repression of early gene expression. To determine whether CTCF and YY1 binding to the viral genome are interdependent, we depleted CTCF protein by shRNA induction and performed ChIP for CTCF and YY1 at the viral LCR and E2–CTCF binding site (Fig 5C). Depletion of CTCF protein resulted in reduced recruitment of CTCF to the E2–CTCF binding site and also a reduction in binding in the viral LCR. Notably, CTCF depletion also resulted in reduced YY1 recruitment to the LCR, and YY1 depletion resulted in reduced CTCF binding at the E2–CTCF binding site (Fig 5D). These data suggest that CTCF binding at the E2 ORF stabilises YY1 binding at the LCR and vice versa and that the enrichment of these proteins within the viral genome is interdependent.
It has previously been shown using transcriptional reporter plasmids that YY1 plays a pivotal role in the repression of HPV enhancer activity [7,8]. To confirm that YY1 is an essential repressor of HPV18 early gene expression in the context of the HPV genome, YY1 protein was depleted by YY1-specific shRNA expression as previously described and E6/E7 encoding viral transcripts quantified by qRT-PCR. Depletion of YY1 resulted in an over 20-fold increase in E6/E7 transcript levels (Fig 6A), confirming the role of YY1 as a transcriptional repressor in the HPV life cycle.
HPV gene expression during the virus life cycle is dependent on keratinocyte differentiation. Differentiation of infected keratinocytes in the midlayers of epithelia corresponds to an increase in early promoter activity [34–37]. In agreement with these studies, synchronous differentiation of keratinocytes by suspension of keratinocytes in semisolid medium for 48 hr resulted in a 2.6-fold increase in E6/E7 encoding transcripts in both keratinocyte donors (Fig 6B) and protein expression of the intermediate–early keratinocyte differentiation marker involucrin and a marker of the productive phase of the HPV life cycle, E1^E4 (Fig 6C). It has previously been shown that CTCF protein is localised to the nucleus of human keratinocytes and that expression is reduced in differentiated layers of human skin [38] and following morphological differentiation of human corneal epithelial cells [39]. Because of the known interaction between CTCF and YY1 and the functional role of this interaction in 3D chromatin loop formation, we also analysed YY1 expression in HPV18-genome–containing keratinocytes grown in monolayer (undifferentiated) or synchronously differentiated by suspension in methylcellulose for 48 hr. Western blot analysis of lysates of undifferentiated and differentiated cells revealed no difference in CTCF protein expression (Fig 6C), and importantly, CTCF protein was expressed at similar levels in HPV18 WT- and ΔCTCF-genome–containing cells (expression in HPV18 ΔCTCF compared to WT was 0.94-fold ± 0.18 SD; p = 0.97). However, a significant reduction of YY1 protein expression was observed in both HPV18 WT- and ΔCTCF-genome–containing cultures (Fig 6C and 6D), and this was consistent in two independent keratinocyte donors.
Since keratinocyte differentiation results in a marked reduction in YY1 protein expression, we examined whether cellular differentiation results in reduced CTCF and YY1 recruitment to the HPV18 genome. HPV18 WT-genome–containing cells were differentiated in methylcellulose and CTCF, and YY1 recruitment was analysed by ChIP. Differentiation of the cells resulted in reduced CTCF recruitment throughout the viral genome (Fig 6E) and a dramatic and complete loss of YY1 recruitment to the viral LCR (Fig 6F). To determine the effect of the differentiation-induced reduction in YY1 expression and recruitment to the HPV18 genome on the interaction between the E2–CTCF binding site and the viral LCR, 3C analysis was carried out using undifferentiated HPV18 WT-genome–containing cells harvested after growth in monolayer culture or genome-containing cells differentiated through incubation in methylcellulose for 48 hrs. We found that E2–LCR interactions were significantly reduced following cellular differentiation in two independent keratinocyte donors (Fig 6G and 6H).
We next tested whether the decreased looping between the E2 ORF and the LCR we observed on keratinocyte differentiation resulted in changes in the epigenetic status and chromatin accessibility of HPV18 WT genomes. ChIP-qPCR analysis of H3K4Me3 and H3K27Me3 levels revealed enrichment of the active H3K4Me3 mark and loss of H3K27Me3 modifications, indicative of transcriptional de-repression following differentiation (Fig 7A). In contrast, differentiation of ΔCTCF HPV18-genome–containing cells that we previously showed were in enriched H3K4Me3 histone marks when undifferentiated (Fig 2B) did not result in any further enrichment of H3K4Me3 following differentiation, indicating aberrant regulation of epigenetic changes upon cellular differentiation in genomes unable to recruit CTCF (Fig 7B). We next examined whether this switching of epigenetic modifications results in increased accessibility of the chromatin using FAIRE. Our data demonstrated that differentiation of HPV18-genome–containing cells resulted in significant depletion of nucleosomes in the HPV18 LCR (Fig 7C). Taken together, our data demonstrate that upon cellular differentiation, reduced YY1 protein expression and recruitment to the viral LCR leads to the loss of chromatin loop formation, depletion of repressive epigenetic marks, and an associated increase in LCR chromatin accessibility. These observations therefore elucidate the mechanism underlying the progressive up-regulation of HPV18 E6 and E7 expression during keratinocyte differentiation, a mechanism likely to play a critical role in the successful completion of the viral life cycle.
CTCF is a major regulator of host and virus transcription and mediates many of its functions by the coordination of dynamic long-range chromosomal interactions [15]. In this study, we show that mutation of the CTCF binding site with the E2 ORF of HPV18 results in a significant depletion of CTCF binding throughout the HPV genome. This interesting phenomenon suggested that intramolecular interactions occur between distinct regions of the HPV18 episome and that these interactions are stabilised by CTCF bound at the E2 ORF. For example, CTCF association with LCR-specific sequences, devoid of CTCF consensus binding sites [13], could occur via indirect interaction with the CTCF-bound E2 ORF, suggesting that the viral genome is organised by distinct intramolecular interactions. The global reduction of CTCF binding, by either mutation of the E2 binding site or by induction of CTCF-specific shRNA, resulted in increased E6/E7 transcript and protein production. We therefore hypothesised that CTCF mediates intrachromosomal interactions that are important for controlling the activity of the viral early promoter (depicted in Fig 8).
Analysis of the epigenetic status of the HPV18 WT episome revealed several important features. H3K4Me3 enrichment, indicative of active promoter regions, was observed at the early and late promoter regions in comparison to other regions in the HPV18 genome. In contrast, repressive H3K27Me3 levels were low at the early promoter, consistent with a previous study in HPV31 [10]. However, higher levels of H3K27Me3 were observed in the viral LCR, suggesting epigenetic repression of enhancer activity in undifferentiated cells, which has not previously been shown. Enhanced enrichment of H3K27Me3 was also observed at the late promoter. Quantitative analysis of these epigenetic marks in HPV18 ΔCTCF episomes demonstrated that attenuation of CTCF binding resulted in dramatic epigenetic switching in the viral genome, as evidenced by increased accessibility of the viral enhancer within the LCR and an enrichment of H3K4Me3 alongside a global loss of H3K27Me3 marks. This alteration in chromatin accessibility and epigenetic status correlated with enhanced recruitment of RNA Pol II. We previously reported that mutation of the CTCF binding within the E2 ORF of HPV18 results in a reduction in exon 416 to 929 inclusion [13]. Cotranscriptional splicing of RNA is physically linked to RNA Pol II activity, and it has been clearly demonstrated that transcription elongation dynamics influence intron identification and processing by the spliceosome (reviewed by [40]). We therefore hypothesise that the observed alteration of HPV transcript splicing [13] is due to altered RNA Pol II dynamics, and we will formally test this hypothesis in future studies.
YY1 in part functions as a cellular transcriptional repressor by mediating the recruitment of PRC1 and PRC2 to specific enhancer loci [41,42]. PRC2 catalyses H3K27Me3 deposition while PRC1 catalyses H2AK119Ub deposition, together resulting in transcriptional repression. Since a dramatic loss of H3K27Me3 enrichment was observed in HPV18 genomes unable to bind CTCF, we hypothesised that PRC2 was depleted. We demonstrated significant loss of the PRC2 component EED, providing evidence that PRC2 recruitment is reduced in HPV18 ΔCTCF genomes, resulting in reduced H3K27Me3 deposition. In addition, a significant reduction of the PRC1 catalytic subunit Ring1B to the viral early promoter was observed, resulting in reduced H2AK119Ub deposition. The reduction in both PRC1 and PRC2 recruitment to the viral LCR following abrogation of CTCF binding explains the loss of repressive epigenetic marks and increased chromatin accessibility and activity of the P105 early promoter.
CTCF and YY1 can physically associate to stabilise chromatin loops, and organisation of the host cell genome in this manner controls specific gene-expression switching in X chromosome inactivation and neural cell differentiation [32,33]. Combined with our data showing the global loss of CTCF recruitment within the HPV18 genome when the dominant E2 ORF binding site was mutated and CTCF-mediated regulation of LCR topology and YY1 enrichment, we hypothesised that CTCF and YY1 mediate an intramolecular interaction between the E2 ORF and the LCR to stabilise an epigenetically repressed chromatin domain. We demonstrated a specific interaction between the E2 ORF and LCR that was dependent on CTCF and YY1 expression. We show that disruption of the E2–LCR interaction results in transcriptional de-repression of the viral early promoter through a dramatic alteration of the epigenetic status of the viral episome and increased chromatin accessibility. Such short-range interactions have been previously identified in cellular loci using similar methods, including interactions between the insulin gene promoter and distal enhancer and within the 2.5 kbp CD68 gene [43,44]. In addition, a recent study has demonstrated genomic interactions within the KSHV genome, ranging from 5 kbp to >80 kbp in size [45]. These studies provide evidence that short-range genomic loci are important in the regulation of host cell and episomal virus transcription regulation. However, it is important to note that the 3C analysis used in our studies could also detect interchromosomal interactions between multiple viral episomes in the same cell. Such interactions between viral episomes could stabilise the formation of viral super enhancers [46] and/or function in the homologous recombination-dependent replication of episomes in replication centres [47].
Our results provide important insight into YY1 and CTCF function in transcriptional control. We demonstrate that depletion of CTCF reduces YY1 recruitment and vice versa, suggesting that CTCF and YY1 bind to the viral genome in a cooperative manner. CTCF and YY1 have been shown to physically interact [33], and studies have shown that YY1 and CTCF can anchor loops via homo- and heterodimerisation [32,48]. It has previously been shown that over 30% of YY1-occupied sites in the human genome are at locations directly adjacent to CTCF-occupied sites, but that the binding of these factors do not directly colocalise, suggesting that these factors work together to cooperatively influence occupancy at adjacent binding sites [32, 49]. YY1 is enriched at sites within the host chromatin that engage in 3D looping, and YY1 enrichment at these sites is reduced when these elements are not connected, suggesting that YY1 binding is stabilised by 3D chromatin interactions [32]. It has therefore been suggested that CTCF binding initially serves as an architectural ‘seed’ and that YY1 binding then connects CTCF-bound nearby regulated genes and enhancers. Our results show that depletion of CTCF reduces YY1 recruitment, consistent with this hypothesis, but also suggest that CTCF binding is also influenced by YY1 recruitment.
In the HPV life cycle, the E6 and E7 oncoproteins are expressed at low levels in basal keratinocytes, presumably to limit host immune activation and because their combined functions in the cell cycle to maintain expression of the cellular DNA replication machinery are less important in these undifferentiated, cycling cells. Host cell differentiation is associated with increased viral early transcript production, resulting in increased E6/E7 protein expression as well as activation of the late promoter [35]. E6 and E7 act to maintain host cell proliferation, maintaining viral access to the host cell DNA replication machinery, and it has been shown that viral genome amplification in differentiated epithelia requires robust E6 expression [50]. Our data demonstrated a significant reduction in E2–LCR loop formation in differentiated keratinocytes and an associated reduction in epigenetic repression of the viral genome and increased accessibility of the LCR. Our results support a model in which the level of YY1 protein expression controls viral oncogene expression during differentiation.
The in-depth analysis of the epigenetic status and topology of HPV18 episomes in a physiological model of the HPV18 life cycle has provided mechanistic insight into the underlying differentiation-dependent control of HPV18 early gene expression. We have demonstrated that CTCF and YY1 together function in coordinating a transcriptional switch that is directly linked to host cell differentiation (Fig 8). This ensures low-level expression of viral proteins in the basal cells, which presumably facilitates persistence in vivo. As infected cells differentiate, the epigenetically repressed chromatin loop responsible for attenuating activity of the viral enhancer is disrupted, and the repressive epigenetic marks are lost. We show that this mechanism of controlling viral gene expression is regulated by CTCF-YY1 interactions within the HPV18 episome; as cells differentiate, YY1 protein expression is repressed and loop formation is disrupted, promoting enhancer activation.
Whether this mechanism of differentiation-dependent regulation of HPV oncogene expression plays a role in HPV-driven cancer is not clear, but it is tempting to speculate that this is the case. YY1 binding sites are often mutated in the HPV16 genome in cervical cancer [51,52], and a recent study has demonstrated that an open chromatin state of the viral LCR correlates with high E6/E7 expression in a model of HPV16-driven carcinogenesis [12]. Since the CTCF binding site within the E2 ORF is conserved in HPV16 [13], we predict that a similar mechanism of oncogene repression exists in HPV16. In addition, CTCF recruitment to the E2 ORF within integrated HPV18 DNA in HeLa cells is very low even though the CTCF binding site is intact [53]. Low CTCF binding in HeLa cells is coincident with low H3K27Me3 and high H3K4Me3 marks at the viral LCR and early promoter, combined with high E6/E7 transcript production, which is in agreement with our findings in HPV18 ΔCTCF episomes. It will therefore be of importance to determine whether CTCF-mediated attenuation of viral oncogene expression is disrupted in HPV-driven cancers. To begin to answer this question, we have analysed CTCF binding-site mutations in a cohort of 3,215 HPV16 positive lesions and correlated our findings with clinical outcome [54]. A variation in the CTCF binding-site motif 2 was discovered in 357 individual HPV16 sequences (A2938 to G), which we predict would enhance CTCF binding [55]. Interestingly, the presence of this genetic variation in the HPV16 genome is significantly associated with decreased cancer incidence when compared to the cases with no variation in the vicinity of the binding site (p = 0.050, one-tailed Fisher’s test). We therefore speculate that in lesions that contain this variant of HPV16, CTCF may bind with higher affinity and have a more significant effect on the attenuation of E6/E7 expression, thereby reducing the risk of cancer development. In addition to genetic variations within the CTCF binding site, we also found many sequences that had no sequence information at and around the E2 ORF CTCF binding site. This could be due to integration of the virus such that the E2 coding sequence is disrupted, but in addition to this widely reported mechanism of HPV-driven carcinogenesis, it will be important to determine the mechanism and consequence of CTCF exclusion in cancers with integrated HPV DNA.
The collection of circumcised foreskin tissue from newborns for the isolation of primary HFKs for investigation of HPV biology was approved by Southampton and South West Hampshire Research Ethics Committee A (REC Reference number 06/Q1702/45). Written consent was obtained from the parent or guardian. The study was approved by the University of Birmingham Ethical Review process (ERN_16–0540).
pGEMII-HPV18 (gift from F. Stubenrauch, University of Tübingen, Germany) contains the complete HPV18 genome cloned into the EcoRI site of pGEMII and was used to create pGEMII-HPV18-ΔCTCF, which contains three conservative nucleotide substitutions (C2993T, G3005A, T3020C) within the E2 coding region to abolish CTCF binding as previously described [13].
CTCF (61311), H3K4Me3 (39915), H3K27Me3 (39155), RNA Pol II (61081), SP1 (39058), TEF1 (61644), EZH2 (39901), EED (61203), and Ring1B (39663) and antibodies were purchased from Active Motif (La Hulpe, Belgium). YY1 antibody (SC-7341X) was purchased from Santa Cruz Biotechnology (Dallas, TX, United States of America). H2AK119Ub (D27C4) was purchased from Cell Signaling Technology, Inc (Danvers, MA, USA). FLAG M2 was purchased from Sigma-Aldrich (Gillingham, United Kingdom), E7 clone 8E2 (ab100953) was purchased from Abcam (Cambridge, UK), and E6 clone G7 (SC-365089) and Glyceraldehyde-3-phosphate dehydrogenase (GAPDH) antibody were purchased from Santa Cruz Biotechnology (Dallas, TX, USA). Monoclonal HPV18 E1^E4 antibody 1D11 was produced by S. Roberts [56]. All horseradish-peroxidase–conjugated secondary antibodies were purchased from Jackson Laboratories (Bar Harbor, ME, USA).
The transfection of normal primary HFKs from neonatal foreskin epithelia with recircularised HPV18 WT and ΔCTCF genomes was performed in S. Roberts’ laboratory by J. Parish as previously described [13,57]. To eliminate donor-specific effects, primary cells from two foreskin donors were used: one isolated in house and one commercially available (Lonza, Basel, Switzerland). Episome copy number in each cell line was determined by Southern blotting as described previously [26] and calculated by densitometry of three technical repetitions of each HPV18-transfected keratinocyte donor compared to the 50 copies per cell loading control. Digestion of the DNA with EcoRI results in linearisation of episomes, whereas digestion with BglII restricts the host cell DNA but not the viral DNA to reveal integrated or multimeric virus. DpnI digests input DNA only.
Organotypic raft cultures were prepared as previously described [13,57,58] and cultured for 14 d in E medium [58] without epidermal growth factor to allow cellular stratification. Rafts were fixed in 3.7% formaldehyde and paraffin embedded prior to sectioning (Propath Ltd., Hereford, UK).
HPV18-genome–containing keratinocytes (3 × 106 cells) were suspended in E medium containing 10% FBS and 1.5% methylcellulose and incubated at 37°C, 5% CO2 for 48 h. Cells were then harvested by centrifugation at 250 × g and then thoroughly washed with ice-cold PBS. Cells were then either resuspended in medium containing 1% formaldehyde to cross-link for ChIP and 3C or in urea lysis buffer for protein extraction.
ChIP assays were carried out using the ChIP-IT Express kit (Active Motif) following the manufacturer’s instructions. Briefly, cells were fixed in 1% formaldehyde for 3 min at room temperature, quenched in 0.25 M glycine, and washed in ice-cold PBS. Nuclei were released by 40 strokes in a tight dounce homogeniser. Samples were sonicated at 25% amplitude for 30 s on/30 s off for a total of 15 min using a Sonics Vibracell sonicator fitted with a microprobe. ChIP efficiency was assessed by quantitative PCR (qPCR) using SensiMix SYBR master mix using an MXPro 3000 (Agilent Technologies, Santa Clara, CA, USA). Primer sequences for ChIP experiments are shown in Table 1. Cycle threshold (CT) values were calculated at a constant threshold for each experiment, and fold-enrichment–compared to negative control FLAG antibody was calculated using the following formula:
Fold binding over IgG = (2ΔCT target)/(2ΔCT IgG),
where ΔCT target = Input CT−Target CT and ΔCT IgG = Input CT−IgG CT. Each ChIP experiment was performed in triplicate, and data shown are the mean ± SD of a representative experiment. Biological repeats were performed for each experiment a minimum of three times with similar results.
Cells were lysed in urea lysis buffer (8 M Urea, 100 mM Tris-HCl, pH 7.4, 14 mM β-mercaptoethanol, protease inhibitors) and protein concentration determined by Bradford assay. Equal amounts of protein were separated by SDS-PAGE and western blotting carried out using conventional methods.
RNA was extracted with an RNeasy Mini Kit (Qiagen, Hilden, Germany) according to the manufacturer’s protocol and DNase treated. For RNA-Seq, libraries were prepared using TruSeq Stranded mRNA Library Prep kit for NeoPrep (Illumina, San Diego, CA, USA) using 100 ng total RNA input according to manufacturer’s instructions. Libraries were pooled and run as 75-cycle–pair end reads on a NextSeq 550 (Illumina) using a high-output flow cell.
cDNA was synthesised using Superscript III (Invitrogen, Carlsbad, CA, USA) according to the manufacturer’s instructions. qPCR was performed using a Stratagene Mx3005P detection system with SyBr Green incorporation and the primers listed in Table 2.
Cells were fixed and chromatin extracted and sheared as described above for ChIP. FAIRE analysis was then carried out as previously described [29]. Briefly, two aliquots of chromatin were taken, each containing chromatin from approximately 2 × 106 cells, one for Input and one for FAIRE. To the FAIRE samples, 150 μl water was added. To the Input sample, 150 μl water and 10 μl of 5 M NaCl were added, and the samples were incubated at 95°C for 15 min to reverse the crosslinks. RNaseA (10 μg/μl) was added, and the samples incubated at 37°C for 15 min. Proteinase K (0.5 μg/μl) was added followed by incubation at 67°C for 15 min.
Both Input and FAIRE samples were then extracted with 200 μl phenol:chloroform:isoamylalcohol (25:24:1) and the aqueous layer retained. DNA was precipitated by conventional methods and the pellet resuspended in 50–150 μl 50 mM Tris-HCl, pH 7.4, 10 mM EDTA. Recovery of FAIRE-extracted DNA in comparison to Input DNA was then determined by qPCR using the ΔΔCT method. Primer sequences are shown in Table 1.
A total of 1–1.5 × 107 cells were trypsinised and resuspended in 1 ml 10% (v/v) FCS/PBS. Cells were passed through a 70 μm cell strainer and 9.5 ml of 1% formaldehyde in 10% FCS/PBS added before incubation for 10 min at RT with end-to-end rotation. Glycine was added to a final concentration of 125 μM before the cells were pelleted at 4°C. Cells were resuspended in 5 ml of ice cold lysis buffer (10 mM Tris-HCl, pH 7.7, 10 mM NaCl, 5 mM MgCl2, 0.1 mM EGTA, protease inhibitors) and incubated on ice for 10 min. Samples were centrifuged at 400 × g for 5 min at 4°C to pellet the nuclei. Five hundred μl 1.2× restriction enzyme buffer and 0.3% (final concentration) SDS were added and samples incubated at 37°C for 1 hr while shaking at 900 rpm. Fifty μl of 20% Triton X-100 was then added, followed by incubation at 37°C for 1 hr with shaking at 900 rpm. Prior to digestion, an aliquot was removed for assessment of digestion efficiency. Four hundred units of NlaIII restriction enzyme were added, and the samples were incubated at 37°C overnight with shaking at 900 rpm. An aliquot of each sample was removed and assessment of digestion efficiency performed by adding 500 μl of 5 mM EDTA, pH 8.0, 10 mM Tris-HCl, pH 8.0, 0.5% SDS, and 20 μg proteinase K and incubating at 65°C for 30 min. One μg RNase A was added, followed by incubation at 37°C for 2 hr. The DNA was extracted with PCI and ethanol precipitated using conventional methods and the pellet resuspended in 60 μl dH20. Digestion efficiency of the viral genomes was assessed by qPCR of genome regions sensitive and insensitive to containing NlaIII restriction sites (Table 3) and comparison of CT values as described in [59].
For ligation, 40 μl of 20% SDS was added to the samples, followed by incubation for 25 min at 65°C with shaking at 900 rpm. A total of 6.125 ml 1.15× ligation buffer and 1% (final concentration) Triton X-100 was added. Samples were incubated for 1 hr at 37°C with gentle shaking. One hundred units of T4 DNA ligase were added and the samples incubated for 4 hr at 16°C, followed by 30 min at RT. Three hundred μg proteinase K was added, and the samples were incubated at 65°C overnight.
To purify the digested DNA, 300 μg RNase A was added and the samples incubated for 45 min at 37°C. The DNA was extracted with PCI twice and ethanol precipitated. Finally, the DNA pellet was resuspended in 10 mM Tris-HCl, pH 7.5. Ligation of specific regions of the HPV18 genome was assessed by PCR using sense primers specific to the L1 and E2 ORFs to detect interactions between the CTCF-bound E2 ORF and the viral LCR and sense primers specific for the L2 and E2 ORFs to detect ligation events that could occur by chance (Table 3). PCR products were assessed by agarose gel electrophoresis and compared to products obtained with a synthesised DNA template equivalent to the predicted ligation product (GeneStrings). Products were sequenced and quantified with a Fusion FX imaging system.
Human embryonic kidney 293T (HEK293T) cells were grown in Dulbecco’s modified Eagle’s medium (DMEM) supplemented with 10% FBS and transfected with second-generation lentiviral packaging plasmids, psPAX2 and pMD2.G, and pTRIPZ-shRNA expressing plasmid using polyethylenimine (PEI) Max at a DNA-to-reagent ratio of 1:3. Medium containing lentiviral particles was recovered at 48 and 72 hr post transfection and passed through a 0.45 μM filter. The resulting recombinant lentiviruses were concentrated using Vivaspin ultrafiltration spin columns (50,000 MWCO PES) and used to spin infect HPV18-genome–containing keratinocytes growing in 6-well plates in E medium containing 8 μg/ml polybrene after feeders had been removed. Plates were spun at 3,220 × g for 90 min to facilitate infection, after which media were replaced with E medium. Twenty-four hr later, lentiviral infection was repeated as described above before cells were detached and seeded onto fresh irradiated feeder cells in 10 cm dishes. Puromycin to a final concentration of 1 μg/ml was added to the cells 72 hr later to select infected cells. shRNA expression was induced with 1 μg/ml doxycycline for 48 hr.
|
10.1371/journal.pntd.0001350 | Re-Emergence of Crimean-Congo Hemorrhagic Fever Virus in Central Africa | Crimean-Congo hemorrhagic fever (CCHF) is a severe tick-borne disease well recognized through Europe and Asia where diagnostic tests and medical surveillance are available. However, it is largely neglected in Africa, especially in the tropical rainforest of Central Africa where only sporadic human cases have been reported and date back to more than 30 years. We describe here an isolated human case that occurred in the Democratic Republic of the Congo in 2008 and performed phylogenetic analysis to investigate whether it resulted from a regional re-emergence or from the introduction of a novel virus in the area.
Near complete segment S and partial segment M sequences were characterized. Bayesian phylogenetic analysis and datation were performed to investigate the relationship between this new strain and viral strains from Africa, Europe and Asia. The new strain is phylogenetically close to the previously described regional genotype (II) that appears to be specific to Central Africa. Phylogenetic discrepancy between segment S and M suggested genetic exchange among local sublineages, which was dated back to 130–590 years before present.
The phylogenetic analyses presented here suggest ongoing CCHF virus circulation in Central Africa for a long time despite the absence of reported human cases. Many infections have most probably been overlooked, due to the weakness of healthcare structures and the absence of available diagnostic procedure. However, despite the lack of accurate ecological data, the sporadic reporting of human cases could also be partly associated with a specific sylvatic cycle in Central Africa where deforestation may raise the risks of re-emergence. For these reasons, together with the high risk of nosocomial transmission, public health authorities' attention should be drawn to this etiological agent.
| Crimean-Congo hemorrhagic fever virus (CCHFV) is transmitted to humans through tick-bite or contact with infected blood or tissues from livestock, the main vertebrate hosts in a peri-domestic natural cycle. With numerous outbreaks, a high case fatality rate (3%–30%) and a high risk for nosocomial transmission, CCHFV became a public health concern in Europe and Asia. However virus surveillance in Africa is difficult due to the limited sanitary facilities. Especially, CCHFV occurrence in Central Africa is very poorly described and seems highly in contrast with the temperate to dry environments to which the virus is usually associated with. We described a single human infection that occurred in Democratic Republic of the Congo after nearly 50 years of absence. The phylogenetic analysis suggests that CCHFV enzootic circulation in the area is still ongoing despite the absence of notification, and thus reinforces the need for the medical workers and authorities to be aware of the outbreak risk. The source of infection seemed associated with a forest environment while no link with the usual agro-pastoral risk factors could be identified. More accurate ecological data about CCHFV enzootic cycle are required to assess the risk of emergence in developing countries subjected to deforestation.
| Crimean-Congo hemorrhagic fever virus (CCHFV, family Bunyaviridae, genus Nairovirus) is a tick-borne virus. It causes severe illness throughout Africa, Asia, Southeast Europe and the Middle East, with case fatality rates ranging from 3% to 30%. Its worldwide distribution closely matches that of its main arthropod vector, ixodid ticks belonging to the genus Hyalomma. Human infection occurs through tick bites, contact with infected livestock, or nosocomial transmission. The CCHFV negative-stranded RNA genome is divided into a small (S), medium (M) and large (L) segment.
Previous phylogenetic analysis of the S segment clustered strains into 6 to 7 distinct phylogeographic groups: West Africa in group I, Central Africa (Uganda and Democratic Republic of Congo (DRC)) in group II, South Africa and West Africa in group III, Middle East and Asia (that may be split into 2 distinct groups Asia 1 and Asia 2 [4]) in group IV, Europe and Turkey in group V, and finally Greece in group VI [1]–[5]. However, some of these phylogenetic lineages include strains separated by large spatial distances (such as South Africa and West Africa) suggesting viral migration, most likely via migratory birds transporting infected ticks, or secondary introductions following importation of commercial livestock. Comparative phylogenetic analysis revealed, with a few exceptions, parallel clustering of the S and L segments, while M segment reassortment seems more frequent [1], [4]–[6].
During the last 60 years, CCHFV outbreaks have been described in Asia, the Middle East and the Balkans, where the virus has become endemic and caused several thousand human cases. During the last decade, CCHFV has caused human disease in previously unaffected countries (Turkey 2002, Iran 2003, Greece 2008, Georgia 2009) and has re-emerged in countries located southwest of the Russian Federation after an absence of nearly 30 years [7]. By contrast, fewer than 100 cases have been recorded in Africa [8], most of them in South Africa [9], [10]. In East and West Africa, enzootic CCHFV circulation has been shown by serological surveys of cattle and virus isolation from ticks since the 1970s [11], [12] but until the outbreaks in Mauritania in 2004 [13] and Sudan in 2008 [14], only sporadic human cases had been reported. In Central African Republic (CAR), limited serological evidences of CCHFV circulation in Zebu cattle has been provided [15] and three viral strains were isolated from ticks between 1973 and 1976, one of which lead to accidental infection of a laboratory worker [11]. Subsequent isolations from ticks were performed in the 80's [16] but no human case was reported. Despite the early identification of human CCHFV infection in DRC (Kisangani, 1956), CCHFV occurrence in Central Africa has not been much described and only sporadic human cases have been reported. One month after having isolated the first CCHFV strain (strain Congo 3011) in newborn mice, Dr. Courtois became infected and this was the last notified case in DRC, from which the strain Congo3010 was isolated [17], [18]. The virus was next identified in Uganda between 1958 and 1978. Fifteen CCHFV strains were isolated from febrile patients, of which nearly half were laboratory workers having handled infectious samples [11], [17], [18]. From the geographic information associated with the other patients, it can be inferred that CCHFV was present both in the Entebbe area and in the Arua district (previous West Nile district) located 350 km North, near the border of Sudan. Three CCHFV strains were also isolated from ticks and an early serological survey suggested cattle infection [11]. No other epidemiological or ecological information is available on CCHFV in Central Africa or its borders, and no further cases have been recorded.
In 2008, CIRMF (Centre International de Recherches Médicales de Franceville, Gabon) identified CCHFV in a serum sample received for etiological diagnosis of a case of hemorrhagic fever in DRC. This is the only identification of CCHV in DRC for more than 50 years. To determine whether it was due to introduction of a novel virus or to re-emergence of a local genotype, we determined the phylogenetic relationships between this virus (hereafter referred to as Beruwe-2008) and previously described isolates. Phylogenetic analysis showed that the Beruwe-2008 strain belonged to the genotype previously identified in this area and thus suggested that it had re-emerged. Local CCHFV persistence may have been supported by a sylvatic natural cycle specific to Central Africa, indicating that countries subject to major deforestation may note an increasing number of human infections.
Laboratory investigations were performed subsequently to the WHO request for surveillance and early alert of hemorrhagic fever outbreak in Central Africa. Because of the emergency settings associated with the suspicion of such acute illnesses, no ethics committee approval or written consent was deemed necessary. The blood sample was taken by national healthcare workers of the Lubutu hospital where the patient came for medical care. He was informed that his blood sample will be further used for diagnostic investigation and gave his verbal consent. The patient described here is anonymous. The blood sample was next sent to the Institut National de Recherche Biomédicale (Kinshasa, DRC) and then to CIRMF upon WHO authorities. The study was approved by the scientific committee of CIRMF.
The patient was a 26-year-old man living in Beruwe (Nord Kivu province) in DRC, 325 km from Kisangani (Figure 1). He became ill in the mining area where he worked. He complained of fever and headache on day 1 and developed moderate bloody diarrhea on day 2. Epistaxis, oral bleeding and hematuria occurred on day 3. He treated himself with ibuprofen and paracetamol during the first three days. On day 4 after onset he additionally took quinine and finally presented with severe asthenia and persistent bleeding to Lubutu hospital, where the serum sample was taken. At this stage the patient was subicteric, with bleeding at the venipuncture site, but had only low-grade fever (37.6°C). He declared no contact with wild animals during the previous three weeks but he had slept in the forest. No further information on his outcome was available.
The patient's serum was manipulated in biosafety level 4 (BSL-4) conditions. The serum was first tested for Ebola and Marburg viruses. As results were negative, investigations were next performed for CCHFV. RNA was extracted with the QIAamp viral RNA mini kit (Qiagen, Courtaboeuf, France) according to manufacturer's instructions. Reverse transcription (RT) and real-time PCR amplification were performed with the High Capacity cDNA RT kit and Taqman universal PCR master mix (Applied Biosystems - Life Technologies Corporation, Carlsbad, California), and previously reported primers and probes [19]. Conventional one-step RT-PCR was performed with CCHFV primers as previously reported [20] and with SuperScript III one-step RT-PCR system and Platinum Taq DNA polymerase (Invitrogen -Life Technologies Corporation, Carlsbad, California). This yielded a 536-nucleotide fragment in the S segment, sequencing of which confirmed CCHFV identification.
As viral isolation on Vero cells was unsuccessful, viral RNA was extracted from the patient's serum as described above, and was used for RT-PCR amplification with Platinum Taq DNA polymerase (Invitrogen). Primers were derived from nucleotide alignments (Table 1). Three overlapping PCR products allowed near-complete characterization of the S segment coding sequence (GenBank accession number HQ849545) and partial characterization of the M segment (GenBank accession number HQ849546). Amplification of the L segment was unsuccessful, being limited by the sample quantity.
A total of 44 complete sequences for segment S and 38 complete sequences for segment M were retrieved from GenBank (Table S1 in online supporting information). Nucleotides were aligned according to the amino-acid profile using the Translation Align algorithm implemented in Geneious software [21]. Initial phylogenetic analyses were performed with MrBayes V3.1 [22], [23] using a GTR+gamma+invariant site substitution model for 4 million MCMC chain iterations sampled every 100 generations, corresponding to 40 000 trees (data not shown). Following confirmation of the tree topology from MrBayes, the tip-dated coding alignments were submitted to Bayesian inference of node ages by using BEAST V1.4.7 [24] under the assumption of a codon-based substitution model (SRD06) and an uncorrelated relaxed lognormal molecular clock and expansion, exponential and constant population growth models. The Expansion model yielded the best results, as indicated by ESS statistics and Bayes factor analysis of the posterior probability trace in TRACER. Sixty million generations were sampled every 1000 states, corresponding to 60 000 trees, that were annotated with TreeAnnotator and visualized with FigTree V1.3.1 from the BEAST package.
In 2008 we received a serum sample for etiological diagnosis of a case of hemorrhagic fever in DRC. The patient' serum was handled under BSL-4 facilities for RNA purification and tested positive for CCHFV by real-time PCR and conventional amplification with previously described detection systems [19], [20]. The patient became ill in Beruwe, approximately 325 km from Kisangani, where the only 2 previously reported cases of CCHFV in DRC occurred in 1956 (Figure 1). The patient worked in a mining area near a forest environment and didn't seem linked to agro pastoral activities. As this was the only identified case of CCHV in DRC for more than 50 years, we performed a phylogenetic analysis to determine whether it was due to introduction of a novel virus or re-emergence of a local genotype.
Virus isolation in Vero cells was unsuccessful, presumably owing to virus degradation subsequently to difficulties and delays of transportation. Genetic characterization was thus based on RT-PCR of RNA extracted from the patient's serum. As reassortment usually affects the M segment, priority was given to sequencing segments S and M, while segment L amplification was limited by sample quantity and was unsuccessful. Near-complete characterization of the segment S coding sequence was achieved, yielding 1501 contiguous nucleotides; the 5′ end was missing, presumably owing to RNA degradation. A 1001-nucleotide fragment was generated for segment M, corresponding to nucleotide positions 2382 to 3380 of the Congo3010-1956 glycoprotein coding sequence (DRC strain).
Pairwise nucleotide comparison of the Beruwe-2008 segment S sequence with those of the most closely related strains Congo3010-1956 (DRC) and Semunya-1958 (Uganda) – showed 92.4% and 92.0% similarity, respectively. In segment M the pairwise identities were 96.1% and 93.8% respectively. Identity between the Beruwe-2008 strain and strains belonging to other genetic groups ranged from 82.2% to 87.6% in segment S and from 72.5% to 81.3% in segment M (Table 2).
Bayesian phylogenetic analysis with a molecular clock assumption was applied to segment S (Figure 2A) and M (Figure 2B) datasets. Both methods yielded tree topologies largely matching the phylogeographic groups previously defined from complete segments S and M [1]–[3]. In both segments, and with posterior probabilities reaching 1, the Beruwe-2008 sequence grouped with the aforementioned DRC and Uganda strains forming lineage II (Central Africa group). Although we cannot rule out the possibility of segment L reassortment, the Beruwe-2008 strain most likely belongs to the genotype previously identified in Central Africa, thus representing viral re-emergence rather than introduction of another genotype. In addition, the phylogenetic position of the Beruwe-2008 strain inside this Central African clade differed between the two segments, lying at the most ancestral branch in segment S while sharing a more recent common ancestry with the Congo3010-1956 strain from Kisangani in segment M. This is highly suggestive of intra-genotypic reassortment, thus implying co-circulation of these two DRC sub-lineages at this time. However, though it may be less probable, we cannot exclude definitively recombination between the two strains.
Our dating analysis of the S segment resulted in time estimations slightly more recent than previously reported, but nonetheless within the same range and in keeping with an ancient origin of CCHF viruses [2]. The MRCA (most recent common ancestor) for the whole CCHFV species was estimated to have arisen 2518 years before present (BP) (95% High Posterior Density (HPD): 820-5281), the lineage II split-off was dated to 1484 years BP (95% HPD: 583-3389) and the MRCA of the three Central African strains was estimated at 587 years BP (95% HPD: 200-1327). In the M segment, the MRCA estimates were slightly more recent, most probably owing to the use of partial rather than complete coding sequences and to different evolution of the two genes. This resulted in MRCA estimates of 1955 years BP for the whole species (95%HPD: 886-3844), 221 years BP for the three Central African strains (95%HPD: 114-407) and 129 years BP (95% HPD: 75-228) for the two DRC strains. The genotype II split-off was estimated to have occurred 646 years BP, but the differences in the tree topologies prevented a true node age comparison with segment S.
CCHV genotype II has been identified only in DRC and Uganda, while different CCHV lineages have been identified in neighboring countries to the north. Multiple genotypes have been identified in CAR, belonging to groups IV and III [2], [20], the latter also being encountered in Sudan [14]. By contrast no other genotype has been identified in Central Africa, for which reports on CCHFV are scarce and date back to 30 years. Hence, the data currently available suggest that genotype II is specific to central Africa. In DRC, CCHV has been reported only once, 50 years ago, but our data strongly suggest that the same genotype is still actively circulating.
Of note, the MRCA estimates presented here are in agreement with ancient divergence of this lineage (around 1000 years ago), but whether or not this split-off was linked to virus adaptation to Central Africa cannot be assessed. However, as the MRCA of the three strains was dated back to 683 to 243 years BP (Figure 2A and B respectively), one might reasonably assume that the association of genotype II with this area goes back to this time period and thus did not result from very recent introduction. In addition, the co-circulation of different sub-lineages supports the possibility that ongoing CCHFV circulation occurred in the same area for some time. However, as the reassortment event would have taken place approximately 120 years BP, there is no evidence that CCHFV has been permanently circulating inside the Beruwe microhabitat, and we cannot exclude the possibility that this virus was very recently (re)introduced.
In addition to the CCHFV genotypic specificity for Central Africa, its occurrence in the tropical rainforest contrasts strongly with the ecological characteristics of other areas in which CCHFV has been isolated [11]. Indeed, the enzootic distribution of CCHFV mostly coincides with temperate to dry or semi-dry climates in the forests, steppes and savannahs of Eurasia and West, East and South Africa. In these environments, domestic animals and their associated ticks are major agents of rural enzootic cycles affecting nearby human populations [11], [25]. Despite the lack of accurate ecological data, the occurrence of CCHFV in Central Africa and its apparent genotypic specificity may suggest a distinctive sylvatic natural cycle in the deep tropical forest characterized by high rainfall, specific wildlife species, and a low density of domestic animals. Interestingly, co-speciation or long-term association with specific tick species has been previously suggested to explain the geographical distribution of CCHV genetic variants in Russia and Central Asia [26]. Such a sylvatic cycle, involving specific vectors and hosts with few contacts with human populations, could partly explain the lack of outbreaks and the sporadic nature of recorded human cases. In addition, as CCHFV is known to have been present in Central Africa for decades, and as human populations often live in isolated villages, many human infections may have been overlooked. However increasing invasion and destruction of rainforest habitats may lead to a higher risk of human CCHFV cases in future.
Hence, despite 30 years without a single reported case, the data presented here suggest that CCHFV continues to circulate in Central Africa. More information on the epidemiology and the natural cycle of CCHFV in this ecosystem is required to assess its potential for emergence, notably in Gabon and Republic of the Congo. However health authorities and medical staff should be aware of the possibility of viral (re)emergence and of the high risk of nosocomial transmission.
|
10.1371/journal.ppat.1007154 | Mononuclear cell dynamics in M. tuberculosis infection provide opportunities for therapeutic intervention | Mycobacterium tuberculosis causes chronic infection of mononuclear phagocytes, especially resident (alveolar) macrophages, recruited macrophages, and dendritic cells. Despite the importance of these cells in tuberculosis (TB) pathogenesis and immunity, little is known about the population dynamics of these cells at the sites of infection. We used a combination of congenic monocyte adoptive transfer, and pulse-chase labeling of DNA, to determine the kinetics and characteristics of trafficking, differentiation, and infection of mononuclear phagocytes during the chronic, adaptive immune phase of M. tuberculosis infection in mice. We found that Ly6Chi monocytes traffic rapidly to the lungs, where a subpopulation become Ly6Clo and remain in the lung vascular space, while the remainder migrate into the lung parenchyma and differentiate into Ly6Chi dendritic cells, CD11b+ dendritic cells, and recruited macrophages. As in humans with TB, M. tuberculosis-infected mice have increased numbers of blood monocytes; this is due to increased egress from the bone marrow, and not delayed egress from the blood. Pulse-chase labeling of dividing cells and flow cytometry analysis revealed a T1/2 of ~15 hrs for Ly6Chi monocytes, indicating that they differentiate rapidly upon entry to the parenchyma of infected lungs; in contrast, cells that differentiate from Ly6Chi monocytes turn over more slowly, but diminish in frequency in less than one week. New cells (identified by pulse-chase labeling) acquire bacteria within 1–3 days of appearance in the lungs, indicating that bacteria regularly encounter new cellular niches, even during the chronic stage of infection. Our findings that mononuclear phagocyte populations at the site of M. tuberculosis infection are highly dynamic provide support for specific approaches for host-directed therapies directed at monocytes, including trained immunity, as potential interventions in TB, by replacing cells with limited antimycobacterial capabilities with newly-recruited cells better able to restrict and kill M. tuberculosis.
| During certain chronic infections such as tuberculosis, inflammatory cells, including macrophages and dendritic cells, are recruited to infected tissues where they aggregate to form tissue lesions known as granulomas. Although granulomas can persist long term, the dynamics of the cell populations that comprise granulomas are not well understood. We used a combination of methods to discover that, during chronic infection of mice with Mycobacterium tuberculosis, the monocyte, macrophage, and dendritic cell populations are highly dynamic: recently-proliferated cells traffic rapidly to infected lung tissues, yet they persist with a half-life of less than one week. We also found that recently-proliferated cells become infected with M. tuberculosis as soon as one day after their arrival in the lungs, indicating that the bacteria are regularly moving to new cellular niches, even during the chronic stage of infection. The dynamic nature of the cell populations that encounter M. tuberculosis suggests that interventions such as trained immunity have potential therapeutic roles, by replacing cells that have poor antimycobacterial activity with cells with enhanced antimycobacterial activity. These interventions could improve the outcomes of treatment of drug resistant tuberculosis.
| Mononuclear phagocytes (MNP) harbor Mycobacterium tuberculosis in tissues of humans [1] and experimental animals [2–4]; and MNP are essential elements of granulomas, the characteristic tissue lesions in tuberculosis [5, 6]. Although macrophages have been characterized as prominent cellular hosts for M. tuberculosis in vivo, recent studies have revealed the roles of distinct populations of MNP, including tissue-resident (i.e., lung alveolar) macrophages, monocyte-derived recruited macrophages, and dendritic cells, as host cells for the bacteria in experimental animals [4] and humans [7]. Although cells in these subsets exhibit functional differences during M. tuberculosis infection, including the ability to transport bacteria from the lungs to the local lymph nodes [8–10] and their ability to present antigens for activation of CD4 T cells [11], there is little known regarding the population dynamics of MNP in tuberculosis or any other chronic infection.
Recent studies of blood monocytes that emigrate from the bone marrow during homeostasis have revealed the potential for these cells to differentiate from Ly6Chi monocytes to several distinct subsets of intravascular and tissue parenchymal cells. A proportion of Ly6Chi monocytes differentiate into Ly6Clo monocytes, which remain in the blood and vascular space of peripheral tissues, where they are considered to 'patrol' the vascular space and respond to inflammatory stimuli [12]. In addition, Ly6Chi monocytes emigrate from the vascular space during homeostasis and differentiate into lung macrophages and dendritic cells [13]. M. tuberculosis infection markedly increases accumulation of recruited macrophages and dendritic cells in the lungs [2, 4, 9, 14, 15], but it is unclear whether the recruited cells are long-lived, or whether they require continuous replenishment by recruitment, local proliferation, or both. Since M. tuberculosis infection is accompanied by apoptosis [16], necrosis [17], and egress from the lungs to the local lymph node [8–10], we hypothesized that mononuclear cell populations in the lungs are dynamic, and their abundance and differentiation may contribute to the outcomes of infection.
Initial evidence that MNP populations can be dynamic at a tissue site was obtained in studies by Dannenberg, who used intravenous injection of tritiated thymidine to track the recruitment and decay of MNP at sites of M. bovis BCG injection in the skin of rabbits [18]. Although these experiments revealed evidence for MNP turnover, they could not reveal the differentiation of the labeled cells, and the decay of labeled cells at the site of BCG injection may have been at least partly due to clearing of the attenuated bacteria used as the stimulus.
Here, we studied monocyte trafficking, differentiation, and turnover in multiple compartments: blood, the lung vasculature, lung parenchyma, lung-draining lymph nodes, and distant lymph nodes, in uninfected and M. tuberculosis-infected mice, using adoptive transfer and in vivo pulse-chase labeling, coupled with flow cytometry and immunostaining of tissue sections. We selected time points that followed the rapid accumulation of cells that occurs during the early response to infection and that followed the development of adaptive immune responses. Our results extend previously-published results [15] by revealing distinct characteristics of endovascular and tissue parenchymal cells, by revealing that the MNP population is actively turning over during the chronic phase of M. tuberculosis infection, and by finding that new MNP rapidly acquire bacteria at the site of infection. The findings that MNP populations are dynamic during the chronic stage of infection suggests that mechanisms that disrupt MNP trafficking or differentiation may underlie reactivation and progression of TB, and also indicate the potential to employ trained immunity [19] as a therapeutic intervention in tuberculosis, by replacing cells that support intracellular bacterial growth with new cells that are more capable of restricting bacterial growth.
To characterize the patterns and kinetics of trafficking and differentiation of monocyte-derived cells during the chronic phase of infection in a non-lymphoid tissue (the lungs), we purified Ly6Chi and Ly6Clo monocytes from the blood and bone marrow (BM) of CD45.1+ mice and transferred them intravenously (i.v.) to CD45.2+ recipients either 4 or 8 weeks after aerosol infection with M. tuberculosis H37Rv. Mononuclear cells were ~92–99% Ly6Chi monocytes; the remainder were Ly6Clo monocytes, dendritic cells (DC) and cells expressing markers of stem cell progenitors. Utilizing multi-color flow cytometry panels, live/dead discrimination and Boolean gating we were able to identify multiple subsets of live mononuclear cells and granulocytes from lungs of infected and uninfected mice (S1 and S2 Figs), we then characterized the adoptively-transferred CD45.1+ cells as they trafficked and differentiated in five compartments: peripheral blood, the lung vasculature, the lung parenchyma, the lung-draining mediastinal lymph nodes (MLN) and pooled peripheral lymph nodes that do not drain the lungs (pLN) (S1 and S3A Figs). Cells within the lung parenchyma were distinguished from intravascular cells by the i.v. administration of anti-CD45 antibody immediately prior to euthanasia. All alveolar macrophages (Siglec F+ CD11blo CD11chi SSChi CD103−) (AvM) were CD45 IV− and contained no CD45.1+ adoptively transferred monocytes at any of the time points we observed.
At the time of the earliest sampling (~40 h following transfer), the transferred cells in the blood and lung vasculature resembled the population of transferred cells: most were Ly6Chi monocytes (Ly6Chi CD11b+ CD11c− MHCII−); the remainder were Ly6Clo monocytes (Ly6Clo CD11b+ CD11clo MHCII−), while no other monocyte-derived cell subset was detectable in the blood or the lung vascular space (Figs 1A, 1B and S3A).
The composition of transferred mononuclear cell populations in the lung parenchyma and MLN were similar, but notably different between weeks 4 and 8 of M. tuberculosis infection. At the first time point after transfer into mice infected for 4 weeks, >50% of the transferred cells in these tissues had differentiated into Ly6Chi DCs (Ly6Chi CD11b+ CD11c+ MHCIIhi); the remainder of the transferred cells in these tissues were predominantly Ly6Chi monocytes, while a small fraction had differentiated into CD11b+ DC (Ly6Clo CD11b+ CD11c+ MHCIIhi) in both tissues or into Ly6Clo CD11b+ CD11clo MHCIIlo recruited parenchymal macrophages (RPM) in the lungs. Transferred cells differentiated almost identically in the MLN when transferred at 4 and 8 weeks post infection. However, at 8 weeks post M. tuberculosis infection, both Ly6Chi DC and CD11b+ DC fractions were diminished in the lung parenchyma while the portion of transferred monocytes that became RPM was substantially increased.
While both Ly6Clo monocytes and RPM express the phenotypic markers previously used to define lung recruited macrophages [4], they differ in their location within the lung vasculature or parenchyma. In the steady state, ~10% of the Ly6Clo monocytes/RPM are within the lung parenchyma (S3B Fig). However, as disease progresses, an increasing proportion of these cells are in the parenchyma, which is further evidenced by the increased number and proportion of donor-derived RPM during the later phase of infection (Fig 1A).
Recent publications have characterized heterogeneity within both the Ly6Chi and Ly6Clo monocyte populations in the bone marrow and circulation. Ly6Clo monocytes are believed to be derived from Ly6Chi monocytes, and in the steady state, the majority of the Ly6Clo monocytes differentiate into I-Ab−vascular-patrolling monocytes in an NR4A1-dependent manner [20, 21]. Ly6Chi monocytes, on the other hand, have been shown to differentiate into monocyte-derived DCs and tissues macrophages in an NR4A1- and Flt3L-independent manner, with high expression of the transcription factor PU.1 predisposing these cells to differentiate into DC [22]. Intranuclear staining for the PU.1 revealed that DC populations in the lung parenchyma expressed the highest levels of PU.1 amongst mononuclear cells, while RPM expressed the lowest levels of PU.1 amongst mononuclear phagocytes (S3C Fig). These data support the assertion that Ly6Chi monocytes contribute to bona fide DC and RPM cell subsets in the lungs of M. tuberculosis-infected mice.
In contrast to the MLN, Ly6Chi monocytes were the dominant transferred cell subset in the pLN week 4 after infection, and only a small fraction of the cells were Ly6Chi DC or CD11b+ DC. However, by week 8 of M. tuberculosis infection, transferred cells in the pLN differentiated into Ly6Chi DC, suggesting that the inflammatory environment within pLN is distinct from MLN 4 weeks post infection, but the inflammatory milieu is more similar by week 8 of infection. The presence of M. tuberculosis in the pLN has been previously reported [10] and the inflammation from infection and the effector cytokines produced by M. tuberculosis-reactive T cells may contribute to the differentiation of Ly6Chi DC in the LN. These results indicate that circulating bone marrow-derived monocytes enter an infected peripheral tissue and differentiate rapidly, and that differentiation into DC or RPM happens during or after egress from the vascular space in the lungs of M. tuberculosis-infected mice.
At later time points after transfer during both phases of M. tuberculosis infection, donor cells underwent further differentiation in tissue-dependent patterns, followed by gradual declines in their abundance in each compartment. In the peripheral blood and lung vascular space, Ly6Chi monocytes progressively became less frequent as a fraction of the recovered cells, with nearly all of the remaining transferred cells present as Ly6Clo monocytes 6–8 days after transfer (Fig 1A). In both of these compartments, the number of Ly6Clo monocytes exhibited an inverse pattern relative to that of Ly6Chi monocytes, with Ly6Clo cells reaching a peak by days 3–4 followed by a gradual decline to approximately 30% of the peak level by day 8. This pattern is consistent with differentiation of Ly6Chi to Ly6Clo monocytes in the blood, as previously reported [12].
Over time, Ly6Chi and CD11b+ DC exhibited an inverse relationship in the lung parenchyma and MLN. The total number of Ly6Chi DC and their proportion of transferred cells continuously ebbed over time while CD11b+ DC numbers increased until day 4 before diminishing in number, while also constituting the majority of transferred cells. Surprisingly, although the kinetics of Ly6Chi monocytes and CD11b+ DC turnover were similar in the pLN and MLN, only a small fraction of transferred cells differentiated into Ly6Chi DC in the pLN.
The pattern and kinetics of adoptively transferred monocyte trafficking and differentiation in the MLN were similar at later phases of infection (8 weeks post infection), though, in the lung parenchyma and in pLN there was a delay in the peak number of Ly6Chi DC to nearly 3 days after transfer (Fig 1). Transferred monocytes also differentiated into CD11b+ DC more rapidly during this later phase of infection, peaking at day 3 instead of 4 days after transfer. The most striking difference between the two infectious phases, however, was the greater increase in RPM in the lung parenchyma at the later phase of infection. This suggests the inflammatory environment within the lung is changing throughout infection, leading to an evolving population of mononuclear cells with different phenotypes.
We determined the half-life of the Ly6Chi monocytes as well as the total transferred CD45.1+ mononuclear cell population in each of the 5 compartments examined by fitting a first order exponential decay function to the data on the cell numbers identified by flow cytometry. There was substantial variation in the R2 coefficients of determination and in the half-lives of the CD45.1+ cells in each compartment and we utilized a comparison of fits analysis to determine the quality of each equation for each tissue. This analysis operates with a null hypothesis that the data are best represented by an equation with the same slope for all the cell numbers across all tissues combined. This analysis determined that the null hypothesis was correct for the total CD45.1+ cells and the best-fit line for all the compartments combined yielded a consensus half-life of 112.18 hours (S3D and S3E Fig). In contrast, the decay of Ly6Chi monocytes (by differentiation, egress, death, or other mechanisms) was more rapid than total transferred cells and appeared most rapid in the pLN and was similar to that in the lung vasculature and parenchyma. However, when the qualities of the non-linear equations for each tissue were analyzed by comparison of fits analysis, it was determined that a single best-fit equation for all the Ly6Chi monocytes in all tissues was a better fit than equations for each tissue and this best-fit equation yielded a half-life of 15.31 hours (S3D and S3E Fig). Together, these results indicate that Ly6Chi monocytes have a half-life of less than one day, while cells that differentiate from Ly6Chi monocytes have a half-life of approximately 5 days during the chronic stage of infection with M. tuberculosis.
The results of these monocyte adoptive transfer studies are consistent with studies of the homeostatic state in mice indicating that Ly6Chi monocytes that egress the bone marrow can differentiate into Ly6Clo monocytes that remain in the blood and vascular space for prolonged periods [12]. In addition, they reveal that during the chronic phase of M. tuberculosis infection, monocytes that enter the lung parenchyma and lung-draining MLN preferentially differentiate into Ly6Chi DC, followed by CD11b+ DC. In contrast, monocytes that traffic to a lymph node distant from the lungs undergo more diverse pathways of differentiation, suggesting that the local tissue environment contributes to monocyte differentiation into distinct cell subsets.
Our adoptive transfer experiments established the kinetics and potential for monocytes to traffic and differentiate in distinct compartments in the context of chronic infection with M. tuberculosis but could not provide insight into the quantitative contributions of monocytes compared with resident cells in the lungs or lymph nodes during infection. To further understand the dynamics of mononuclear cell trafficking, proliferation, and differentiation, we briefly pulsed uninfected and M. tuberculosis-infected mice at multiple stages of chronic infection with the nucleoside EdU (5-ethynyl-2’-deoxyuridine) and characterized monocytic cells with EdU-labeled DNA at multiple time points in the bone marrow, blood, lung vascular space, lung parenchyma, and MLN. EdU incorporation was compared to the staining of cells from infected or uninfected mice that did not receive EdU injections (S4 and S5 Figs).
Quantitation of total Ly6Chi cells in the bone marrow revealed no differences between uninfected mice and M. tuberculosis-infected mice, at any stage of chronic infection (up to 16 weeks) (Fig 2A and S1A Table). In contrast to the bone marrow, in the blood we found that chronic M. tuberculosis infection was associated with increases in Ly6Chi monocytes, with the most marked increase observed at the latest time point of infection examined (16 weeks post infection) (Figs 2A, S6A and S1A Table). The finding in the blood is consistent with observations with human tuberculosis patients, in which peripheral blood monocytosis is associated with active tuberculosis and with the severity of tuberculosis in children [23]. We also observed increased ratios of blood Ly6Chi monocytes/total lymphocytes (S6B Fig) consistent with recent reports of increased monocyte/lymphocyte ratios in the blood of patients with active TB [24] or at high risk of progression to active TB [25].
The finding of increased numbers of blood monocytes during infection without an increase in the number of mature monocytes at their site of production in the bone marrow is consistent with a shorter dwell time of mature monocytes in the bone marrow, and/or a longer dwell time in the blood. Analysis of the frequency of labeled monocytes after a brief pulse of EdU revealed a significant increase in the rate of decay or egress of monocytes from the bone marrow associated with infection (Fig 2B and 2C). The statistical differences in %EdU+ and total numbers of EdU+ Ly6Chi monocytes between uninfected or M. tuberculosis-infected mice are summarized in S1B and S1C Table. The half-life of Ly6Chi monocytes in the BM decreased from 47.59 h in uninfected mice to 26.06 h at the most acute phase of infection examined, though, the half-life had increased to 34.27 h by the 16th week of infection (S6C Fig and S2A Table). We also determined the half-lives of Ly6Chi monocytes by fitting the equation to the total numbers of EdU+ Ly6Chi monocytes; this revealed a similar trend of increased monocyte turnover during M. tuberculosis infection (S6D Fig and S2B Table). These results suggest that acute and chronic M. tuberculosis infection are associated with increased production of monocytes in the bone marrow, which is balanced by increased egress from the bone marrow to the blood and to peripheral tissues. This interpretation was supported by analysis of the frequency of EdU-labeled monocytes in the blood. First, comparison of the frequency of EdU pulse-labeled monocytes in the blood of uninfected and infected mice revealed a lag before EdU-labeled monocytes peaked in the blood of uninfected mice; this lag was not evident in the blood of M. tuberculosis-infected mice, consistent with more rapid entry of monocytes into the blood from the bone marrow (Fig 2B and 2D and S1B Table). Second, the rate of decay of the frequency of EdU-labeled monocytes from the blood was greater in infected mice (ranging from 17.6–23.6 h) compared with that in uninfected mice (34.1 h) (Figs S6C and 2C, and S2A Table). Together, these results show that increased numbers of monocytes in the blood of M. tuberculosis-infected mice result from increased production of monocytes in the bone marrow, balanced by increased release of mature monocytes from the bone marrow into the blood, and not by prolonged retention of monocytes in the blood.
Because the total numbers of Ly6Chi monocytes are substantially different between phases and a logarithmic scale makes it difficult to observe the change in EdU+ monocytes over time, we sought to normalize the number of new monocytes and display them on a linear scale. We determined the frequency of EdU+ Ly6Chi monocytes amongst total live cells in the lungs, determined the peak frequency of EdU+ monocytes of live cells and then calculated the % of maximum EdU+ monocytes (Fig 2D). This allowed all phases to be shown on a linear scale and helped normalize for variations in total lung cell yields between time points, as well as more clearly demonstrating the turnover of monocytes.
Ly6Chi monocytes accumulated in the peripheral tissues during M. tuberculosis infection, with ~30-fold increase in the lung vasculature over that in uninfected mice (S3 Table). The kinetics of EdU pulse-labeled monocytes in the lung vascular space closely mimicked those in the blood, including the delay to the peak in the frequency of labeled Ly6Chi monocytes in uninfected mice that was absent in infected mice (Fig 2B and 2D), and the shorter half-life of monocytes in blood of infected mice.
In the lung parenchyma, M. tuberculosis infection was accompanied by a striking increase in the total number of monocytes, as previously reported [4, 15] the number of Ly6Chi monocytes in the lung parenchyma progressively increased between 4 and 16 weeks of infection where they were more than 550 times more numerous than in uninfected mice (Fig 2A and S3 Table). The peak in the frequency of EdU pulse-labeled monocytes was lower in the lung parenchyma of infected mice compared with that of uninfected mice; this is most likely due to the larger population of monocytes in the lungs of infected mice into which the EdU+ cells are diluted. After peaking on day 2 post EdU pulse, the rate of decay of Ly6Chi EdU+ cells was increased in M. tuberculosis-infected mice (Figs 2B, 2C, S6C and S6D and S2 Table).
The behavior of Ly6Chi EdU+ monocytes in the lung-draining MLN closely mimicked those in the lung parenchyma: the number of cells in both tissues increased markedly and progressively with M. tuberculosis infection, and the kinetics of appearance and decay of Ly6Chi EdU+ cells were similar. In both tissues, infection accelerated the rate of decay of Ly6Chi cells, and by 7 days after the EdU pulse, the number of Ly6Chi cells decreased 10 to 100-fold compared to the peak value reached at 2 days post-pulse. During both acute and chronic phases of M. tuberculosis infection, Ly6Chi monocytes rapidly proliferate in the BM, and enter the periphery to differentiate into various cell subsets at an increased rate relative to uninfected mice.
Since the quantitative decay of Ly6Chi EdU+ monocytes in tissues can be due to differentiation, egress, cell division, and/or death, we analyzed the differentiation of EdU+ cells over time in the lung vasculature and parenchyma of uninfected and M. tuberculosis-infected mice. Total numbers of Ly6Clo monocytes in the peripheral blood were modestly increased only at the later chronic phase of M. tuberculosis infection that we examined (S7A Fig and S4 Table). In the lungs, however, the numbers of the Ly6Clo CD11b+ CD11clo MHCIIlo subset we have previously characterized as recruited macrophages [4, 9, 11] and Ly6Clo monocytes (CD45 IV–) continuously increased in the lungs of M. tuberculosis-infected mice as disease progressed (Figs 3A and S7A). As with our adoptive transfer data, the majority of these cells were in the lung vasculature but the proportion of these cells that were in the parenchyma as RPM increased from ~10% in uninfected mice to ~25% in mice infected with M. tuberculosis for 16 weeks (S3B Fig).
By 24 hours after EdU pulse, only a small fraction of these Ly6Clo monocyte and RPM subsets were EdU+, matching the delayed differentiation of Ly6Chi monocytes to Ly6Clo monocytes and RPM we observed in our adoptive transfer experiments (Figs 3B, 3D, S7B and S7D). The statistical differences in %EdU+ and total numbers of EdU+ Ly6Clo monocytes and RPM between uninfected or M. tuberculosis-infected mice are summarized in S4 Table. The peak in the frequency of EdU+ RPM appeared later after the EdU pulse than did the peak of Ly6Chi cells in the lung parenchyma (3 d vs 2 d, respectively), consistent with Ly6Chi cells as precursors of at least a fraction of the RPM. Although the peak of EdU labeling was delayed in these cells, EdU+ RPM declined in frequency and total number, indicating that this cell population undergoes continuous turnover during both acute and chronic phases of infection. EdU+ RPM exhibited several differences later in infection (16 weeks) compared with the pattern 4–8 weeks post infection; the peak frequency of EdU+ RPM was delayed and significantly lower than in other phases of infection. At this time point, the total number of RPM and the number of EdU+ RPM was higher than observed 4 or 8 weeks post infection. Together, these results indicate that the dynamics of mononuclear cells at the site of infection change during the chronic phase of infection, resulting in a larger population of RPM that turns over more slowly and comprises a larger proportion of the parenchymal cells (S7E and S7F Fig).
Consistent with the data from adoptive transfer of bone marrow monocytes, the EdU pulse labeled up to 10–15% of the Ly6Chi DC in the lung parenchyma with peak labeling observed 3 days after the EdU pulse (Fig 3B and 3C). The statistical differences in total cell number, %EdU+ and total numbers of EdU+ in the multiple DC and MP populations in the lung parenchyma of uninfected or M. tuberculosis-infected mice are summarized in S5 Table. Also consistent with the results of adoptive transfer, Ly6Chi DC are transient, with the proportion of EdU+ cells decaying to ~30% of the peak level one week after EdU pulse. Furthermore, these cells did not accumulate with disease progression as observed with other mononuclear cell populations (RPM, CD11b+ DC, CD103+/XCR1+ DC), displayed similar kinetics of EdU incorporation and decay between infection phases, and were undetectable in the lungs of uninfected mice. These cells likely represent a transient transition state from Ly6Chi monocytes to CD11b+ DC or RPM and are only detectable during infection.
CD11b+ DC progressively increased in number within the lung parenchyma of M. tuberculosis-infected mice and EdU+ CD11b+ DC reached a peak frequency ~3 days after the EdU pulse (Fig 3B). A substantial fraction of CD11b+ DC were EdU+ within 24 h of EdU administration, suggesting that a large fraction of these cells can proliferate locally in the lungs, as previously reported [26], while the rest of the EdU+ DCs are derived from the differentiation of monocytes or other subsets of recruited cells (Fig 3C). The frequency of CD11b+ DC that were EdU+ was continuously diminished as disease progressed, conversely, the total numbers of EdU+ CD11b+ DC increased with infection progression. This is largely due to the accumulation and retention of increasing numbers of CD11b+ DC over time during infection, and while recently proliferated cells comprising a smaller portion of the total number, they still represent a sizable population of cells that is turning over (Fig 3C and 3D).
We examined two additional subsets of mononuclear phagocytes in the lung parenchyma: CD103+/XCR1+ DC and alveolar macrophages (AvM). For some of the experiments we utilized antibodies specific for XCR1 instead of CD103 as XCR1+ DC have been shown to largely consist of CD103+ CD11b– DC in the LN and peripheral organs [27, 28]. Up to 12–15% of the CD103+/XCR1+ DC in the lung parenchyma were labeled with an EdU pulse; unlike the other cell subsets studied, the frequency of EdU+ CD103+/XCR1+ DC did not differ in uninfected compared with infected mice, although the total number of CD103+ DC and the number of EdU+ CD103+/XCR1+ DC increased in a time-dependent manner with M. tuberculosis infection. Notably, the peak of EdU-labeled CD103+/XCR1+ DC did not exhibit the lag observed with RPM, Ly6Chi DC, or CD11b+ DC (Fig 3), suggesting that this subset of cells in the lung parenchyma does not depend on differentiation from a cell with another phenotype, and that CD103+/XCR1+ DC proliferate locally in the lungs, even in the absence of infection, as previously reported [26]. Compared with the other cell subsets examined, a smaller fraction (<10%) of AvM incorporated EdU, and the frequency and number of EdU+ AvM was unaffected by M. tuberculosis infection at any time point examined. Also, unlike the other cell subsets examined, there was no detectable increase in the total number of AvM in lungs of infected mice compared with uninfected mice (Fig 3). The majority of mononuclear cell subsets in the lung parenchyma of infected mice had low (1–2%) expression of Ki67, while ~10% of CD103+/XCR1+ DC were Ki67+ (S7G Fig). Thus, the majority of professional phagocyte subsets we identified were derived by differentiation from Ly6Chi monocytes, while CD103+/XCR1+ DC and AvM proliferated in situ with minimal contribution from monocytes.
Finally, we calculated the total number of potential monocyte precursors within the lung vasculature and compared it to the total number of putative monocyte-derived populations in the parenchyma. We determined the area-under-the-curve for lung vasculature Ly6Chi and Ly6Clo monocytes as well as DCs throughout 7 days post EdU pulse for each phase and compared it to the area-under-the-curve for parenchymal monocytes, RPM, Ly6Chi DC, CD11b+ DC and CD103+/XCR1+ DC (S7H Fig). The turnover of circulating EdU+ precursors of vascular monocytes, MP and DC populations was greater than the combined number of these parenchymal populations in uninfected mice and at all phases of M. tuberculosis infection. Thus, the turnover of recently-proliferated vascular monocytes is sufficient to account for the turnover in monocyte-derived cells in the lung parenchyma.
We further characterized the differentiation of recently-proliferated monocytes and their descendants into cells with antigen-presenting potential by comparing the expression of MHCII in new (EdU+) cells relative to all cells of each subset. Ly6Chi monocytes within in the lungs increased their expression of MHCII as they moved from the vasculature into the parenchyma and differentiated into Ly6Chi and CD11b+ DC (Fig 4). At 12h after an EdU pulse, in the lung parenchyma and vasculature EdU+ Ly6Chi monocytes expressed significantly less surface MHCII than the total population of Ly6Chi monocytes. New Ly6Chi DC, CD11b+ DC and CD103+ DC also had significantly lower expression of MHCII when compared to total MHCII expression of these subsets. However, by 24h all populations of new EdU+ had upregulated MHCII to the same level as their respective total population. These data support previous work demonstrating the upregulation of MHCII by Ly6Chi monocytes as they transition from the vasculature into peripheral tissues [29].
The lung-draining lymph nodes play pivotal roles in TB immunity. Live M. tuberculosis bacteria are transported from the lungs to the local lymph nodes by DC [10, 30], where infected migratory DC transfer antigens to uninfected resident DC to initiate T cell priming [8, 9]. The M. tuberculosis population within the MLN also represents a substantial bacterial burden [10, 31] and in humans may act as reservoirs of the pathogen for activation of latent TB [32]. Priming and development of CD4 T cell responses in the MLN is key to the eventual control of M. tuberculosis replication within the lungs, but also necessary for generating the proinflammatory adaptive immune responses that facilitate active TB disease and transmission. Understanding the relationships of antigen presenting cells within the MLN with those in the lungs may be essential for optimal development of effective vaccines or pharmaceutical treatments of TB.
We identified multiple mononuclear cell subsets in the MLN following M. tuberculosis infection and characterized their incorporation of EdU over time (S8A Fig). M. tuberculosis infection causes large increases in the population size of several mononuclear cell subsets in the MLN (Fig 5A and S6 Table. The marked expansion of mononuclear cells affects at least 6 distinct subsets, develops by 4 weeks post-infection, and persists for at least 16 weeks (Fig 5A).
Amongst all myeloid cells in the MLN that contain bacteria, Ly6C− CD11bhi CD11chi MHCII+ cells constitute the largest population infected with M. tuberculosis at all time points examined (S8B Fig), in agreement with previous studies that were limited to early time points [4, 9]. There were also minor populations of Ly6Chi DC (CD11b+ CD11c+ MHCII+) infected with M. tuberculosis, while few of the CD11b− DC (Ly6C− CD11blo CD11c+ MHCIIhi and Ly6C− CD11blo CD11chi MHCII+) [9] contained bacteria as detected by flow cytometry.
Ly6C− CD11blo CD11chi MHCII+ DC rapidly incorporated EdU following the pulse (Fig 5), suggesting they proliferated within the MLN and supporting their classification as resident DC. The other mononuclear cell subsets examined all showed increasing EdU incorporation over time, albeit with different kinetics between populations and infection phases. Furthermore, each of these populations had different levels of EdU incorporation 24 h post pulse, potentially reflecting varied contributions of proliferation within the MLN. Of the mononuclear cell populations subsets examined, Ly6Chi DC decayed most rapidly, consistent with their differentiation into other DC subsets. Other mononuclear subsets also decayed in the MLN during the 7 days following the EdU pulse, indicating that their populations remain dynamic as late as 16 weeks post-infection.
Mononuclear cells present in the MLN during M. tuberculosis infection might arrive from the blood via high endothelial venules (HEV) or from the lungs through afferent lymphatics [33, 34]. Monocytes expressing CD62L bind to heavily glycosylated molecules expressed on HEV and enter into lymph nodes from the blood [29, 35]. To determine the contribution of blood-derived, HEV-recruited monocytes to the mononuclear cell subsets present during M. tuberculosis infection we blocked CD62L in conjunction with an EdU pulse to identify recently-proliferated and/or recruited cells (S8C Fig).
As expected, there were profound reductions in total cell numbers in the MLN and peripheral LNs following anti-CD62L treatment (S8D Fig). However, there were no significant differences in the total numbers of Ly6Chi monocytes in any of the compartments we evaluated (S8E Fig). To compare the relative contribution of HEV-migration on monocytes and mononuclear cell subsets within the LN that had drastically different total cell numbers we evaluated the relative fraction of EdU incorporation between anti-CD62L-treated and control mice. There was a small but significant reduction in the frequency of EdU+ Ly6Chi monocytes in the MLN following CD62L blockade while there were no differences in the BM, blood, pLN or lungs (S8F Fig). In contrast to the small reduction in proportion of new Ly6Chi monocytes in the MLN, there were no differences in EdU incorporation in Ly6Chi DC, total CD11b+ DC in the lungs or in the multiple mononuclear cell subsets in the lymph nodes (S8G–S8I Fig). Thus, a small fraction of Ly6Chi monocytes are directly entering into the MLN, while the majority of these cells arrive from the lungs through the afferent lymphatics. The vast majority of potential M. tuberculosis-harboring mononuclear subsets within the MLN are derived from cells that have passed through the lung lymphatics or proliferated within the MLN.
Since only a small minority of the cells that accumulate at the site of M. tuberculosis infection in the lungs harbor bacteria [4], we investigated whether the bacteria remain in the same cells for long periods, or whether they are acquired by new cells. We took advantage of EdU labeling to identify new cells and to localize and quantitate the presence of fluorescent protein-expressing bacteria in the recently-proliferated cells.
Immunofluorescence microscopy of lung sections revealed that by day three following the EdU pulse recently-proliferated (EdU+) CD11b+ professional phagocytes are evident within granulomas containing GFP-expressing M. tuberculosis, 4 weeks (Fig 6A) or 8 weeks (S9 Fig) after infection. Consistent with the evidence from flow cytometry studies (Fig 3B), EdU+ CD11b+ cells were most apparent at day 3 post pulse and became subsequently less abundant in lung lesions over time (S9 Fig). This timing corresponds to the peak numbers of EdU+ CD11b+ cells in the lung parenchyma, which were comprised of both mononuclear cell subsets and neutrophils by flow cytometry (Figs 3, S7 and S10A). As M. tuberculosis infection progressed, CD11b+ DC increasingly became the dominant cell type containing bacteria in the lung parenchyma, particularly at the expense of Ly6Chi DC and to a lesser extent, neutrophils (S10B Fig). Recently-proliferated CD11b+ cells were also found to contain GFP-expressing M. tuberculosis in the lung-draining MLN (Fig 5B) after infection. Although EdU+ nuclei within CD11b+ lesions are in close proximity to GFP+ bacilli (Figs 5A, 5B and S9), the indistinct cell boundaries in the lung sections precluded quantitation of new, EdU-labeled CD11b+ cells that have acquired bacteria.
To quantitate new (EdU pulse-labeled) cells that become infected soon after they appear in the lungs, we used flow cytometry to detect the coincidence of EdU labeling and M. tuberculosis H37Rv-mCherry in the cell subsets of interest. Because the numbers of EdU+ Rv+ cells in the MLN were so infrequent, we were only able to quantify newly infected cells within the lung. Examination at either 4 weeks or 16 weeks post infection revealed that a fraction of the EdU+ Ly6Chi monocytes, Ly6Chi DC, CD11b+ DC, RPM, neutrophils and AvM contained bacteria as early as one day after the EdU pulse (Figs 5C, 5D, S10C and S10D). At both time points post infection, the frequency of infected EdU+ cells increased progressively in each of these cell subsets, up to 7 days post EdU (Figs 5C and S5C).
We also characterized the dynamics of infection with an alternative approach, by determining the fraction of bacteria-containing cell subsets that were comprised of EdU+ new cells. Most infected cell subsets contained increasing amounts of EdU+ cells over the week post EdU pulse, with up to 10–20% recently proliferated cells at both 4 and 16 weeks after infection (Figs 5E and S10E). In contrast, M. tuberculosis-infected Ly6Chi monocytes week 16 post infection and neutrophils at both infection phases had a substantially higher proportion of new cells, up to 50% and 40% EdU+ respectively. Instead of continuously increasing, the proportion of these infected cells peaked in EdU staining before rapidly diminishing to 20–25% of the peak the following day. The large fraction of these particular subsets that is composed of new cells and the drastic loss of EdU staining is indicative of their relative short half-lives within granulomas before differentiating, migrating or dying.
These data demonstrate that newly-proliferated and -differentiated mononuclear cells are constantly trafficking to the infected lung where they are able to enter granulomas during the early and the later phase of infection. Importantly, these results also reveal that new mononuclear cells become infected shortly after their appearance at the site of infection, indicating that while the overall bacterial population remains at a steady state, the bacteria themselves are regularly entering new cellular environments. Since these data are only able to reveal the acquisition of bacteria by the new/EdU+ cells which are a small fraction of the total population at a given time, our results likely underestimate the frequency of bacterial transfer from cell to cell that is ongoing at the site of infection, even after development of adaptive immune responses.
In this work, we advanced the understanding of host cell and bacterial dynamics during the chronic stage of M. tuberculosis infection, after the development of adaptive immunity. Consistent with reports in sterile inflammation models, we found that Ly6Chi blood monocytes differentiate into multiple subsets of dendritic cells and macrophages in the lungs and lymph nodes of M. tuberculosis-infected mice. Using a recently-described method of distinguishing cells that reside in the vascular space from those in the tissue parenchyma, we determined that Ly6Chi monocytes differentiate into Ly6Clo monocytes that remain in the lung vascular space in M. tuberculosis-infected mice. Using the same method, we determined that differentiation of Ly6Chi cells into other subsets happens after egress from the vascular compartment, without transitioning through a Ly6Clo state. In particular, we determined that Ly6Chi monocytes rapidly differentiate into Ly6Chi CD11c+ MHCII+ population of DC, which is a transient intermediate state followed by differentiation into CD11b+ DC in the lungs. Although our evidence is indirect, we found kinetic evidence that both Ly6Chi and CD11b+ DC migrate and transport bacteria from the lungs to the lung-draining mediastinal lymph node, consistent with previous reports [8–10].
We also determined that the response to M. tuberculosis infection in mice includes peripheral blood monocytosis, as observed in children and adult humans with TB [23–25], and using pulse-chase DNA labeling, we determined that peripheral blood monocytosis is due to enhanced production and egress from the bone marrow and not due to delayed egress from the blood. Pulse-chase DNA labeling further confirmed the capacity and kinetics of differentiation of monocyte-derived cells and established that the monocyte-derived professional phagocyte populations in the lungs and lymph nodes of M. tuberculosis-infected mice are dynamic and turn over with a half-life of less than one week. The finding that the monocyte-derived cell subsets that harbor M. tuberculosis turn over frequently suggested that the bacteria must exchange their cellular niches frequently, and we confirmed this prediction by detecting M. tuberculosis in newly-arrived cells in the lungs, indicating that the bacteria must respond regularly and rapidly to distinct cellular environments.
Although the dynamics of mononuclear cells in localized mycobacterial infections has gotten little recent attention, there are precedents in other animal models, and evidence to suggest similar events in humans. Nearly 50 years ago, administration of radiolabeled thymidine was used to characterize mononuclear cell recruitment and death at skin sites of injection of M. bovis BCG in rabbits, and revealed that in contrast to prior assumptions, granuloma mononuclear cell populations are dynamic [18, 36]. However, at the time those experiments were performed, it was not possible to distinguish distinct subsets of mononuclear cells or their differentiation other than by their morphology. More recent studies used adoptive transfer of bone marrow monocytes during M. tuberculosis infection of mice, and revealed tissue-dependent differentiation into multiple subsets of macrophages and dendritic cells, with distinct patterns of differentiation in the lungs and the lung-draining lymph nodes [15]. Our findings extend those results by revealing that Ly6Chi CD11c+ cells appear after Ly6Chi monocytes enter the lung parenchyma and are transient intermediates that subsequently differentiate into CD11chi CD11b+ dendritic cells. This observation helps reconcile apparently different results of analysis of the cells that transport M. tuberculosis from the lungs to the lung-draining mediastinal lymph node: one study attributed this property to inflammatory monocytes [8], while another reported that CD11b+ dendritic cells are responsible for bacterial transport [9]. Our present results indicate that the inflammatory monocytes credited with bacterial transport likely differentiate into Ly6Chi CD11c+ dendritic cells after entering the lung parenchyma, where they acquire bacteria and differentiate further into CD11b+ dendritic cells, which contain bacteria and appear in the draining lymph node [9, 10].
Other recent studies have emphasized the importance of tissue-dependent differentiation of monocytes in tuberculosis, and certain of these have revealed important functional differences of monocyte-derived cell subsets. In particular, administration of poly-ICLC, a potent inducer of type I interferon expression, impaired control of M. tuberculosis in the lungs of mice, due to CCR2-dependent recruitment of monocyte-derived cells that are especially permissive for intracellular bacterial survival and/or growth [14]. In contrast, a recent study reported that interstitial macrophages that depended on CCL2 for their recruitment, are more able to constrain M. tuberculosis fitness as reflected by use of fluorescent reporter plasmids [2]. These seemingly contradictory results emphasize the potential plasticity of cells that differentiate from monocytes, which is likely the product of early programming of monocytes as well as of signals for differentiation and activation at the site of infection. Notably, studies in zebrafish embryos infected with virulent M. marinum have revealed that recruited macrophages are less able to control intracellular mycobacterial growth compared with tissue-resident macrophages [37]. As bacterial spread to recruited macrophages was a function of bacterial phenolic glycolipid, these results indicate that pathogen virulent factors can also act by directing bacteria to cells that are poorly equipped to kill them.
Although studies similar to ours have not been performed in humans or with human samples, it is noteworthy that cell subsets comparable to the ones studied here can be detected in human lung tissue [38]. Likewise, recent studies of human monocytes have revealed a comparable pattern of cell turnover and differentiation [39], indicating that in humans with tuberculosis, cell populations localized to the site of infection in granulomas may exhibit dynamic states of turnover and differentiation similar to those that we have found in M. tuberculosis-infected mice.
The finding that mononuclear cell populations are in a dynamic state during the chronic stage of M. tuberculosis infection may have clinical value. First, our findings that MNP populations are dynamic, even during the chronic phase of M. tuberculosis infection suggests that mechanisms that perturb the dynamics of MNP trafficking and differentiation may underlie or contribute to the transition from latent TB infection to active TB disease and/or the severity of TB disease. Evidence consistent with this hypothesis is that pulmonary TB in patients with diabetes mellitus, which has poorer outcomes compared with those of pulmonary TB in those without diabetes [40–42] is characterized by decreased frequencies of blood monocyte and dendritic cell populations compared with those in patients without diabetes [43]. Second, a recent study revealed that 'trained hematopoietic stem cells' that produce epigenetically-modified monocytes and macrophages are capable of contributing to control of M. tuberculosis infection in mice, in a manner that is independent of T lymphocytes [19]. Our finding that recruited macrophage populations are dynamic and that newly-arrived cells have access to bacteria even during the chronic stage of infection, implies that modification of cell populations to better restrict and/or kill M. tuberculosis may have therapeutic value in humans with drug-resistant tuberculosis, and may shorten the length of effective treatment, even in those with drug-susceptible tuberculosis.
C57/BL6 mice were bred in the New York University School of Medicine (New York, NY) animal facilities or purchased from The Jackson Laboratory and maintained under specific pathogen-free conditions. Ptpcra (CD45.1) mice utilized in adoptive transfer experiments were either bred in the New York University School of Medicine animal facilities or purchased from Taconic Farms, Inc. Mice were infected with M. tuberculosis at 8–12 weeks of age. Uninfected or M. tuberculosis-infected mice were twice injected intraperitoneally with 2 mg of EdU (Thermo Fisher Scientific) in 400 ul of PBS, two hours apart.
All experiments were conducted in accordance with procedures approved by the NYU School of Medicine Institutional Animal Care and Use Committee and in accordance to the recommendations in the Guide for the Care and Use of Laboratory animals of the National Institutes of Health, operating under Animal Welfare Assurance number D16-00274. The protocol approval number was s15-01412.
M. tuberculosis (H37Rv) derivatives used to infect mice constitutively expressed either EGFP or mCherry. Bacteria were grown in Middlebrook 7H9 medium with 10% (v/v) albumin dextrose catalase (ADC) enrichment, 0.05% Tween80 and 50 μg/ml kanamycin or 25μg/ml hygromycin. Mice were infected via aerosol utilizing an inhalation exposure unit from Glas-Col. Mid-log cultures of M. tuberculosis were pelleted at 4000 g, resuspended in 7ml of PBS+0.05% Tween 80 and then serially centrifuged at 800 g to remove clumps. Clump-free cultures were then diluted to 1.5x106 in dH20 and 5 mL of the inoculum was added to the nebulizer. Mice were infected with a program of 900s of preheating, 2400s of nebulization, 2400s of cloud decay, and 900s of decontamination. For controls of fluorescent-protein-expressing M. tuberculosis, mice were infected with wild-type H37Rv by the same procedure on the same day. Infection dose was determined by euthanizing mice within 24 h of infection, plating lung homogenates on 7H11 agar plates, and counting CFU within 2–3 weeks of incubation at 37° C.
Mice were anesthetized by inhalation of 10% Isoflurane in mineral oil (v/v) and then retro-orbitally injected with 2.5 μg of anti-CD45-APC-Cy7 antibody (clone 30-F11, Biolegend) in 200 μl of PBS. Three minutes after antibody injection mice were euthanized by Isoflurane inhalation and cervical dislocation. Lungs, MLN, peripheral LN, BM and blood were collected from the mice depending on the experiment. Solid tissues were placed in complete media (RPMI1640, 5% FCS, 10 mM HEPES, 1x NEAA, 1 mM Sodium Pyruvate, 55 μM 2-mercaptoethanol) containing 1 mg/ml type 2 collagenase and 30 μg/ml DNase, minced with scissors and incubated for 30 minutes at 37° C. Tissues were then forced through a 70-uM cell strainer (BD) and washed twice in collagenase wash buffer (PBS + 2% FBS (v/v), 2 mM EDTA). Mouse femur and tibia leg bones were cleaned of tissue before removing their ends with scissors. Marrow was then extracted by perfusing with 5 mL of complete media using a 27G needle. Blood was collected in complete media. RBC were removed from all tissues using ACK lysis buffer (155 mM NH3Cl, 10 mM KHCO3, and 88 μM EDTA) and live cells were counted using trypan blue exclusion and a Countess cell counter (Thermo Fischer Scientific)
Single cell suspensions from each tissue and during each experiment were first washed with PBS and stained Zombie-Aqua (Biolegend) live/dead dye for 20 min at 4°C. Cells were then incubated with FcγRIII/I blocking antibody (clone 2.4G2)(Biolegend) and fluorescently labeled antibodies in PBS + 3% BSA (w/v) for 30 minutes at 4° C. From Biolegend, anti-CD11b (clone M1/70), CD11c (N418), MHCII I-A and I-E (M5/114.15.2), CD19 (6D5), Ly6C (AL-21), Ly6G (1A8), NK1.1 (PK136), Thy1.2 (30-H12), and CD103 (2E7) or XCR1 (ZET). From BD, anti-B220 (RA3-6B2), CD45 (30-F11) and Siglec F (E50-2440). Cells were then washed twice with PBS+3% FCS and fixed overnight in 1% paraformaldehyde (v/v).
For EdU staining, cell suspensions were stained with live/dead dye and incubated with blocking and fluorescently-labeled antibodies, except for antibodies with PE or PE-conjugated fluorophores. Cells were washed and fixed in 4% paraformaldehyde before permeablization in Click-iT saponin-based perm-wash buffer (Molecular Probes). EdU was then labeled with Alexa Fluor 647 azide in Click-iT buffer as per the manufacturers protocol. At each time point and for each experiment, cells from the tissues of uninfected or infected mice that did not receive EdU injections were also labeled with Alexa Fluor 647 using the Click-iT reaction and used as negative controls for EdU labeling. After further washing, cells were then incubated with antibodies labeled with PE and PE-conjugates. Cells were then washed twice with PBS+3% FCS and fixed overnight in 1% paraformaldehyde (v/v). Samples were acquired on a BD Biosciences LSR II and analyzed using FlowJo (TreeStar, Inc.).
Mice infected with GFP-expressing M. tuberculosis were injected with EdU and euthanized as described. Lungs were instilled with OCT by syringe via the trachea prior to embedding in OCT and freezing in liquid N2. Frozen tissues were then cut into 6uM with a cryostat, mounted on slides and fixed immediately with ice-cold 100% acetone for 15 minutes then stored at −20° C. Slides were rehydrated in PBS then blocked for PBS + 5% FBS for 15 min. EdU was labeled with Alexa Fluor 647 in Click-iT buffer as per the manufacturers protocol. Tissue sections were blocked with anti-FcγRIII/I (2.4G2) and labeled with anti-CD11b and CD4 fluorescently conjugated antibodies. Slides were mounted with Prolong-Diamond (Life sciences). Images were captured using the 60x oil-immersion objective on a Nikon Eclipse Ti fluorescent microscope. Images were captured over multiple frames and z-stacks and were stitched together and deconvoluted using NIS Elements software (Nikon). Images were further analyzed using ImageJ (National Institutes of Health).
Bone marrow and blood were collected from 30 naïve CD45.1 mice and processed into single cell suspensions. Cells were incubated with biotin-conjugated Monocyte Isolation Kit antibodies (Miltenyi Biotec) and anti-biotin microbeads as per the manufacturers instructions. Monocytes were then enriched negative selection on either an AutoMACS or MultiMACS magnetic column. Stem cells were further depleted by incubating enriched monocytes with biotin-conjugated Sca-1 (D7), CD117 (2B8) and CD135 (A2F10) antibodies and anti-biotin microbeads. Monocytes were then washed twice in PBS, counted and 1.0–2.25 x 106 monocytes in 100 μl of PBS were retro-orbitally intravenously injected into each M. tuberculosis-infected mouse.
Mice infected with aerosolized M. tuberculosis were intravenously injected with 250 μg of isotype (clone 2A3) or anti-CD62L antibody (Mel-14)(BioXCell) on day 28 post infection. Two hours after administering antibodies, mice were intraperitoneally injected with 1 mg of EdU. Mice were given an additional 100 μg of isotype or anti-CD62L antibody by intraperitoneal injection 24 h after the first antibody injection. Two days after the first antibody treatment, mice were injected with anti-CD45 antibody, sacrificed and organs were collected and processed as described previously.
|
10.1371/journal.pgen.1002038 | Mouse Genome-Wide Association and Systems Genetics Identify Asxl2 As a Regulator of Bone Mineral Density and Osteoclastogenesis | Significant advances have been made in the discovery of genes affecting bone mineral density (BMD); however, our understanding of its genetic basis remains incomplete. In the current study, genome-wide association (GWA) and co-expression network analysis were used in the recently described Hybrid Mouse Diversity Panel (HMDP) to identify and functionally characterize novel BMD genes. In the HMDP, a GWA of total body, spinal, and femoral BMD revealed four significant associations (−log10P>5.39) affecting at least one BMD trait on chromosomes (Chrs.) 7, 11, 12, and 17. The associations implicated a total of 163 genes with each association harboring between 14 and 112 genes. This list was reduced to 26 functional candidates by identifying those genes that were regulated by local eQTL in bone or harbored potentially functional non-synonymous (NS) SNPs. This analysis revealed that the most significant BMD SNP on Chr. 12 was a NS SNP in the additional sex combs like-2 (Asxl2) gene that was predicted to be functional. The involvement of Asxl2 in the regulation of bone mass was confirmed by the observation that Asxl2 knockout mice had reduced BMD. To begin to unravel the mechanism through which Asxl2 influenced BMD, a gene co-expression network was created using cortical bone gene expression microarray data from the HMDP strains. Asxl2 was identified as a member of a co-expression module enriched for genes involved in the differentiation of myeloid cells. In bone, osteoclasts are bone-resorbing cells of myeloid origin, suggesting that Asxl2 may play a role in osteoclast differentiation. In agreement, the knockdown of Asxl2 in bone marrow macrophages impaired their ability to form osteoclasts. This study identifies a new regulator of BMD and osteoclastogenesis and highlights the power of GWA and systems genetics in the mouse for dissecting complex genetic traits.
| Osteoporosis is a disease of weak and fracture-prone bones. The characteristic of bone that is most predictive of fractures is low bone mineral density (BMD), a trait primarily controlled by genetics. In recent years, significant advances have been made in the discovery of genes affecting BMD; however, our understanding of its genetic basis is still primitive. In this study, we used genome-wide association in the mouse to identify additional sex combs like-2 (Asxl2) as a novel BMD gene. In confirmation of our genetic analysis, mice deficient in Asxl2 had reduced BMD. To evaluate its function in bone, the expression levels of Asxl2 and tens of thousands of other genes were measured in bone in a large number of inbred mouse strains. Asxl2 demonstrated a pattern of expression indicative of genes that play a critical role in osteoclasts, the cells that are responsible for bone resorption. Further study of Asxl2 may reveal novel therapeutic targets for the treatment and prevention of osteoporosis.
| Osteoporosis is a common disease characterized by bone fragility and an increased risk of fracture [1]. One of strongest predictors of fracture is low bone mineral density (BMD) [2] and while BMD is influenced by both genetic and environmental factors, most (between 60% and 80%) of its variance is heritable [3]. Thus, the identification of novel BMD genes is critical for the discovery of new pathways and gene networks that will advance our understanding of basic bone biology and identify new therapeutic targets with the potential to combat osteoporosis.
Due in large part to its many advantages, such as the ability to experimentally cross genetically defined strains and perturb candidate genes, the mouse has played an instrumental role in the genetic analysis of BMD [4]. However, progress has been limited by low-resolution linkage-based quantitative trait loci (QTL) mapping approaches and the difficulties inherent to QTL cloning [5]. As a result, mouse linkage approaches have lead to the identification of only three BMD quantitative trait genes, Alox12 [6], Sfrp4 [7] and Darc [8], even though hundreds of QTL have been mapped [9]. In part due to the success of genome-wide association (GWA) in humans, several groups have investigated the prospects of high-resolution association mapping approaches in the mouse. One of the main challenges facing GWA in the mouse has been determining the most ideal population for such analyses. To date, mouse GWA studies have been performed with varying success using small sets of classical laboratory strains [10], advanced intercross lines [11], heterogeneous stocks [12], outbred mice [13], [14] and the Hybrid Mouse Diversity Panel (HMDP) [15]. One of the most promising is the HMDP, a collection of ∼100 classical laboratory and recombinant inbred (RI) strains that have been genotyped at ∼135,000 SNPs [15]. The primary advantage of the HMDP is that mapping resolution is over an order of magnitude higher than with linkage. We have recently demonstrated through simulations that associations explaining 5% of the phenotypic variance in a trait have 95% confidence intervals (CIs) of ∼2.6 megabases (Mb) [15]. This is in comparison to CIs for mouse linkage studies that are typically in the range of 40–60 Mb [13]. Additionally, statistical power in the HDMP has been found to be adequate to map variants affecting complex traits [15]. Moreover, phenotypes can be mapped in the HMDP without the need for costly genotyping and one can collect multiple specimens (e.g. tissues, individual cell-types, etc.) from strains for molecular profiling that are difficult or impossible to collect from a single mouse [15].
A drawback of both mouse and human GWA studies (and all “genotype to phenotype” mapping approaches) is their inability to provide information on how associated genes actually influence disease [16]. In many cases, it takes years to decipher the underlying biology of novel gene discoveries. One way to begin to provide functional information is through the use of systems genetics [17]. Systems genetics is an approach that incorporates molecular phenotypes, most commonly gene expression microarray profiles, into the genetic analysis of clinical phenotypes. One way that systems genetics can be used to functionally annotate genes of unknown function is through the generation of gene co-expression networks. Co-expression networks are created by clustering genes based on patterns of co-expression across a series of perturbations, such as the differing genetic backgrounds in the HMDP [18]. Co-expressed gene clusters or “modules” have been shown to be enriched for genes involved in the same general function [19], [20], [21], allowing one to annotate genes through “guilt-by-association” [22]. For example, if an uncharacterized gene is co-expressed with genes known to be involved in a biological process such as “apoptosis”, then it is more likely than not the unknown gene is also involved in “apoptosis”. The use of network analysis of systems genetic data can help to inform GWA discoveries by providing clues as to a gene's function in a physiologically-relevant context.
The goal of the current study was to identify and functionally characterize novel BMD genes using GWA and systems genetics in the HMDP. This approach implicated additional sex combs like-2 (Asxl2) as the gene responsible for a BMD association on chromosome (Chr.) 12. This was further strengthened by the observation that Asxl2 knockout mice had reduced BMD. Furthermore, gene co-expression analysis of bone transcriptomic data predicted that Asxl2 was involved in the differentiation of bone-resorbing osteoclasts. In support of this prediction, osteoclastogenesis was impaired in bone marrow macrophages in which Asxl2 expression was reduced by RNA interference. Together, these data are consistent with the hypothesis that Asxl2 is a novel regulator of BMD and osteoclastogenesis.
A detailed description of the HMDP, including strain selection and evaluation of statistical power and mapping resolution, is provided in [15]. Towards identifying genomic regions associated with BMD, we phenotyped 16-week old male mice (N = 879) from 97 HMDP strains (N = 9.1 mice/strain) for total body (TBMD), lumbar spine (SBMD) and femur (FBMD) areal BMD (Table S1). A wide range of BMD values were observed across the HMDP with differences of 1.4, 1.6 and 1.6-fold between the lowest and highest strains for TBMD, SBMD and FBMD, respectively (Figure 1).
To identify associations for the three BMD phenotypes we used the Efficient Mixed-Model Association (EMMA) algorithm [23]. Adjusted association P-values were calculated for 108,064 SNPs with minor allele frequencies >5%. We have previously demonstrated that the P<0.05 genome-wide equivalent for GWA using EMMA in the HMDP is P = 4.1×10−6 (−log10P = 5.39) [23]. At this threshold, associations on chromosomes (Chrs.) 7, 12 and 17 were identified influencing TBMD (Figure 2A). A fourth unique association on Chr. 11 was identified for SBMD (Figure 2B) and the only significant locus for FBMD was the Chr. 7 association also affecting TBMD (Figure 2C). The details of each association are provided in Table 1. Importantly, each of these regions overlap with the location of QTL for areal BMD measures previously identified by linkage in F2 crosses [9].
We have previously shown that the 95% confidence interval (CI) for the distribution of distances between the most significant and true causal SNPs, for simulated associations that explain 5% of the variance in the HMDP, is ∼2.6 Mb [24]. Therefore, we used this interval to conservatively define the boundaries of the four associations (Table 1). Within each association there were a total of 112 (Chr. 7), 14 (Chr. 11), 18 (Chr. 12) and 19 (Chr. 17) unique RefSeq genes (a full list of genes is provided in Table S2). We next identified those genes possessing functional alterations that might underlie the associations. Genes were selected based on whether they were regulated by a local expression QTL (eQTL) in the HMDP or if they harbored a non-synonymous (NS) SNP that was predicted to have functional consequences. For the eQTL analysis, we generated gene expression microarray profiles using RNA isolated from cortical bone in 95 of the 97 HMDP strains (N = 1–3/arrays per strain). EMMA was then used to perform an association analysis between all SNPs and array probes mapping within each region. A total of 74 genes were represented by at least one probe, after excluding probes that overlapped SNPs present among the classical inbred strains used in the HMDP (see Methods). Of these, 11 genes (8 within the Chr. 7 association and 3 within the Chr. 17 association) were identified with at least one probe whose expression was regulated by a significant (P≤5.1×10−4; Bonferroni corrected for the number of probes tested) local eQTL (Table 2 and data for all genes is provided in Table S2). In addition, we identified a total of 19 NS SNPs in 14 genes that were predicted to be either “Possibly Damaging” or “Probably Damaging” by PolyPhen [25], [26] (Table 3 and a list of all NS SNPs is provided in Table S3). A nonsense SNP was also identified in the meprin A, alpha (Mep1a) gene (Table 3). Therefore, of the 163 positional candidate genes, 26 were found to be regulated by a local eQTL in bone or harbored potentially functional NS SNPs. The number of functional candidate genes within each association was 12 (Chr. 7), 2 (Chr. 11), 3 (Chr. 12) and 9 (Chr. 17).
We also determined if any of the 163 genes implicated by GWA have been previously implicated in bone development. The associations on Chrs. 11 and 12 did not harbor known bone genes. In contrast, three (Fosb [27], RelB [28] and Apoe [29]) of the Chr. 7 genes have been linked to the regulation of bone mass. All three were located in close proximity to the association peak. Fosb was 124 kilobases (Kb) upstream, Relb was 180 Kb downstream and Apoe was 270 Kb downstream of rs32149600, the most significantly associated SNP on Chr. 7. Additionally, of the 19 Chr. 17 genes, Runx2, the “master regulator” of osteoblast differentiation [30], was located 1.0 Mbp downstream of rs33294019, the most significant SNP. None of the known bone genes were regulated by local bone eQTL or harbored potentially functional NS SNPs.
All 26 of the genes identified above are candidates for the BMD associations and warrant further investigation. However, the goal of this analysis was to identify a gene(s) that was the most likely causal gene for an association. Due to Chr. 7 and Chr. 17 possessing multiple functional candidates (12 and 9, respectively), we could not identify the most likely candidate based on the existing data for either of the associations; therefore, we focused on the Chr. 11 and Chr. 12 associations due to the presence of only 2 and 3 functional candidates, respectively. All five of these genes harbored potentially functional NS SNPs. The expression of these genes was not regulated by local eQTL. Of the five, the additional sex-combs like 2 (Asxl2) was the most compelling candidate due to the fact that rs29131970, a NS SNP in Asxl2 that was predicted to be functional, was also the peak SNP for Chr. 12 BMD association (Table 3 and Figure 3). Rs29131970 results in a phenylalanine (F) to serine (S) substitution at amino acid 1191 of the mouse ASXL2 protein. The mouse reference genome (the C57BL/6J strain) harbors the “C” allele, which codes for S; whereas, the rat, human, orangutan, dog, horse, opossum and chicken reference genomes all have the “T” allele, which codes for F. HMDP strains homozygous for the “C” allele had lower BMD relative to strains with the “T” allele (data not shown). A role for the other four candidates possessing NS SNPs on Chr. 11 and Chr. 12 appear to be less likely because of the modest linkage disequilibrium (LD) (r2 between 0.09 and 0.36) among the HMDP classical inbred strains between these SNPs and the SNPs most significantly associated with BMD for each region (Table 3). Therefore, based on the existing data we hypothesized that Asxl2 was the causal gene for the Chr. 12 association.
To directly test the hypothesis that Asxl2 was involved in the regulation of BMD we characterized TBMD, SBMD and FBMD in Asxl2 knockout mice (Asxl2−/−). The mice used for this experiment were between the ages of 2–4 months and as previously reported [31] a significant (P<0.05) reduction in body weight in Asxl2−/− mice was observed (data not shown). To evaluate bone mass in the absence of these confounding effects we evaluated the BMD residuals across genotype after adjusting for age and body weight within each sex. This analysis revealed significant (P<0.05) reductions in TBMD, SBMD and FBMD residuals in male Asxl2−/− mice as compared to wild-type (Asxl2+/+) controls (Figure 4A–4C). In addition, male heterozygous (Asxl2+/−) mice demonstrated an intermediate phenotype for all BMD measures, although the differences were not statistically different from either homozygous genotype (Figure 4A–4C). We also observed similar decreases in TBMD, SBMD and FBMD residuals in female Asxl2−/− mice as compared to Asxl2+/+ controls (Figure 4D–4F). These data confirm that Asxl2 is a regulator of BMD and are consistent with the hypothesis that Asxl2 is responsible for the genetic association on Chr. 12 in the HMDP.
We next used systems genetics to begin to identify a potential function for Asxl2 in bone. For this analysis we utilized the cortical bone gene expression data from the HMDP strains. An analysis of Asxl2 expression revealed that while its expression was not under regulation by detectable local (described above) or distant eQTL (data not shown), Asxl2 was expressed in cortical bone (in the top 10% of probes based on average expression) and most importantly, its expression differed by 1.5-fold between the lowest and highest expressing HMDP strains (Figure S1). We reasoned that its high level of expression and variation in expression among strains would allow us to identify biologically meaningful co-expression relationships between Asxl2 and genes sharing similar functions, even though a difference in its expression does not underlie the Chr. 12 association. To identify the genes that were co-expressed with Asxl2 in bone, genes were grouped into “co-expression modules” using Weighted Gene Co-expression Network Analysis (WGCNA) [20]. We used WGCNA to generate a bone co-expression network comprised of the 3600 most variable and highly connected genes (see Methods). The 3600 genes were subsequently partitioned into eight gene modules (Figure 5A). Of the eight, we focused our attention on the blue module that contained Asxl2 along with 1334 other genes (full list is provided in Table S4). The DAVID knowledge base (http://david.abcc.ncifcrf.gov/) was used to determine if the blue module was enriched for specific gene ontology (GO) categories. We were most interested in identifying enrichments in specific gene functions; thus, we restricted the analysis to the GO biological process and molecular function categories and excluded the cellular component category. DAVID's functional annotation clustering tool was used to identify 14 significant (enrichment score (ES) >3.0; see Methods) gene clusters containing highly related GO terms (Table S5). The top four clusters contained genes involved in: 1) the cell cycle/chromosome/DNA replication/cell division, 2) hematopoiesis/myeloid cell differentiation, 3) ATP binding and 4) chromosome organization/chromatin organization (Table 4). Asxl2 is thought to regulate the function of Polycomb (PcG) and Trithorax (trxG) protein complexes, which are involved in the establishment and maintenance of chromatin [31]. Its membership in a module enriched for genes involved in the GO terms “chromosome” (Bonferroni corrected enrichment P = 3.2×10−7) and “chromatin organization” (Bonferroni corrected enrichment P = 9.2×10−3) is consistent with its known function and suggests that this module is comprised of biologically meaningful co-expression relationships.
We next characterized the nature of the genes most closely connected to Asxl2 in the blue module. For this purpose, a network view was generated for the 34,690 strongest connections among 1256 (94%) of the 1335 blue module genes (Figure 5B). In this view of the blue module, Asxl2 was connected to a single node, the apoptotic peptidase-activating factor 1 (Apaf1) gene. Apaf1, in turn, was connected to 149 genes; all located within the cluster of genes on the left side of the network depiction Figure 5B. To examine the gene composition of the cluster connected to Asxl2, genes were colored based on their membership in three of the four top enrichments described above, including the “cell cycle”, “myeloid cell differentiation” and “chromatin organization” (Table 4). We excluded the “ATP-binding category” since genes in this category are involved in a wide-range of biological processes. The blue module as a whole was enriched for all three categories (Table 4); however, the cluster of genes most closely connected to Asxl2 contained most (75%; enrichment P = 2.0×10−3) of the genes involved in myeloid cell differentiation (red nodes in Figure 5B) and very few cell cycle (green nodes) or chromatin organization (yellow nodes) genes. These data indicate that in our bone network, Asxl2 is most closely connected with genes involved in myeloid cell differentiation and may play a role in this process.
The GO category “myeloid cell differentiation” is comprised of genes that are involved in the general process of myeloid precursors acquiring characteristics of downstream cell lineages. With respect to bone cells, osteoclasts are bone-resorbing cells of myeloid origin; thus, we hypothesized that Asxl2 played a role in the differentiation of pre-osteoclasts. Additionally, many of the blue module genes in this category have been implicated in osteoclastogenesis, such as Inpp5d (aka SHIP) [32], Smad5 [33], Id2 [34], Plcg2 [35], Twsg1 [36], among others. To test the hypothesis that Asxl2 was involved in osteoclastogenesis, we infected bone marrow macrophages (BMMs) with lentiviral constructs expressing short-hairpin RNAs (shRNA) targeting Asxl2. BMMs were infected with a vector only control (NC) or one of five lentiviral constructs (A1–A5) expressing distinct shRNAs targeting Asxl2. After five days of culture, Asxl2 expression was particularly lower in cells infected with the A3 and A5 lentiviral constructs (Figure 6A). Osteoclastogenesis was induced in infected BMMs by culturing in the presence of M-CSF and RANKL. After five days of culture the number of TRAP+ (tartrate-resistant acid phosphatase, a marker of osteoclasts) multinuclear cells (MNCs) was significantly (P<0.05) reduced in cultures infected with lentiviral constructs A3 and A5 compared to NC treated cells (Figure 6B and 6C). We observed a strong positive correlation (r = 0.74, P = 0.04) between the relative expression of Asxl2 and TRAP+ MNCs across the six treatments (Figure 6D). These data are consistent with our network inference and confirm that Asxl2 is a regulator of osteoclastogenesis and strengthen the hypothesis that Asxl2 is a regulator of BMD.
The mouse has numerous advantages for the genetic analysis of BMD; however, historically mapping approaches in the mouse have been plagued by the lack of resolution. Additionally, GWA approaches provide no information on the function of associated genes. We have addressed both limitations through the use of high-resolution GWA in the HMDP to identify associations confined to narrow genomic intervals and gene co-expression analysis of bone microarray data to provide insight on gene function. This novel analytical paradigm resulted in the discovery of Asxl2 as a regulator of BMD and osteoclastogenesis. This study identified a new gene and possibly an entire network of genes that play an important role in BMD and osteoclast function.
Polycomb (PcG) and trithorax (trxG) are highly conserved protein complexes that are involved in the repression and activation of gene expression, respectively, through the establishment and maintenance of chromatin modifications at specific target genes [37]. With respect to bone cells, PcG and trxG have been implicated in cell proliferation [38], myeloid precursor maturation [39] and osteoblast differentiation [40]. In Drosophila, Additional sex combs (Asx) belongs to a group of proteins known as the Enhancers of trxG and PcG (ETP) [41]. Although its specific mechanism is unknown, Asx is thought to promote both PcG-mediated silencing and trxG-mediated activation of gene expression. In humans and mice there are three Asx homologues, Asx-like 1, 2 and 3 [42], [43], [44]. Mutations in ASXL1 result in myleoproliferative neoplasms [45] and little is known regarding the function of ASXL3. Recently described Asxl2−/− knockout mice display a global reduction in the PcG-associated histone modification trimethylation of histone H3 lysine 27. This is consistent with a conserved function as an ETP in mammals [31]. Additionally, Asxl2−/− mice develop anterior and posterior transformations of the axial skeleton [31], which further supports our observation that Asxl2 is involved in bone development.
An important open question is whether Asxl2 impacts bone development through its expression in other cell-types, such as osteoblasts. If Asxl2 functioned exclusively in osteoclasts, we would expect based on the in vitro osteoclastogenesis data, that loss of Asxl2 function would impair osteoclast function and bone resorption in vivo, resulting in increased BMD. In contrast, we observed a decrease in BMD in Asxl2−/− mice. It has been shown in a number of mouse models that loss of specific genes can lead to an impairment of both osteoblast and osteoclast function. This results in a condition referred to as low-turnover osteopenia in which bone formation by osteoblasts and resorption by osteoclasts are impaired with a net loss of bone. As examples, osteoblasts and osteoclasts from mice deficient in klotho [46], JunB [47] and Akt1 [48] have impaired in vitro differentiation. These mice also have reduced BMD due to low-turnover osteopenia. In addition, Synaptotagmin VII (Syt11) has been shown to alter protein secretion in osteoblasts and osteoclasts, resulting in decreased bone mass in Syt−/− mice [49]. In data not presented we observed (using publically available data from the BioGPS browser; http://biogps.gnf.org, probes 1460597_at and 1439063_at) high and ubiquitous expression of Asxl2 in 96 different mouse tissues and cell-lines, including primary osteoclasts and osteoblasts. In addition, while the blue module was highly enriched for genes involved in myeloid differentiation there were also a number of genes such as Bmp4, Chrd, Hdac5 and Igf2 that play a role in osteoblast differentiation (Table S3). These data suggest that the decreases in BMD seen in Asxl2−/− mice may be due to deficiencies in both osteoclast and osteoblast function. Further work is needed to clarify the precise role of Asxl2 in bone.
We have previously characterized the HMDP as a novel population for GWA in the mouse [15]. This study extends our original observations by demonstrating the feasibility of identifying associations affecting additional complex phenotypes. In contrast to more traditional mouse linkage mapping strategies, we used association in the HMDP to identify four associations containing a relatively small number of genes. Based on prior work, the boundaries of the associations were defined as the 2.6 Mb region surrounding the most significant SNP. We expect that these intervals are conservative and would likely be smaller if based on region specific LD patterns, as shown for Chr. 12 (Figure 3). However, defining the associations in this way allowed us to be confident that the regions contained the causal gene(s). This study also highlights the other key advantage of the HMDP; the ability to collect and molecularly profile tissues, such as bone, that are difficult or impossible to collect from a single mouse or in human populations. In the future, the accumulation of many bone-related clinical and molecular phenotypes in the HMDP will enable the large-scale systems-level analyses that should provide significant insight on bone physiology and genetics. Additionally, the HMDP will provide the opportunity to address the genetic basis of extremely important phenotypes, such as bone loss, nanostrucutal properties of bone and bone cell aging that are difficult to address in humans. Moreover, as we have used systems genetic to gain functional insight for a gene identified by mouse GWA, it is also possible that the HMDP could be used to dissect the function of genes identified in human GWA studies.
However, the HMDP is not without limitations. First, the statistical power to detect effects of subtle variants is modest. We have previously estimated that for highly heritable phenotypes, such as BMD, the power to detect variants explaining 10% of the trait variance is 50% [15]. The power drops precipitously for variants explaining less than 10% of the variance. Thus, this version of the HMDP (with ∼100 strains) is unable to identify the many more variants with subtle effects that are undoubtedly affecting BMD in this population. This problem will be less of an issue for future HMDP panels containing a larger number of strains. Additionally, one of the side effects of the breeding history of inbred mouse strains is the presence of LD between markers on separate chromosomes (i.e. non-syntenic LD). It is thought that this is due to selection for allelic combinations that confer increased fitness during the inbreeding process [50]. False-positive associations can arise if a region associated with a phenotype is in LD with other regions of the genome. Although this is always a potential pitfall when using the HMDP, it is easily identifiable and we did not observe LD (r2<0.4) between any of the four BMD associations (data not shown).
Most GWA studies stop at gene discovery. However, without physiologically relevant functional information it is often difficult to be confident that the true causal gene has been identified and to begin unraveling the mechanistic underpinnings of significant associations. Recently, some GWA studies have begun to incorporate gene expression information to determine if a significant variant regulates expression. A positive finding in such an analysis suggests that the genotype dependent differential gene expression is the basis of the association. This approach has recently been used to discover that variants in the promoter of the serine racemase (SRR) gene regulate its expression and BMD [51]. We have taken that application one step further and developed a bone transcriptional network to aid in the functional annotation of genes of unknown function. Using this network, we were able to determine that Asxl2 was closely connected to genes with links to osteoclast differentiation. This simple connection allowed us to test the hypothesis that Asxl2 was involved in a bone specific function. Another important point is that, as we demonstrated for Asxl2, a gene's expression does not have to be under genetic regulation for this approach to work.
In conclusion, we have used mouse GWA and gene co-expression network analysis to identify Asxl2 as a novel regulator of BMD and osteoclastogenesis. Our analysis has revealed a new gene and pathway that play an important role in bone development. Additionally, this study demonstrates the feasibility of using the HMDP for the dissection of complex genetic traits.
The animal protocol for the HMDP mice was approved by the Institutional Care and Use Committee (IACUC) at the University of California, Los Angeles. The animal protocol for the Asxl2−/− mice was approved by the Animal Care Committee (ACC) at the University of Illinois at Chicago.
Approximately nine male mice for each HMDP strain (Table S1) were purchased from the Jackson Labs (Bar Harbor, ME). Mice were between 6 and 10 weeks of age and to ensure adequate acclimatization to a common environment the mice were aged until 16 weeks before sacrifice. All mice were maintained on a chow diet (Ralston-Purina Co, St. Louis, Mo) until sacrifice.
Inbred strains were previously genotyped by the Broad Institute (available via www.mousehapmap.org). Genotypes of recombinant inbred strains were imputed as previously described [15]. Of the 140,000 SNPs available at the Broad Institute, 108,064 were informative with an allele frequency ≥5% and less than 20% missing data, and were used for the association analysis.
All carcasses were stored at −20°C after sacrifice and then thawed overnight at 4°C prior to BMD scans. The entire thawed carcass was scanned. BMD scans were performed using a Lunar PIXImus II Densitometer (GE Healthcare, Piscataway, NJ). The PIXImus II was calibrated daily using a phantom of known BMD. BMD was calculated for the entire carcass minus the skull, the lumbar spine and the left femur.
At sacrifice the diaphysis of the right femur was excised and cleaned free of soft tissue. Bone marrow was removed by flushing with PBS using a 22-guage needle and 3 ml syringe. The bone was then flash frozen in LN2 and stored at −80C. Total RNA was isolated using the Trizol Plus RNA Purification Kit (Invitrogen, Carlsbad, CA) following homogenization of the whole bone sample. RNA integrity was confirmed using the Agilent 2100 Bioanalyzer (Agilent, Palo Alto, CA). Microarray expression profiles were generated (N = 1–3 per strain) using the Illumina MouseWG-6 v1.1 BeadChips (Illumina, San Diego, CA) by the Southern Genotyping Consortium at UCLA. Biotin-labeled cRNA was synthesized by the total prep RNA amplification kit from Ambion (Austin, TX). cRNA was quantified and normalized to 77 ng/µl, and then 850 ng was hybridized to Beadchips.
The expression values were transformed using the Variance Stabilizing Transformation (VST) [52], and normalized with the Robust Spline Normalization (RSN) algorithm using the LumiR R package [53]. After normalization, the ComBat software was used to adjust for batch effects using an empirical Bayes methods [54]. Microarray data has been submitted the the NCBI Gene Expression Omnibus (GEO) database (GSE27483).
Population structure is a major confounding for genome-wide association analyses in the HMDP. This is due to the fact that many phenotypes correlate with the phylogeny of HMDP strains (i.e. genetically similar strains have similar phenotypes) and any SNP that correlates with these strain relationships will be falsely associated with the phenotype. The Efficient Mixed-Model Association (EMMA) algorithm has been shown to effectively reduce this confounding [23]. We applied the following linear mixed model to perform association mapping between BMD or a gene's expression and a marker under the confounding effect from population structure [23], [55], [56]; , where is a phenotypes of each mouse, is their genotype, is an incidence matrix mapping each mouse to corresponding strain, is a random effect accounting for population structure effect with , and is the uncorrelated random effect with . The Efficient Mixed Model Association (EMMA) [23] is used for efficient and reliable estimation of restricted maximum likelihood (REML) parameters and hypothesis testing under the linear mixed model. After estimating REML parameters, a standard F test is used to test the statistical significance of the marker-phenotype association.
To characterize the genes located in each BMD association, we downloaded all RefSeq genes in the four regions from the USCS genome browser (http://genome.ucsc.edu/cgi-bin/hgGateway) using the NCBI Build37 genome assembly. From the Illumina MouseWG-6 microarray we identified all probes corresponding to the 163 RefSeq genes. Probes were excluded if they overlapped with SNPs (dbSNP 128) to avoid the hybridization artifacts that can arise due [57], [58]. EMMA was used to calculate association P-values for all probes corresponding to the 163 RefSeq genes. Only SNPs mapping to each associated region were used in this analysis. Known non-synonymous SNPs within each region were downloaded from the Mouse Phenome Database (http://phenome.jax.org/) using a set of over 7 million genotyped and imputed SNPs. We only selected SNPs that were variant in at least one of the classical inbred strains represented in the HMDP. Prediction of the functional effect of these SNPs was performed using the PolyPhen tool (http://genetics.bwh.harvard.edu/pph/). R2 was calculated for each non-synonymous SNP and the peak BMD SNP within the 30 classical inbred strains in the HMDP.
The generation and initial characterization of Asxl2−/− gene trap mice has been previously described [31]. These mice harbor a gene trap cassette downstream of exon 1. Homozygotes for the gene trap allele show little (∼3%) Asxl2 expression whereas heterozygotes expressed Asxl2 at ∼50% of wild type levels. BMD in littermate male and female mice (2–4 months of age) of varying Asxl2 genotype was measured as described above. Age and weight at sacrifice were recorded. To assess the effects of Asxl2 deficiency on BMD independent of age and weight we generated BMD residuals using a simple linear regression. A Student's t-test was used to test the significance of the differences in BMD residuals in the different genotypes.
Network analysis was performed using the WGCNA R package [59]. An extensive overview of WGCNA, including numerous tutorials, can be found at http://www.genetics.ucla.edu/labs/horvath/CoexpressionNetwork/. To begin, we filtered the array data to remove lowly and non-expressed genes by selecting probes based on a detection P-value of <0.05 in 95% of the samples. Next, we selected the 8000 most varying genes based on variance across the 95 samples and then selected the most connected (based on k.total described below) 3600 genes for network analysis. Our group and others have used this number of genes previously (as examples [19], [60]), mainly because most other genes have very low k.total values. If multiple probes existed for a given gene only the most connected probe per gene was included in the list of 3600. To generate a co-expression network for the selected probes, we first calculated Pearson correlation coefficients for all gene-gene comparisons across the 95 microarray samples. The matrix of correlations was then converted to an adjacency matrix of connection strengths. The adjacencies were defined as where and are the and gene expression traits. The power was selected using the scale-free topology criterion previously outlined by Zhang and Horvath [18]. Network connectivity (k.total) of the gene was calculated as the sum of the connection strengths with all other network genes, . This summation performed over all genes in a particular module was defined as the intramodular connectivity (k.in). Modules were defined as sets of genes with high topological overlap [18]. The topological overlap measure (TOM) between the and gene expression traits was taken as , where denotes the number of nodes to which both and are connected, and indexes the nodes of the network. A TOM-based dissimilarity measure was used for hierarchical clustering. Gene modules corresponded to the branches of the resulting dendrogram and were precisely defined using the “Dynamic Hybrid” branch cutting algorithm [61]. Highly similar modules were identified by clustering and merged together. In order to distinguish modules each was assigned a unique color.
Whole bone marrow was extracted from femora of mice with α-MEM and cultured overnight in α-MEM (Sigma-Aldrich, St Louis, MO) containing 10% heat-inactivated FBS, 100 IU/ml penicillin, and 100 µg/ml streptomycin (α10 medium). The nonadherent cells were collected by centrifugation and re-plated in a new 10-cm petri dish in α10 medium. To generate osteoclasts, 100 ng/ml RANKL and 1/100 vol CMG 14-12 culture supernatant were added to α10 medium for 4–5 days. Osteoclasts were stained for TRAP as described by the manufacturer's instructions (Sigma-Aldrich). The SHCLNG-NM_172421 MISSION lentiviral shRNA (Sigma-Aldrich) clone set (containing five separate lentiviral shRNA clones each expressing a distinct Asxl2 targeting sequence) was used to transduce BMMs according to manufacturer's specifications. The MISSION pLKO.1-puro control vector was used as a negative control.
Total RNA was isolated from cultures using the RNeasy Mini Kit (Qiagen). Purified RNA was DNase treated using the DNA-free kit (Ambion). The High Capacity cDNA Reverse Transcription Kit (Applied Biosystems) was used to synthesize cDNA in a volume of 20 µl. The reaction mixture was adjusted to 200 µl with dH2O and 5 µl of the dilute cDNA was used for PCR. PCR was performed for 22 cycles for Asxl2 and Actb using the following primers (Asxl2-F, 5′-ACCCACCATTCCAGCAAGTA-3′, Asxl2-R, 5′-TGGCTGCTTTGACAGTCTTG-3′, Actb-F, 5′-CCAACCGTGAAAAGATGACC-3′, Actb-R, 5′-ACCAGAGGCATACAGGGACA-3′). PCR products were separated on a 1.5% agarose gel containing 0.5 mg/ml ethidium bromide. Band densitometry was performed using the Image J software (NIH). Normalized Asxl2 expression levels were determined by subtracting lane specific backgrounds for each sample and dividing by Actb intensities.
|
10.1371/journal.ppat.1003635 | Human Cytomegalovirus Latency-Associated Proteins Elicit Immune-Suppressive IL-10 Producing CD4+ T Cells | Human cytomegalovirus (HCMV) is a widely prevalent human herpesvirus, which, after primary infection, persists in the host for life. In healthy individuals, the virus is well controlled by the HCMV-specific T cell response. A key feature of this persistence, in the face of a normally robust host immune response, is the establishment of viral latency. In contrast to lytic infection, which is characterised by extensive viral gene expression and virus production, long-term latency in cells of the myeloid lineage is characterised by highly restricted expression of viral genes, including UL138 and LUNA. Here we report that both UL138 and LUNA-specific T cells were detectable directly ex vivo in healthy HCMV seropositive subjects and that this response is principally CD4+ T cell mediated. These UL138-specific CD4+ T cells are able to mediate MHC class II restricted cytotoxicity and, importantly, show IFNγ effector function in the context of both lytic and latent infection. Furthermore, in contrast to CD4+ T cells specific to antigens expressed solely during lytic infection, both the UL138 and LUNA-specific CD4+ T cell responses included CD4+ T cells that secreted the immunosuppressive cytokine cIL-10. We also show that cIL-10 expressing CD4+ T-cells are directed against latently expressed US28 and UL111A. Taken together, our data show that latency-associated gene products of HCMV generate CD4+ T cell responses in vivo, which are able to elicit effector function in response to both lytic and latently infected cells. Importantly and in contrast to CD4+ T cell populations, which recognise antigens solely expressed during lytic infection, include a subset of cells that secrete the immunosuppressive cytokine cIL-10. This suggests that HCMV skews the T cell responses to latency-associated antigens to one that is overall suppressive in order to sustain latent carriage in vivo.
| Human cytomegalovirus (HCMV) is a widely prevalent virus, which is normally carried without clinical symptoms, but often causes severe clinical disease in individuals with compromised immune responses. In healthy HCMV carriers, the immune response to HCMV is robust and includes large numbers of virus-specific T-cells that control viral replication during active infection. Despite this prodigious immune response, HCMV is never cleared after primary infection but persists in the host for life: a key feature of persistence is the ability of the virus to establish a type of viral quiescence, termed latency. Although much is known about T-cell responses to viral proteins expressed solely during lytic infection, the interplay between the T-cell response and latent HCMV is not well understood. Here we report the first comprehensive characterisation of the T-cell response to latent HCMV and show that it is composed principally of CD4+ T-cells, which are specific for viral proteins expressed during latency, and are able to detect latent virus in vitro. We further show that these CD4+ T-cell responses to latency-associated viral gene products include T-cells that secrete cIL-10, an immunosuppressive cytokine, which may function to suppress antiviral immune responses and thereby maintain lifelong carriage of the latent virus.
| Human cytomegalovirus (HCMV) is widely prevalent, with an estimated 50–60% of the world population being seropositive [1]. Primary infection in the immunocompetent host is usually asymptomatic and overt disease is seen almost exclusively in the immunocompromised and immuno-naive host. For example, placental transmission of HCMV is a leading infective cause of congenital abnormalities [2]. Primary infection with HCMV induces a robust innate and adaptive immune response, which includes a substantial CD4+ and CD8+ T cell response [2]–[4], which is essential for the control of HCMV disease [5]–[8]. However, despite this extensive immune response, HCMV is not cleared but persists for the lifetime of the host due, at least in part, to the establishment of viral latency in certain cell types, where the viral genome is carried in the absence of the production of infectious viral progeny [9], [10].
One defined site of latency of HCMV in vivo is in cells of the myeloid lineage, including CD34+ haematopoietic progenitor cells of the bone marrow [11]–[14]. Furthermore, differentiation of CD34+ cells to terminally differentiated cells of the myeloid lineage, such as macrophages or dendritic cells, results in reactivation of infectious virus [12], [15]–[17]. Despite probable frequent occurrences of reactivation events in vivo, these events are likely asymptomatic due to an active and robust immune response. In contrast, in an immunocompromised setting, the lack of a functional T cell response results in uncontrolled virus replication, which can occur during primary infection, super-infection or reactivation [2], [18], [19] and result in significant clinical disease [2], [18].
The viral proteins recognised by HCMV-specific T cells during lytic infection have been extensively investigated. The immediate early proteins IE1 and IE2 (UL123 and UL122, respectively), as well as the tegument protein pp65 (UL83), were all recognized by both CD4+ and CD8+ T cells in the majority of individuals, regardless of HLA type. In contrast, although T cells specific for glycoprotein B (UL55) are also frequently generated they are predominantly CD4+ T cells [20]–[23]. Using synthetic peptide libraries spanning the entire predicted HCMV proteome, a comprehensive analysis of the breadth and frequency of the CD4+ and CD8+ T cell response to HCMV has been carried out. These data showed that, in a large cohort of healthy seropositive donors with a diverse range of HLA types, a CD4+ or CD8+ T cell response was detectable to over 150 viral ORFs which included both structural and non-structural proteins expressed during all phases of lytic infection [20], [21]. Furthermore, they also observed that in any given donor CD4+ and CD8+ T cell responses recognised a median of 8 ORFs and 12 ORFs, respectively, suggesting that a large repertoire of HCMV ORFs were recognised by the immune response.
In contrast to lytic infection, during which pivotal genes such as the viral immediate early (IE) genes drive expression of early and late genes such as pp65 and gB, viral gene expression during latency is highly restricted. In the absence of IE gene expression, an accepted characteristic of latent infection, only a handful of viral genes have been shown to be expressed during natural and experimental models of latent infection [24]–[31]. These include: UL138, LUNA (latency-associated unidentified nuclear antigen) an antisense transcript to the UL81–82 region, UL111A (vIL-10) a viral homologue of cellular IL-10 (cIL-10) and US28, a chemokine receptor homologue [28], [32].
Although the functional role of these latency-associated genes during latency is far from clear, we hypothesised that the robust and antigenically broad T cell immune response elicited by HCMV would likely include responses to these latency-associated antigens, since they are expressed during lytic infection also. Consistent with this, recent evidence has suggested a limited CD8+ T cell response to UL138 is present in healthy individuals [33]. However, if this was generally the case it would raise the question as to why such responses do not expose the latently infected cell to immune recognition and eventual clearance. Clearly, understanding this is of crucial importance to understanding the mechanism of latent carriage.
In this study we have measured and characterised UL138 and LUNA specific T cell responses in a cohort of healthy HCMV seropositive donors. Here we show, that they elicit predominantly CD4+ Th1 type responses, characterised by IFNγ production. Additionally, we observe that UL138 specific CD4+ T cells mediated MHC class II restricted cytotoxicity. These CD4+ T cells are able to recognize antigen both in lytically infected dendritic cells and, importantly, they also recognize latently infected monocytes. Intriguingly, a proportion of the UL138 and LUNA specific CD4+ T cells also secreted the immunomodulatory cytokines cIL10 and transforming growth factor β (TGF-β). Importantly, additional analysis of two other latency-associated proteins (US28 and vIL-10), showed that these antigens also generated CD4+ T cell responses which again were able to secrete cIL-10. To our knowledge this is the first description of HCMV specific cIL-10 producing CD4+ T cells in normal healthy HCMV seropositive individuals. We hypothesise that this T cell response, which includes potentially suppressive T cells specific to latent antigens, may function to assist in the lifelong carriage and maintenance of the latent reservoir of infection by preventing efficient Th1 T cell recognition of latently infected cells.
In order to determine whether T cell responses to UL138 or LUNA proteins could be detected, IFNγ specific ELISPOT assays were performed on whole freshly isolated peripheral blood mononuclear cells (PBMC) or PBMC depleted of either CD4+ or CD8+ T cells from both HCMV seropositive (n = 17) and seronegative donors (n = 6). Isolated cells were stimulated with overlapping pools of peptides spanning either the UL138 or LUNA predicted open reading frames and T cell recognition was determined by IFNγ production. In parallel, we performed a concomitant analysis of the response to pools of peptides spanning the HCMV ORFs IE1/2 (UL123 and UL122), pp65 (UL83) and gB (UL55), which are well defined CD4+ and CD8+ T cell antigens in HCMV seropositive individuals (Table 1).
As expected, substantial spot forming units (SFU) (SFU/106 >100) CD4+ and CD8+ T cell responses to pp65 (CD8+ T cells 15/17 donors and CD4+ T cells 12/17 donors), IE (CD8+ T cells 17/17 and CD4+ T cells 11/17 donors) and gB (CD8+ T cells 3/17 and CD4+ T cells 12/17 donors) were readily detected in the HCMV seropositive, but not seronegative donors.
Interestingly, T cell responses to UL138 and LUNA were also detected in many of the donors tested with responses greater than 100 SFU/106 cells to UL138 in 5/17 donors (range 120–1295 SFU/106 cells) and to LUNA in 8/17 donors (range 100–446 SFU/106 cells) seropositive donors tested. Furthermore, responses >100 SFU/106 cells were mediated exclusively by CD4+ T cells in all the UL138 and LUNA reactive donors (Table 1); none of the seronegative donors tested made responses (data not shown).
We confirmed that all donor PBMC and CD4/CD8 –depleted PBMCs tested retained IFNγ producing effector function following polyclonal stimulation using phytohaemagglutinin (PHA) (data not shown). Taken together, these results clearly show that both UL138 and LUNA generate a CD4+ T cell response and, furthermore, the frequency of the response was in the lower to middle range compared to CD4+ T cell responses to gB (Table 1). In order to expand any UL138 specific T cells that were below the detection limit of the ex-vivo ELISPOT assays purified CD8+ T cells from three different donors were stimulated with the UL138 peptide pool and cultured for 12 days. These cultures were then retested for UL138 specific responses by ELISPOT assays and were negative in all cases (data not shown).
In order to confirm the frequencies of these CD4+ T cell responses as well as determine more precisely the epitope specificity, individual 15 amino acid overlapping peptides that composed the UL138 and LUNA ORF pools were tested for T cell reactivity against each donor. All 17 HCMV seropositive donors were re-tested using the individual peptides from each ORF, irrespective of whether they had a detectable response to UL138 or LUNA ORF pools in the initial screen.
The results show that no additional UL138 or LUNA IFNγ secreting T cell responses were detected by stimulation with the individual peptides. All five donors previously shown to elicit a CD4+ T cell responses to the UL138 ORF pool were restricted to an immunogenic 15mer peptide or clusters of overlapping peptides (Figure 1) including 3 donors (CMV300, 301, and 305) who elicited high level IFNγ production to the same 15mer peptide; UL138 peptide 2 (LNVGLPIIGVMLVLI). Similarly, six of the seven donors with a detectable LUNA specific T cell response were mapped to either a 15mer peptide or alternatively to a clusters of overlapping 15mer peptides (Figure 2). Again 3 donors (CMV300, 302 and 317) showed IFNγ+ T cell responses to the same 15mer peptide, LUNA peptide 18 (RLILSGLPGVRVQNP).
UL138 and LUNA specific T cells are clearly generated in response to HCMV infection and can be re-stimulated with synthetic peptides. Consequently, we next addressed whether UL138 and LUNA specific CD4+ T cells could be activated in response to cells lytically infected with HCMV. CD4+ T cell lines specific for UL138, LUNA as well as gB and IE were incubated with either HCMV infected autologous monocyte derived dendritic cells (moDC) or mock infected cells (Phenotypic analysis of moDC shown in Figure S1). Following infection with HCMV, approximately 10% of moDC cells were IE positive (Figure 3A). Lack of availability of antibodies against UL138 and LUNA meant an RT-PCR analysis was performed instead. Using this approach we confirmed the expression of IE, UL138 and LUNA mRNAs expression in lytically infected moDCs (Figure 3B). As expected, all samples were positive for GAPDH and mock-infected moDCs expressed no viral transcripts.
The results clearly show that both the UL138 and LUNA specific T cells from donor CMV300 were activated following stimulation with lytically infected moDC (Figure 3C). As expected, gB and IE specific CD4+ T cells were also stimulated by virus infected cells (Figure 3C). Consistent with specificity, all CD4+ T cell lines were activated by the relevant gB, IE, UL138 and LUNA peptide pulsed moDC. In contrast, no IFNγ was detected following stimulation with uninfected/peptide untreated moDC. Similarly, CD4+ T cells specific for UL138 from a second donor (CMV305) were tested for recognition of HCMV infected autologous moDC. Again, we observed activation of all three specificities in this donors' CD4+ T cells (Figure S2). Furthermore, repetition of the analyses on independent occasions confirmed that, from both donors, latent specific T cells were activated by lytically infected moDC (data not shown).
The previous experiments clearly demonstrated that HCMV lytic infection of MHC Class II positive dendritic cells lead to antigen presentation of both UL138 and LUNA peptides and subsequent CD4+ T cell recognition. An important question was whether these T cells could also be activated by latently infected autologous monocytes. In order to determine if UL138 specific CD4+ T cells could recognize latently infected cells, autologous monocytes were prepared and latently infected with HCMV strain TB40e for 10 days. The establishment of latency was confirmed by RT-PCR analysis; latently infected monocytes were shown to express the UL138 latent transcript with an absence of IE mRNA expression (Figure 4A). Mock-infected monocytes did not express UL138, while both mock and latently infected cells were GAPDH positive.
UL138 and gB specific CD4+ T cells from donor CMV305 were generated as previously described and co-incubated with autologous latently or mock infected monocytes in IFNγ ELISPOT assays. As expected, no gB specific CD4+ T cell response against HCMV latently infected monocytes was detected. In contrast, gB specific CD4+ T cells clearly produced IFNγ if the mock or latently infected monocytes had been pulsed with gB peptide (Figure 4B). These results are consistent with a lack of gB expression during latent infection of monocytes but confirm that failure to detect gB is not due to a defect in antigen presentation to gB-specific CD4+ T cells by autologous monocytes.
Interestingly, in contrast to our observations with gB, our results show that UL138 specific CD4+ T cells produce IFNγ in response to incubation with latently infected monocytes (Figure 4C) as well as when cells were pulsed with UL138 peptide. Identical analyses were also performed using donor CMV300, which confirmed the recognition of latently infected monocytes by CD4+ T cells specific to UL138 (Figure S3). Further validation was achieved with subsequent analyses of T cells and monocytes derived from donors CMV300 and CMV305 which, again, confirmed that UL138 specific T cells recognized HCMV latently infected monocytes (data not shown). To date, we have not been able to demonstrate if LUNA protein expression in latently infected cells can be recognised by LUNA specific T cells. Whether this is experimental failure or intrinsic to LUNA remains an open question and is currently under investigation.
Classically, CD4+ T cells are considered helper T cells, exerting effector functions by cytokine secretion in vivo. Indeed, it is well established that T cell mediated cytotoxicity effector function is associated with CD8+ T cell recognition of antigen presented by MHC Class I. However, recent studies suggest that CD4+ T cells may play a more direct role in viral infection. Pertinent to this study, it has been shown in the context of HCMV infection that Th1 type gB specific CD4+ T cells that secrete IFNγ and TNFα have also been shown to be able to mediate MHC class II restricted cytotoxicity [21], [34]–[36].
Consequently, since both the UL138 and LUNA specific CD4+ T cell response is also characterised by IFNγ production, we next asked if these CD4+ T cells also have cytotoxic effector cell function. To do this, UL138 specific CD4+ T cell lines from donors CMV300 and CMV305 as well LUNA specific CD4+ T cell lines from CMV300 were expanded in vitro for two weeks and then tested for MHC Class II restricted cytotoxicity in vitro using chromium release assays. Again gB specific CD4+ T cells were used as a positive control.
Consistent with previously published findings [34], [35] gB specific CD4+ T cells mediated cytotoxicity (Figure 5A). Furthermore, our results now show that UL138 specific CD4+ T cells are also able to mediate cytotoxicity (Figure 5B). In contrast, LUNA specific CD4+ T cells were not cytotoxic at any E∶T ratio examined (range 10∶1–80∶1) (Figure 5C), although importantly, the LUNA specific T cells remained antigen reactive by IFNγ specific ELISPOT assays (Figure 5D).
The CD4+ T cell response is potentially composed of multiple subsets of CD4+ T cells with distinct functions and characteristic cytokines they produce. The HCMV specific CD4+ T cell response is characterised as being almost exclusively Th1 mediated, secreting IFNγ in response to lytic antigens such as gB and IE. However, it is interesting to note that parallels with other herpes viruses may be apparent: CD4+ T cells specific for a latent protein of EBV (LMP1) are able to secrete the immunosuppressive cytokine cIL-10 which is thought to play a role in evading immune responses during latent infection and maintenance of EBV latency [37]–[39]. Consequently, we analysed the cytokine profile of UL138 and LUNA specific CD4+ T cells following antigen stimulation and compared this to the well characterized response made by gB specific CD4+ T cells using a multi-analyte Th1/Th2 cytokine assay which measured 11 cytokines simultaneously.
CD4+ T cells specific to the lytic protein gB induced high levels of the classic Th1 type cytokines IFNγ, TNFα and IL2 as expected (Figure 6A) [34], [35]. Both donor CMV300 and CMV 305 responded to the same UL138 peptide (LNVGLPIIGVMLVLI). Interestingly, stimulation of UL138 specific CD4+ T cell lines (from donors CMV 300 and CMV305) resulted in a cytokine secretion profile with increased heterogeneity when compared to that observed from gB specific T cells. Specifically we detected the secretion of IFNγ, TNFα, IL-2, IFNβ, IL-6, IL-8 and low levels of IL-4 and IL-5 (Figure 6A). Interestingly, UL138 specific CD4+ T cells also produced high levels of the immunomodulatory cytokine cIL-10 in both these donors – an event not seen in response to gB peptide stimulation. Similarly, stimulation of donor CMV 300s' LUNA specific CD4+ T cells also produced a heterogeneous range of cytokines including, pertinently, cIL-10 (Figure 6A).
The detection of cIL-10 producing T cells that appeared restricted to the recognition of latently expressed HCMV antigens was intriguing. CD4+ T cell production of cIL-10 has been described in a subset of helper T cells associated with immune regulatory functions described as T regulatory cells (Treg). Treg are potent modulators of immune responses and exert their effects at least partly via the production of immunomodulatory cytokines, TGFβ and cIL-10. Consistent with this phenotype, both UL138 and LUNA specific CD4+ T cells (from donor CMV300) produced TGFβ upon peptide stimulation, while gB specific CD4+ T cells from the same donor did not (Figure 6B). A parallel analysis for IFNy and cIL-10 on the T cell lines was performed to assess the frequency of IFNγ and cIL-10 producing cells in the line. As previously, gB, UL138 and LUNA specific CD4+ T cells all produced IFNγ upon antigenic stimulation (Figure 6C). Furthermore, UL138 and LUNA, but not gB-specific, CD4+ T cells were again shown to secrete cIL-10 (Figure 6D). However, the ELISPOT assays also indicated that the frequency of cIL-10 producing UL138 and LUNA specific T cells was less than the IFNγ producing frequency.
What was not clear from these initial analyses was whether the UL138 specific T cell response we observed was composed of polyfunctional T cells which secrete both pro-inflammatory and immunomodulatory cytokines, or whether the IFNγ and cIL-10 producing T cells were actually separate populations.
To address this, PBMC were stimulated with peptide and then assayed for the production and co-expression of IFNγ and cIL-10 from CD4+ T cells by intracellular cytokine staining and flow cytometry. These data show that the stimulation of PBMC from both donors with UL138 peptide resulted in the generation of IFNγ and cIL-10 producing cells, however, the UL138 specific cells were composed of separate populations of CD4+ T cells that secreted either IFNγ or cIL-10 and not both (Figure S4A). These data were recapitulated in PBMC from donor CMV300 that exhibited a LUNA specific response composed of separate IFNγ and cIL-10 producing CD4+ T cells. A further donor, CMV317, with a known response to both UL138 and LUNA (Table 1, Figure 1 and Figure 2) also showed that both UL138 and LUNA specific CD4+ T cells were again composed of separate populations of IFNγ and cIL-10 producing CD4+ T cells. Taken together these data show that different subsets of T cells within the CD4+ T cell response can be detected and characterised by their cytokine expression profile (summarised in Figure S4B).
Finally, CD4+ T cell lines specific to gB and UL138 were also generated from donor CMV305 and, 14 days post in vitro expansion, the production of IFNγ and cIL-10 examined in a similar manner. Consistent with our ex vivo analyses, the expanded gB specific CD4+ T cells were again composed solely of IFNγ producing cells whereas UL138 specific CD4+ T cells were again composed of separate populations that produced either IFNγ or cIL-10 (Figure S4C).
We reasoned that the observation that UL138 specific CD4+ T cells, but not gB specific CD4+ T cells, secreted the immunomodulatory cytokines cIL-10 and TGFβ upon stimulation with cognate peptide could potentially result in suppression of the host T cell response. If this was the case then the supernatants from UL138 specific CD4+ T cells could impact upon the proliferative response of polyclonally activated CD4+ T cells. PBMC from three donors were polyclonally stimulated with anti-CD3/CD28 beads in the presence of supernatant from T cells stimulated for 48 hours with either UL138 or gB specific peptides or media control. Proliferation of CD4+ T cells was measured by dye dilution using flow cytometry (Figure 7). Consistent with the prediction that the secretion of cIL-10 and TGF-β was indicative of Treg phenotype we observed that the supernatant from cells stimulated with UL138 (but not gB) peptide, suppressed CD4+ T cell proliferation (p<0.01; n = 3) (Figures 7B and C). Supernatant from cells stimulated with UL138 peptide were treated with neutralizing antibodies specific for cIL-10 and TGFβ. The results show that proliferation was partially restored by either neutralizing antibody and in combination proliferation was fully restored to the level of the control (Figure 7D).
The production of the immunomodulatory cytokines cIL-10 and TGFβ, by CD4+ T cells has been shown to be associated with the function of a subset of immunosuppressive CD4+ T cells termed regulatory T cells (Treg) [40]. Indeed, it has been shown previously that the EBV specific CD4+ T cell response includes a Treg component specific to a viral gene product expressed during EBV latency [37]. Furthermore, it has also been shown that HCMV can induce the expansion of virus specific CD4+ T cells that express phenotypic markers associated with Treg [41].
Thus our data so far was highly suggestive that the detection of latent antigens was concomitant with the development of subset of T cells with Treg phenotype. To address this definitively we next examined whether the expanded CD4+ T cells specific to UL138, LUNA or gB were populated, in part, with phenotypically defined Treg cells – based on CD4+ CD25hi FoxP3+ expression [42]–[46]. CD4+ T cell lines specific to UL138 (Figure 8A) and gB (Figure 8B) were generated and on day 14 of the expansion protocol assessed for stable expression of CD25. The data clearly showed that UL138 but not gB specific CD4+ T cells expressed CD25 after 14 days in vitro culture (p<0.01, n = 3) (Figure 8C). We further characterised these cells by determining FoxP3 expression as an indicator of Treg cells. Having established the staining conditions to identify CD4+CD25hiFoxP3+ T cells using whole PBMC (Figure 8D) the CD25, FoxP3 phenotype of CD4+ T cells specific to gB, LUNA and UL138, or gB and UL138 from donors CMV300 or CMV305, respectively, were determined (Figure 8E). The results clearly show that CD4+CD25hi cells expressing FoxP3, were detected in in vitro T cell cultures of UL138 and LUNA specific T cells, but not those specific for gB, in both donors tested. These data showed that a subset of the cells specific for UL138 and LUNA, but not gB, present in the expanded T cell cultures, expressed phenotypic markers consistent with Treg cells.
Although unlikely, it was necessary to exclude the possibility that the culture conditions to produce UL138 and LUNA-specific T cell lines induced or favoured cIL-10 producing T cells in vitro. Thus, we assayed for cIL-10 production from T cells isolated directly ex vivo without prior in vitro expansion. In addition, this nature of this analysis allowed us to assess if cIL-10 production by UL138 and LUNA specific T cells was common in a larger panel of donors. Parallel ELISPOT assays detecting IFNγ, cIL-10, IL-4 and IL-17 were thus performed on 13 HCMV seropositive donors using peptides derived from both latent and lytic antigens. The ELISPOT assays for each cytokine (IFNγ, cIL-10, IL-4 and IL-17) were enumerated and the cytokine frequencies were used to determine the percentage of each individual cytokine to the total antigen specific response.
Having first confirmed that we could detect all four cytokines in the PBMC from all 13 donors following stimulation with PHA (Figure 9A) we next assayed the effect of specific peptides on cytokine production. As we have demonstrated repeatedly, stimulation with gB, IE, UL138 or LUNA elicited IFNγ responses. IE stimulation was dominated by IFNγ responses while gB elicited predominantly IFNγ however, we note that we did detect a few donors having a small cIL-10 response and 2 donors having a more substantial cIL-10 responding T cell population. In contrast, the cytokine responses to both LUNA and UL138 were much more heterogeneous with most donors producing both IFNγ and cIL-10 and some donors having a predominantly cIL-10 response (Figure 9B). Absolute values (SFU/106 cells) for the total T cell response (IFNγ, cIL-10, IL-4 and IL-17) to each antigen are also shown (Figure S5).
Finally we expanded these analyses to test whether other viral gene products expressed during latency were responsible for a similar T cell phenotype observed with UL138 and LUNA – namely, UL111A (an HCMV homologue of cIL-10) and US28 (an HCMV chemokine receptor homologue) [27]–[32], [47]–[51]. We stimulated the same 13 donors with overlapping ORF peptide pools to UL111A and US28, and measured IFNγ, cIL-10, IL-4 and IL-17 cytokine production in separate ELISPOT assays (Figure 9B). Absolute values (SFU/106 cells) for the total T cell response (IFNγ, cIL-10, IL-4 and IL-17) to each antigen are also shown (Figure S5). These results clearly show that donors did have both UL111A and US28 specific T cells and, importantly, while IFNγ responses could be measured the dominant cytokine to these antigens was cIL-10. Furthermore, we note that some donors also had small IL-4 or IL-17 cytokine responses. Taken together these data clearly show that there is a circulating population of CD4+ T cells detectable directly ex vivo that recognise latently expressed HCMV antigens and which have a phenotype consistent with the production of the immunomodulatory cytokine cIL-10.
The results we present here represent the first comprehensive analysis of T cell responses to those viral proteins associated with latent HCMV infection. In general, and in contrast to antigens solely associated with virus lytic infection (such as IE, gB and pp65), our results show that healthy seropositive donors have robust T cell responses to all the latency-associated antigens we analysed, which are dominated by CD4+ T cells. As expected, these CD4+ T cells recognise cells lytically infected with HCMV but, importantly, also recognise latently infected monocytes.
There have been few other analyses of T cell responses to HCMV encoded ORFs associated with latent infection. In a total proteome screen for HCMV-specific T cell responses undertaken by Sylwester et al (2005) only one donor out of 33 was identified as having a UL138-specific CD4+ and CD8+ T cell response whereas LUNA was not included in their analysis [21]. Similarly, an independent analysis of UL138 identified UL138-specific CD8+ T cell responses, but only in individuals who expressed the HLA-B3501 haplotype [33]. Our study also included four HCMV seropositive donors who expressed HLA-B3501, but none of these individuals had detectable CD8+ T cell responses to UL138 in our hands. These differences could be due to the methods used to detect antigen specific T cells - our studies used ELISPOT assays to screen directly ex vivo, in contrast to an in vitro antigen stimulation to induce T cell expansion and pre-enrichment prior to detection of IFNγ used in the previous study [33]. Also, in contrast to Tey et al (2010), who reported an absence of CD4+ T cell responses to either UL138 or LUNA, we observed robust CD4+ T cell responses to these antigens in healthy donors. Again, it is likely that experimental differences between the studies, such as the size of the individual peptides, the use of the ELISPOT assays versus intracellular cytokine detection and the duration of the assay (48 hours v 6 hours restimulation) may account for these discrepancies.
Consistent with the knowledge that UL138 and LUNA are also expressed during lytic infection, our data clearly showed that CD4+ T cells specific to these viral proteins were able to recognize lytically infected moDCs. Furthermore, UL138 specific CD4+ T cells expanded in vitro were also able to mediate cytotoxicity against peptide-loaded autologous target cells. This recognition of lytically infected DCs, and concomitant secretion of IFNγ, occurred despite the expression of those viral genes associated with lytic infection that are known to modulate immune responses [3], [52]–[55]. However, this is not inconsistent with numerous studies which have shown potent anti-viral CD4+ T cell responses to other antigens such as gB and IE (expressed only during lytic infection) in HCMV infected cells, despite expression of the known viral immune-evasins [34], [56].
Our observations that CD4+ T cells specific for UL138 could also recognise latently infected cells and secrete IFNγ leads to an obvious conundrum: this ability of the host to recognise latently infected cells carries the risk that the latently infected cells should become targets for immune clearance. However, it is already known that HCMV is able to modify the latently infected cell itself in order to reduce T cell recognition and activation. During latent infection expression of viral UL111A (vIL-10) results in down regulation of MHC class II and diminished CD4+ T cell recognition [49] and latent infection of CD34+ bone marrow progenitor cells induces release of cIL-10 and TGFβ which decreases CD4+ T cell IFNγ production and cytotoxicity [14]. We have not been able to determine if LUNA protein expression in latently infected cells can be recognized by LUNA specific T cells and thus this remains an open question.
The results presented here now demonstrate that both UL138 and LUNA-specific CD4+ T cells also, themselves, secrete the immunomodulatory cytokine cIL-10, in direct contrast to CD4+ T cells specific for gB and IE antigens (which are expressed solely during lytic infection). Interestingly, our analysis of the latency-associated antigens US28 and UL111A also showed a skewing of T cell responses towards CD4+ T cells which secreted cIL-10 suggesting that immune evasion during latency is a complex interplay between the microenvironment around the latently infected cell and the properties of the immune cells recruited to it.
In the donors analysed directly ex vivo, there was a clear bias towards CD4+ T cells which secreted cIL-10 in response to the latency associated HCMV gene products UL138, LUNA, UL111A and US28. Not all donors respond to every latency associated gene product, however, when all four antigens (UL138, LUNA, US28 and UL111a) were taken together for the 13 donors tested two donors made no responses, one donor made an IFNã response and the remaining 10 donors had an cIL-10 response to at least one of the latently expressed antigens. Indeed, if a donor did not make a T cell response to any latently expressed antigen there is no requirement for cIL-10 producing latent antigen specific T cells. Regarding the single donor we found that makes an IFNã response but no associated cIL-10 response to latent antigens it is possible that there are other latent antigens that have yet to be recognised. We have recently published that UL144 is also expressed during latency [57] and it is possible that this donor makes cIL-10 to this antigen. Alternatively, it is possible that some donors do make antiviral responses to latent antigens that are not balanced with a cIL-10 T cell response and that these individuals may turn out to have lower latent viral loads than individuals with higher frequency cIL-10 responses.
Interestingly, EBV also induces high frequencies of CD4+ T cells specific to latent antigens; targeting EBNA1 which suppress the proliferation and cytokine production of both CD4+ and CD8+ T cells [58] and LMP1 that also secrete cIL-10 [37]–[39]. cIL-10 secreting CD4+ T cells are known to perform an immunomodulatory role in the immune response, often functioning to restrict immune activation [59]–[61] and are a classical signature of Treg cells [40], [59]. It was interesting to note that the UL138 and LUNA T cell lines expanded in vitro more slowly than the gB specific T cell lines (data not shown). We were also able to show that the suppression of polyclonally activated CD4+ T cell proliferation was due to cIL-10 and TGFβ secreted in the supernatant from UL138 specific T cells. In MCMV latency, cIL-10 producing CD4+ T cells have been isolated from salivary glands. Furthermore, in cIL-10 knockout mice (or after IL-10R blockade) the latent MCMV load is reduced [62] with a concomitant increase in memory MCMV-specific T cell frequency being observed. These observations are consistent with the view that cytomegalovirus may induce cIL-10 producing CD4+ T cells to prevent latently infected cells from being recognised by the immune system [63].
In contrast to EBV, it was not known whether viral gene products expressed during HCMV latency generate Treg cells. Lytic HCMV (IE and pp65) antigen specific Treg cells have been described, particularly enriched in kidney transplantation patients that had recurring HCMV reactivation events. The authors suggested that frequent episodes of antigen stimulation might drive the Treg phenotype and they further demonstrate that the antigen specific Treg cells had the same TCR as effector T cells, suggesting a common lineage [41]. IE and pp65 specific Treg cells were also isolated from normal healthy donors but at lower frequencies than in recurrent patient groups. In this study we examined IE and gB specific CD4+ T cells and in some healthy individuals we were able to identify cIL-10 producing cells in addition to predominant IFNγ secreting cells, which is in agreement with the Schwele observations. However, key to this study, is the detection of cIL-10 production by UL138 and LUNA specific CD4+ T cells in a larger number of donors than was seen with IE and gB specific CD4+ T cells. Indeed, the number of donors that had cIL-10 secreting T cells specific for US28 and UL111A was striking and within individual donors was often dominant over IFNγ producing T cells of the same specificity.
It is unclear why CD4+ T cell responses to HCMV lytic antigens appear to be dominated by IFNγ producing cells, while CD4+ T cells, which recognise antigens, expressed during latent infection predominantly secrete cIL-10. Many factors are likely to impact upon the type of CD4+ T cell response generated to a particular antigen, and this is known to include the cytokines present in the microenvironment during T cell activation. During latent infection, viral gene expression in CD34+ cells is highly restricted and associated with secretion of immunomodulatory cytokines cIL-10 and TGFβ [14]. This immunosuppressive microenvironment may also have an impact on the generation of CD4+ T cells activated during latent phases of infection. Specifically, the CD34+ mediated secretion of cIL-10 and TGFβ, may result in the generation of cIL-10 producing, immunomodulatory CD4+ T cells specific to latent antigens but not those expressed solely during lytic infection [64]–[67]. It is highly plausible that these effects act in concert with the known functions of latency-associated UL111A (vIL-10) which has been shown to promote MHC class II down-regulation and inhibit CD4+ T cell activation [29]–[31], [49] as well as restrict the ability of latently infected myeloid cells to differentiate. This ability to modulate the ability of the infected cell to function as a professional antigen presenting cell as has been seen during HSV infection of plasmacytoid DCs [68], [69].
It has been shown that that some HCMV seropositive donors generate cIL-10 secreting T cells to lytic HCMV antigens (such as pp65) [41], in agreement with these observations we have also seen cIL-10 producing T cells specific for gB and IE in a small number of donors tested. Schwele et al clearly demonstrated that these cIL-10 producing T cells were generated at a higher frequency in reactivating transplant individuals (speculated to be due to repeated antigenic stimulations) and, as our cohort were normal healthy donors, this probably accounted for the lower frequency of detection in our analysis. It might be expected that in older HCMV seropositive donors, who have carried the virus for many years, that you would see the generation of cIL-10 producing CD4+ T cells to lytic antigens and maybe an increase in frequency to latent antigens. Our data highlights the consistent generation of cIL-10 producing CD4+ T cells, in most normal healthy donors, to antigens expressed in latency and thus the possibility that this is important in preventing the immune clearance of latently infected cells.
We believe that there are potential clinical implication from these findings, in the case of bone marrow transplantation (D+/R−) if it were possible to eliminate or drastically reduce the latent viral load prior to transplantation this could either prevent reactivation or substantially reduce reactivation loads. We speculate that since the T cell response to the latent antigens is composed of both anti-viral (Th1) and immunesuppressive activities the neutralization of cIL-10/TGFβ may allow Th1 type latent specific T cells to recognize and results in the elimination of latent CD34+ cells.
In conclusion, based on this and other studies we suggest that a number of direct and indirect immune suppressive mechanisms may act together to help maintain sites of HCMV latency: Virally encoded UL111A (vIL-10) expressed during latency down regulates MHC class II expression on APCs restricting T cell recognition of latently infected cells [49]; latent infection of CD34+ cells results in increases in cIL-10 and TGFβ in the cell secretome, which act to suppress antiviral immune responses in the microenvironment of latently infected cells [14] and now we show that CD4+ T cells specific for viral latency-associated gene products, themselves secrete cIL-10 which helps suppress antiviral effector functions. This biasing of the immune response by latency-associated antigens, to elicit CD4+ T cells that secrete cIL-10, may assist in the maintenance of the latent reservoir and lifelong carriage of HCMV in vivo.
Ethical permission for this project was granted by the Cambridgeshire 2 Research Ethics Committee (REC reference 97/092). Informed written consent was obtained from all of the volunteers included in this study prior to providing blood samples.
HCMV serostatus of 23 healthy volunteers was determined using a commercial HCMV specific IgG ELISA kit (Captia, Trinity biotech, Ireland). The assay was performed according to manufacturer's instructions.
Human leukocyte antigen typing was performed for donors regardless of HCMV serostatus. All MHC class I alleles (HLA-A, HLA-B, HLA-C) and HLA-DR and HLA-DQ MHC class II alleles were typed by molecular methods by Helen Stevens (Cambridge Institute for Medical Research, UK) (Table S1).
Sequences for viral proteins from the clinical HCMV strain Merlin were used and peptides constructed as sequential 15 amino acid peptides with 10 amino acid overlap, spanning UL138 (Table S2A), LUNA (), US28 (Table S2C) and UL111A (Table S2D) gB, IE and pp65 from Proimmune (UK). Peptides were reconstituted and stored according to manufacturer's instructions to give a storage concentration of 40 mg/ml. Individual peptides were further diluted in RPMI 1640 (PAA laboratories, Austria) to create a stock of 1 mg/ml and a working concentration of each peptide of 40 µg/ml and stored at −80°C. Peptide pools were made for screening purposes from the single peptides, and were constructed to give 2 µg/ml of each individual peptide. These peptide pools were then stored in 100 µl aliquots at −80°C.
Venous blood was collected in heparin sodium (100 IU/ml), diluted 1∶2 with RPMI-1640 containing no serum (PAA laboratories, Austria) supplemented with 100,000 IU/ml penicillin, 100 mg/ml streptomycin, and 2 mmol/ml L-glutamine (RPMI-wash). Peripheral blood mononuclear cells (PBMC) were isolated by Lymphoprep (Axis-Shield, Norway) centrifuged at 800 g for 15 minutes. Autologous serum was removed from the interface and incubated at 60°C for 30 minutes in a water bath to inactivate complement.
LCL lines were established according to published protocols [70].
ELISPOT plates were prepared, coated and blocked according to manufacturer's instruction (EBioscience). PBMC directly ex vivo, previously frozen, or depleted of either CD4+ or CD8+ T cells by magnetic activated cell sorting (MACS), were plated 3.0×105 cells in 100 µl RPMI-10 per well (of a 96 well Multiscreen IP sterile plate (Millipore, UK)). Plates were incubated for 48 hours at 37°C 5% CO2, and developed according to manufacturer's instruction. Plates were read using an ELISPOT plate scanner (ELISPOT Reader System, AID) and spots enumerated using ImageJ (National Institutes of Health).
PBMC were depleted of either CD4+ or CD8+ T cells by MACS using either anti-CD4+ or anti-CD8+ direct beads (Miltenyi, U.K.), according to manufacturer's instructions and separated on LS columns (Miltenyi, U.K.). Efficiency of depletion was determined by staining cells with either anti-CD4 or anti-CD8 antibodies and analysed by flow cytometry. Depletions performed in this manner resulted in 0.1–0.8% CD4+ cells and 0.3–0.8% CD8+ cells, respectively.
CD4+ T cells were purified by MACS using anti-CD4 direct beads (Miltenyi, UK) and separated on LS columns (Miltneyi, UK) according to manufacturer's instructions which resulted in CD3+ CD4+ mean cells purities of 98.1% (range 95.6–99.8%) as determined by flow cytometry. PBMC (5.0×106) were incubated with 100 µl of the peptide of interest (2 µg/ml) for two hours at 37°C before irradiation using a sealed source irradiator for 30 minutes. These cells were then washed in PBS, resuspended in RPMI-1640 supplemented with 10% autologous donor serum, 100,000 IU/ml penicillin, 100 mg/ml streptomycin, and 2 mmol/ml L-glutamine. PBMC were then transferred 5.0×105 cells per well of a 96 well round bottomed microtitre plate. MACS purified CD4+ T cells were then added 1.0×104 cells per well and incubated at 37°C for 14 days. RPMI-1640 supplemented with 10% autologous donor serum, 100,000 IU/ml penicillin, 100 mg/ml streptomycin, and 2 mmol/ml L-glutamine and 15 IU/ml recombinant IL-2 (National Institute of Biological Standards and Control, U.K) was added 50 µl per well on days 2, 8 and 12 of culture.
The production of cytokines by peptide specific T cell lines was determined using Flowcytomix Th1/Th2 11plex kit, or using simplex kits for specific cytokines (Bendermed systems, Netherlands) according to manufacturer's instructions. Samples were analysed using a BD FACSort and data analysed using Flowcytomix pro software (Bendermed systems, Netherlands).
Supernatants from peptide stimulated CD4+ T cells were assayed for the presence of TGFβ by ELISA, according to manufacturer's instructions (R&D systems, U.K.).
PBMC were washed in PBS and 1.5×106 cells were resuspended in 500 ul RPMI-1640 supplemented with 10% autologous donor serum, 100,000 IU/ml penicillin, 100 mg/ml streptomycin, and 2 mmol/ml L-glutamine, in polypropylene FACS tubes (B.D., U.K.). Cells were then incubated with 100 µl (40 µg/ml) of the peptide of interest and incubated for 16 hours at 37°C +5% CO2. Post incubation Brefeldin A and Monensin (Biolegend, U.K.) were added to the cultures according to the manufacturers instructions, and the tubes incubated for a further 4 hours. Post incubation cells were stained with a LIVE/DEAD fixable dead cell stain kit (Invitrogen, U.K.), according to manufacturers instructions. Cells were then surface stained with anti-CD3 PE-Cy7 (B.D., U.K.) and anti-CD4 PerCP Cy5.5 (B.D., U.K.) monoclonal antibodies according to manufacturers instructions. Intracellular cytokines were fixed and permeablised using the FoxP3/Transcription factor staining buffer set (EBioscience, U.K.). Intracellular cytokines were stained using anti-IFNγ alexafluor 488 (Biolegend, U.K.) and anti-IL10 PE (Biolegend, U.K.) and acquired using the FACS Canto II (B.D., U.K.) and analysed using FlowJo software (Treestar, U.S.A).
Expanded T cell lines or PBMC were assayed for the presence of Treg using the Human Regulatory T cell staining kit (EBioscience, U.K.), according to manufacturer's instructions.
Cultured peptide specific T cell lines were used in a standard Cr51-release assay against both HLA matched and mis-matched B cell lines as target cells as previously described [70].
PBMC were labelled with cell trace violet proliferation kit for flow cytometry (Life technologies, U.K.), according to manufacturer's instructions. Cells were then resuspended in RPMI-1640 supplemented with 10% autologous donor serum, 100,000 IU/ml penicillin, 100 mg/ml streptomycin, and 2 mmol/ml L-glutamine. Alternatively, PBMC were resuspended in supernatant from PBMC cultures stimulated for 48 hours with gB or UL138 peptides, and plated 1.0×105 per well of a round bottom 96 well plate and incubated for 1 hour at 37°C +5% CO2. For neutralisation assays, supernatants were treated with neutralising antibodies for cIL-10, TGFβ or both, or with the relevant isotype control antibodies at a final concentration of 20 ng/ml for one hour prior to the addition of the violet labelled PBMC. Post incubation cells were stimulated with Dynabeads® Human T-Activator CD3/CD28 (Life technologies, U.K.) at a bead to cell ratio of 1∶200. Alternatively cells remained unstimulated and all wells were adjusted to a total volume of 200 µl and incubated for 5 days at 37°C +5% CO2. Post incubated cells were harvested and washed in PBS prior to staining with anti-CD3 PE-Cy7 (B.D., U.K.) and anti-CD4 FITC (B.D., U.K.) antibodies. Cells were then stained with 7-amino- Actinomycin D (Calbiochem, U.K.). Cells were then washed in PBS and resuspended in FACS buffer and acquired on the FACS Canto II and analysed FlowJo software.
Monocytes were isolated from PBMC using anti-CD14 MACs beads according to the manufactures instructions (Miltneyi, UK). The purity of isolated monocytes was determined by flowcytometric detection of CD14+ cells, resulting in mean CD14+ populations of 98.1% (range 97.4–98.9%, n = 5). Purified monocytes were adhered to tissue culture plates overnight in X-vivo15, medium was then changed to X-vivo15 supplemented with 2.5 mM L-glutamine, 500 IU/ml IL-4 (Peprotech, UK) and 1000 IU/ml granulocte-macrophage colony stimulating factor (GM-CSF) (Peprotech, UK) and incubated for a further 3 days at 37°C + 5% CO2. Post incubation cells were washed in PBS and fresh X-vivo15 supplemented with 2.5 mM L-glutamine, 500 IU/ml IL-4 (Peprotech, UK) and 1000 IU/ml GM-CSF (Peprotech, UK) was added for a further 3 days. Post incubation, the cells were washed in PBS and matured by the addition of 1 ml X-vivo15 per well supplemented with 2.5 mM L-glutamine and 50 ng/ml LPS for 48 hours. Cell surface phenotype of monocytes and the iDC and mDC cells derived from them was determined by flow cytometry using anti-CD14, CD80, CD86 and HLA-DR, as expected iDC and mDC lost CD14 and gained CD83 expression, MHC Class II, CD80 and CD86 were all upregulated (Figure S1). In vitro differentiated monocyte derived dendritic cells were infected at MOI 5 for 3 hours at 37°C + 5% CO2 in X-vivo15 supplemented with 2.5 mM L-glutamine. Post incubation supernatant was removed and 1 ml X-vivo15 supplemented with 2.5 mM L-glutamine added per well and further incubated at 37°C + 5% CO2. Media was changed every 3 days post infection and cells used as antigen presenting cells 5 days post infection in the presence or absence of cognate peptide. Infection was confirmed by immunofluorescence and RT-PCR (see below).
Monocytes were isolated from PBMC using anti-CD14 MACs beads according to the manufactures instructions (Miltneyi, UK) and adhered in tissue culture plates overnight in X-vivo15 supplemented with 2.5 mM L-glutamine overnight at 37°C + 5% CO2. Post incubation cells were washed in PBS and infected with TB40e UL32GFP at an MOI of 5 for 3 hours at 37°C + 5% CO2. Post incubation cells were washed in PBS and 1 ml fresh X-vivo15 supplemented with 2.5 mM L-glutamine added and incubated 37°C + 5% CO2 for 10 days. Medium was changed every 3 days post infection and cells used as antigen presenting cells 10 days post infection in the presence or absence of peptide. Establishment of latency was confirmed at day ten by RT-PCR.
Cells were fixed in 4% PFA in PBS for 10 minutes at room temperature, washed twice in PBS prior to the addition of 0.1% Triton-X in PBS for ten minutes at room temperature. 300 µl mouse anti-IE antibody (1∶1000; Millipore, UK) per well for 1 hour at room temperature, followed by PBS washes and then stained with anti-mouse Alexafluor 594 (1∶1000; Invitrogen, UK) and DAPI (1∶100; Invitrogen, UK) in 300 µl PBS for 1 hour at room temperature in the dark. After washing cells were analysed immediately by fluorescent microscopy.
RNA was extracted from in vitro cell infections using a previously published method [71]. Briefly, adherent cells were washed in chilled PBS and 1 ml TRIZOL (Invitrogen, UK) added per well (for a maximum of 1×106 cells), adherent cells were removed using a cell scraper. RNA samples were DNase treated using the RQ1 RNase free DNase kit (Promega, UK) according to manufacturer's instructions. Samples were then reversed transcribed using the Reverse transcription system (Promega, UK) according to manufacturer's instructions. PCR was performed to amplify a range of viral transcripts associated with either lytic or latent infection, or cellular genes: GAPDH forward GAGTCAACGGATTTGGTCGT and GAPDH reverse TTGATTTTGGAGGGATCTCG [72]; IE forward GGACCCTGTAATCCTGACG and IE reverse ATCTTTCTCGGGGTTCTCGT [73]; UL138 forward TGCGCATGTTTCTGAGCTC and UL138 reverse ACGGGTTTCACAGATCGAC [28]; LUNA forward ATGACCTCTCCTCCACACC and LUNA reverse GGAAAAACACGCGCGGGGGA [27] (all primers obtained from Sigma-Adlrich, UK). 45 cycle PCR Biomix red (Bioline, UK), was performed using: 95°C 1 minute, 55°C 40 seconds, 72°C 1 minute.
|
10.1371/journal.pgen.1005255 | Genetic Mechanism of Human Neutrophil Antigen 2 Deficiency and Expression Variations | Human neutrophil antigen 2 (HNA-2) deficiency is a common phenotype as 3–5% humans do not express HNA-2. HNA-2 is coded by CD177 gene that associates with human myeloproliferative disorders. HNA-2 deficient individuals are prone to produce HNA-2 alloantibodies that cause a number of disorders including transfusion-related acute lung injury and immune neutropenia. In addition, the percentages of HNA-2 positive neutrophils vary significantly among individuals and HNA-2 expression variations play a role in human diseases such as myelodysplastic syndrome, chronic myelogenous leukemia, and gastric cancer. The underlying genetic mechanism of HNA-2 deficiency and expression variations has remained a mystery. In this study, we identified a novel CD177 nonsense single nucleotide polymorphism (SNP 829A>T) that creates a stop codon within the CD177 coding region. We found that all 829TT homozygous individuals were HNA-2 deficient. In addition, the SNP 829A>T genotypes were significantly associated with the percentage of HNA-2 positive neutrophils. Transfection experiments confirmed that HNA-2 expression was absent on cells expressing the CD177 SNP 829T allele. Our data clearly demonstrate that the CD177 SNP 829A>T is the primary genetic determinant for HNA-2 deficiency and expression variations. The mechanistic delineation of HNA-2 genetics will enable the development of genetic tests for diagnosis and prognosis of HNA-2-related human diseases.
| Human neutrophil antigen 2 (HNA-2) is coded by CD177 gene that involves in human myeloproliferative disorders. HNA-2 expression varies among humans and about 3–5% people lack HNA-2 expression. HNA-2 deficient people are susceptible to produce HNA-2 alloantibodies, which play a pathological role in various human diseases including transfusion-related acute lung injury, neonatal alloimmune neutropenia, autoimmune neutropenia, drug-induced immune neutropenia, and graft failure following marrow transplantation. The level of HNA-2 expression has also been identified as a prognostic biomarker for the gastric cancer. Although HNA-2 is among the most important clinical antigens, the underlying genetic mechanism of HNA-2 deficiency and expression variations has remained unknown. Here, we demonstrate that HNA-2 deficiency and expression variations are primarily caused by a novel CD177 genetic polymorphism that disrupts HNA-2 expression. The illumination of genetic mechanism for HNA-2 deficiency and expression variations will enable the development of effective HNA-2 genetic tests. Our findings will facilitate prognosis and diagnosis of HNA-2-related human disorders.
| Transfusion-related acute lung injury (TRALI) is associated with the transfusion of leukocyte alloantibodies from donors or associated with the presence of alloantibodies in recipients of blood [1,2]. Alloantibodies against human neutrophil alloantigenes (HNAs) are a very strong trigger for the development of TRALI [1,2]. Human neutrophil antigen 2 (HNA-2) alloantibodies have been linked to the induction of TRALI and various pulmonary reactions [3–6] while anti-HNA-3 alloantibodies are frequently implicated in severe and fatal TRALI [7]. Animal models have firmly established a pathological role for HNA-2 alloantibodies in TRALI [8,9]. Furthermore, HNA-2 alloantibodies have been implicated in multiple human disorders such as neonatal alloimmune neutropenia, autoimmune neutropenia, drug-induced immune neutropenia, and graft failure following marrow transplantation [10–13]. Accordingly, HNA-2 is among the most important clinical antigens.
HNA-2 is heterogeneously expressed on subpopulations of neutrophils and approximately 3–5% Americans do not express HNA-2 [14]. HNA-2 deficient subjects are predisposed to the production of HNA-2 alloantibodies when exposed to the HNA-2 antigen during blood transfusion, pregnancy, and bone marrow transplantation. HNA-2 is encoded by the CD177 gene that contains nine exons at Chromosome 19q13.31 region, where a CD177 pseudogene highly homologous to CD177 between exon 4 and 9 is also located (Fig 1A) [15–17]. The genetic studies of CD177 were significantly hampered by the presence of CD177 pseudogene [18,19]. HNA-2 is also known as PRV-1 as CD177 mRNA is over-expressed in polycythemia rubra vera patients [20]. CD177 has an open reading frame of 1311 nucleotides that encode 437 amino acids with a signal peptide of 21 residues. HNA-2 (or CD177) is expressed as a GPI-linked receptor with a mature peptide consisting of residue 22 to 408 [15,21]. HNA-2 plays important roles in neutrophil functions and myeloid cell proliferation. The interaction between HNA-2 and PECAM-1 facilitates neutrophil transendothelial migration [22,23]. In addition, HNA-2 is required for the attachment of proteinase 3 (PR3) to neutrophils [24–27], which plays a pivotal role in PR3-ANCA-mediated neutrophil activation [28]. CD177 mRNA levels are elevated in several conditions associated with increased neutrophil counts [14,29]. Furthermore, elevated levels of neutrophil CD177 mRNA are associated with increased neutrophil production and quantitation of neutrophil CD177 mRNA is a diagnostic tool for polycythemia vera [14]. Moreover, the level of HNA-2 expression has been identified as a prognostic biomarker for gastric cancer [30].
The CD177 non-synonymous coding SNPs (cSNPs) were reported to associate with HNA-2 expression variations, however, the effect of those non-synonymous CD177 coding SNPs on HNA-2 expression was unknown [18,31,32]. CD177 mRNA splicing variants were found in two HNA-2 deficient donors but it remains inconclusive whether CD177 splicing abnormality was actually responsible for HNA-2 deficiency [33]. Therefore, the underlying genetic mechanism of HNA-2 deficiency has remained elusive since the observation of HNA-2 deficiency four decades ago [10]. Elucidation of the molecular genetics and basis of the HNA-2 deficiency is a prerequisite for the use of effective genetic tests in prognosis and diagnosis of HNA-2-related human diseases. In the current study, we demonstrated that a novel nonsense CD177 coding SNP 829A>T is the primary genetic determinant for HNA-2 deficiency and expression variations in humans.
The percentages of neutrophils expressing HNA-2 were heterogeneous among normal healthy blood donors in flow cytometry analysis (Fig 1B). In 294 normal healthy blood donors, the percentage of HNA-2-positive neutrophils ranged from 0.0% to 97.8%. Among 294 blood donors, we have identified 11 donors (or 3.7%) deficient for HNA-2 and the percentage of HNA-2 deficient blood donors is consistent with those previously reported [6,10,34].
Copy number variations (CNVs) are the primary cause of human neutrophil antigen 1 (HNA-1 or FcγRIIIB) deficiency and expression variations [35–38]. To investigate whether CD177 CNVs are involved in HNA-2 deficiency, we determined CD177 CNVs using TaqMan CNV assay kit Hs01327659_cn with the probe targeting the unique CD177 exon 1 region (Fig 1A). Among 294 human subjects, 95.2% (280/294) of subjects were two-copy CD177 carriers and 4.8% (14/294) were three-copy CD177 carriers. No human subjects had CD177 gene deletions among 294 subjects. Notably, all 11 HNA-2 deficient donors identified in the flow cytometry analysis carried two copies of CD177 gene. In addition, those 11 HNA-2 deficient donors produced full-length CD177 mRNAs as demonstrated by RT-PCR (Fig 1C). Our data clearly demonstrated that CD177 gene deletion (or CNVs) and the lack of mRNA expression are not the cause of HNA-2 deficiency.
We subsequently determined CD177 cDNA sequences of all 11 HNA-2 deficient donors along with 119 HNA-2 positive donors. In addition to CD177 coding SNPs (cSNPs) identified previously, we discovered five novel cSNPs (SNP 824G>C or rs17856827G>C, 828A>C or rs70950396A>C, 829A>T or rs70950396A>T, 832G>A, and 841A>G or rs201266439) (S1 Table), which form two haplotypes (Fig 2). Most importantly, the CD177 SNP 829A>T is a nonsense polymorphism that creates a translation stop codon at amino acid position 263 (Lysine → Stop codon change) in CD177 open reading frame. Consequently, those two haplotypes were designated as the open reading frame haplotype (or ORF allele: 824G/828A/829A/832G/841A) and the stop codon haplotype (or STP allele: 824C/828C/829T/832A/841G) (Fig 2). To determine the origin of the novel CD177 cSNP haplotype, we have also sequenced CD177 genomic DNA PCR products. Based on genomic DNA sequencing analysis, 72.1% (212/294) of donors were homozygous 829AA donors and the homozygous 829TT donors accounted for 3.1% (9/294) in our study population. The minor allele (829T) frequency is 15.5% (S2 Table). The distribution of SNP 829A>T genotypes was consistent with the Hardy-Weinberg equilibrium in 294 blood donors (χ2 = 0.76, P = 0.38) (S2 Table).
To examine whether the CD177 SNP 829A>T affects HNA-2 expression, the donor genotypes and HNA-2 expressions were statistically analyzed. As shown in Fig 3A, all nine 829TT homozygous donors were negative for HNA-2 expression in flow cytometry analysis. In addition, the percentages of HNA-2 positive neutrophils from 73 heterozygous donors (829AT) were significantly lower than those from 212 homozygous 829AA donors (P < 0.0001). Western blot analyses also confirmed the absence of HNA-2 protein in 829TT homozygous donors and significantly less HNA-2 protein being expressed in the 829AT donors when compared to the 829AA homozygous donors (Fig 3B). Our data strongly support the notion that the SNP 829A>T allele is a crucial determinant for HNA-2 deficiency and expression variations. To verify our findings, we recruited an independent cohort containing 102 blood donors, among whom nine HNA-2 deficient donors were identified (S1 Fig). Similar to those of the first cohort, all nine HNA-2 deficient donors in the replication cohort were SNP 829TT homozygotes as demonstrated by sequencing analysis of genomic DNA and cDNA (S2 Fig). Again, the SNP 829A>T genotypes were significantly associated with the percentages of HNA-2 positive neutrophils (S1 Fig) and the HNA-2 protein expression (S3 Fig). Our data confirmed that the SNP 829A>T is a crucial genetic determinant for HNA-2 deficiency and expression variations.
Similar to all nine homozygous 829TT donors, two 829AT heterozygous donors were also negative for HNA-2 expression (Fig 3A, empty diamonds in the middle column). Analysis of their CD177 cDNA sequences revealed that both HNA-2 deficient donors who were heterozygous for SNP 829A>T also had a heterozygous deletion of the guanidine nucleotide at nucleotide 997 (997G deletion). To determine haplotypes of the SNP 829A>T and the 997G deletion, we cloned and sequenced cDNA from those two HNA-2 deficient donors. As shown in Fig 4A, two species of CD177 mRNAs were found in those two donors. The SNP 829T (STP) allele is in the linkage disequilibrium with the wild-type CD177 997G allele while the 829A (ORF) allele carries the 997G deletion. Genomic DNA sequence analysis confirmed that the guanidine nucleotide deletion occurs at genomic level (Fig 4B). Our data indicate that the presence of the 829T allele in combination with the deletion mutation at nucleotide 997 on another chromosome could also lead to the HNA-2 expression deficiency in an individual. However, we found that only two out of 294 blood donors carried the 997G deletion mutation at one chromosome with genomic sequencing analysis. Therefore, the allele frequency of the 997G deletion mutation is estimated to be 0.0034 in the study population. In those two 829AT heterozygous donors, the 997G deletion allele was coincidentally paired with the 829T allele, which facilitated the discovery of the rare 997G deletion mutation in the study. We failed to identify any donors with the CD177 997G deletion among 102 additional blood donors of the replication cohort, confirming that the 997G deletion is a rare mutation in the population.
Although the genotypes of CD177 non-synonymous SNPs were reportedly associated with HNA-2 expression variations in several genetic analyses [18,31,32], it is unknown whether those CD177 cSPNs directly affect HNA-2 expression. To examine the effect of non-synonymous CD177 cSNPs on HNA-2 expression and on the binding to HNA-2 alloantibodies, we cloned the full-length CD177 cDNA variants containing common non-conservative cSNPs (SNP 134A>T, 652A>G, 656G>T, and 1084G>A) within the coding region for HNA-2 mature peptide (aa22-408). As shown in Fig 5, there were no significant differences in the expression of HNA-2 (Fig 5A) or in the binding to HNA-2 alloantibodies (Fig 5B) among four CD177 variants consisting of four non-conservative amino acid substitutions (His31Leu, Asn204Asp, Arg205Met, and Ala348Thr). Our data support the notion that non-synonymous CD177 cSNPs do not have a direct role in the HNA-2 alloantibody production and expression variations. However, cells transfected with CD177 variants of either STP haplotype (CD177-STP) or 997G deletion (CD177-997ΔG) failed to express HNA-2 on cell surface (Fig 5C) and had no reactivity with HNA-2 alloantibodies (Fig 5D). Our data confirmed that either STP allele or 997G deletion mutation will lead to the HNA-2 expression deficiency. To further confirm that the nonsense SNP 829A>T in the STP haplotype is the key factor for HNA-2 expression, we generated a CD177 expression construct carrying the sole change at SNP 829A>T position. The T substitution at nucleotide position 829 alone led to the absence of HNA-2 expression in transfection experiments (S4 Fig), confirming that the SNP 829A>T is the sole determinant for HNA-2/CD177 expression in the STP haplotype.
The phenomenon of HNA-2 deficiency was observed more than four decades ago [10], however, the underlying genetic mechanism of HNA-2 deficiency has remained unknown. In the current study, we identified five common CD177 cSNPs (SNP 824G>C, 828A>C, 829A>T, 832G>A, and 841A>G, minor allele frequency = 0.155) in complete linkage disequilibrium. Among five SNPs, the nonsense SNP 829A>T changes the amino acid codon #263 from lysine to a stop codon, which leads to the HNA-2 expression deficiency. Neutrophils from all 829T allele homozygous donors failed to express HNA-2. In addition, the percentages of HNA-2 positive neutrophils from the SNP 829A>T heterozygous donors (ORF/STP) were significantly lower than those from ORF homozygous donors. In vitro, the T substitution at the nucleotide position 829 alone led to HNA-2 expression deficiency in transfection experiments, confirming that the SNP 829A>T is the sole determinant for HNA-2 expression in the STP haplotype. Our study was the first to unravel the genetic mechanism for HNA-2 deficiency, which plays critical roles in human immunological diseases including TRALI, immune neutropenia, and bone marrow graft failure [3–6,10–13]. The delineation of the HNA-2 genetics undoubtedly will enable the development of effective genetic and clinical diagnosis tools in human medicine.
Intriguingly, similar to neutrophils from all homozygous donors of 829T allele, neutrophils from two 829AT heterozygous donors were also negative for HNA-2 expression (Fig 3A). Analysis of cDNA sequences of those two 829AT heterozygous donors deficient for HNA-2 revealed that the 829A allele (or ORF allele) in those two donors had a guanidine deletion at the nucleotide position 997, which leads to the CD177 reading-frame shift starting from the amino acid codon #319 (Fig 4) and the creation of a stop codon at the amino acid codon #342. The CD177 997G deletion also leads to the early termination of HNA-2 peptide translation, similar to the consequence of the 829T allele. Furthermore, the CD177 variant carrying the nucleotide 997G deletion failed to express HNA-2 on cell surface in the transfection experiments (Fig 5C), confirming the contribution of the 997G deletion mutation to HNA-2 deficiency in those two specific individuals. The CD177 nucleotide 997G deletion mutation was extremely rare (mutant allele frequency = 0.0034) and was absent in the replication cohort of 102 donors. The coincidental appearance of the 997G deletion allele and the 829T allele in the HNA-2 deficient donors facilitated the discovery of the rare 997G deletion mutation in the study. Therefore, at the presence of 829T allele, the rare CD177 997G deletion may also contribute to HNA-2 deficiency. However, the 997G deletion mutation with the allele frequency of 0.0034 will have much less impact on overall HNA-2 deficiency as compared to the SNP 829A>T (the 829T allele frequency was 0.155, S2 Table).
Previous genetic studies suggested that the CD177 non-synonymous SNPs might affect HNA-2 expression [18,31,32], however, the effect of those CD177 cSNPs on HNA-2 expression is unclear. In the current study, we carried out transfection experiments to examine whether common non-conservative cSNPs (SNP 134A>T, 652A>G, 656G>T, and 1084G>A) within the HNA-2 mature peptide (aa22-408) affect the HNA-2 expression and the binding of HNA-2 alloantibodies. We found that the expression of HNA-2 and the binding to HNA-2 alloantibodies were not significantly different among those natural CD177 variants containing non-conservative amino acid substitutions (His31Leu, Asn204Asp, Arg205Met, and Ala348Thr) (Fig 5A and 5B). The expression of HNA-2 in normal neutrophils is also affected by methylations of CD177 promoter and the CD177 SNP 42G>C (rs45441892) at the third codon (Pro3Ala) of the HNA-2 signal peptide was associated with methylation levels of CD177 promoter [39]. However, we found no association between the SNP 42G>C genotypes and the percentages of HNA-2 positive neutrophils in our study (ANOVA, P = 0.1209, S5 Fig). Taken together, those non-synonymous CD177 cSNPs do not seem to have a significant effect on HNA-2 deficiency and expression.
The CD177 mRNA splicing abnormality was previously suggested to be the cause of HNA-2 deficiency as alternatively spliced CD177 mRNA species were detected in two HNA-2 deficient donors [33]. However, no further evidence was provided to support the alternative splicing hypothesis of HNA-2 deficiency in the report [33]. Alternative mRNA splicing is a physiological process and is an essential mechanism to produce different products from a single human gene [40–42]. It seems unlikely that HNA-2 deficient subjects have an abnormal mRNA splicing machinery as HNA-2 deficient donors appear healthy [6]. We have detected full-length CD177 mRNAs in all 11 HNA-2 deficient donors in the main study (Fig 1C) and in all nine HNA-2 deficient donors from the replication study. The combination of the alternative spliced CD177 mRNA isoforms and the regular CD177 mRNA isoform occurred only in two out of nine SNP 829TT homozygous donors in our replication cohort (S2 Fig). Our data refute the notion that the alternative splicing is a major cause of HNA-2 deficiency.
Although gene deletions or copy number variations (CNVs) are the primary cause for HNA-1 (or FcγRIIIB) deficiency [35–38], we did not find any CD177 gene deletion in our blood donors. We found that all HNA-2 deficient donors expressed full-length CD177 mRNAs. We also found that the SNP 829T allele was in complete linkage disequilibrium with SNP 134A, 156G, 593G, 652A, 656G, 671C, 782C, 793C, 824C, 828C, 832A, 841G, 1084G, and 1333G. Our data clearly demonstrated that the gene deletion or the lack of mRNA expression is not responsible for HNA-2 deficiency, in contrast to the HNA-1 deficiency. Interestingly, we found that all heterozygous donors of the SNP 829A>T determined by genomic DNA analysis primarily produced the SNP 829A allele (or ORF allele) mRNA based on their cDNA sequences. The nonsense SNP 829T allele tracer peak barely above the background was typically considered as sequence noise in the cDNA sequence analysis for heterozygous donors. This observation suggests that the CD177 mRNAs containing the nonsense 829T allele are much less stable than the CD177 mRNAs containing the common 829A allele within the same donor. This may explain the observation of associations between expression variations and certain CD177 cSNPs and the inability to discover the SNP 829A>T using the cDNA sequencing strategy in previous studies [31–33].
After transcription, the CD177 mRNA of the nonsense 829T allele may be quickly degraded by the mechanism of nonsense-mediated mRNA decay [43], which will lead to the low abundance of CD177 829T allele mRNA and the dominance of CD177 829A allele mRNA in the heterozygous individual. The nonsense-mediated mRNA decay mechanism per se may contribute to the CD177 mRNA expression deficiency in humans with different diseases, which may explain that the partial HNA-2 peptide was undetectable from those HNA-2 deficient donors in a previous study [44] and in the current study using multiple anti N-terminus of HNA-2 mAbs and HNA-alloantibodies (S6 Fig). Therefore, HNA-2 alloantibodies likely target the whole mature peptide of CD177 in HNA-2 deficient subjects. In heterozygous donors for the SNP 829A>T, only the 829A allele is able to express HNA-2. CD177 promoter DNA methylation regulates HNA-2 expression under physiologic conditions [39]. Non-selective methylation on the 829A allele alone is sufficient to effectively abrogate the HNA-2 expression in a specific cell during granulopoiesis, which may explain that the percentages of HNA-2 positive granulocytes were significantly lower in the 829A>T heterozygous donors than those in the 829A (or ORF) allele homozygous donors (Figs 3 and S1). Therefore, our data strongly support the concept that the SNP 829A>T is also a primary genetic factor for HNA-2 expression variations in humans.
As an important biomarker, HNA-2 (CD177) is over-expressed in neutrophils from patients with myeloproliferative disorders including polycythemia vera, essential thromobocythemia, idiopathic myelofibrocythemia, and hypereosinophilic syndrome [6,14]. HNA-2 was an indicator of increased erythropoietic activity in thalassemia syndromes as HNA-2 expression was significantly elevated in β-thalassemia patients compared to healthy controls [45]. HNA-2 overexpression may also have a direct role in the pathogenesis of myeloproliferative disorders as HNA-2 enhances cell proliferation in vitro [46,47]. Not surprisingly, the low percentage of HNA-2 positive neutrophil is significantly associated with myelodysplastic syndrome and chronic myelogenous leukemia [48,49], suggesting that the reduced levels of membrane-bound HNA-2 may decrease the proliferation and differentiation potentials of myeloid cells. It is possible that the selection pressure to limit the spread of myeloproliferative disorders during evolution may be an important factor in maintaining the CD177 nonsense polymorphism in humans. Therefore, the CD177 SNP 829A>T may be an important genetic risk factor for various myeloproliferative disorders. Approximately 3% of Caucasians, 5% of African Americans, and 1–11% of Japanese are HNA-2 deficient [14]. In the current study, we found that between 3.7% (main cohort) and 8.8% (replication cohort) blood donors (>98% of them were Caucasians from the State of Minnesota) were HNA-2 deficient. Our data indicate that percentages of HNA-2 deficient humans may vary in different regions and be affected by sample sizes.
In summary, the elucidation of the molecular mechanism of HNA-2 deficiency and expression variations fills the critical knowledge gap in the genetics of HNA-2 antigen system. Our findings will enable the development of reliable genetic assays for HNA-2 system and will facilitate the diagnosis and prognosis of HNA-2-associated human disorders.
The human study was approved by the Institutional Review Board for Human Use at the University of Minnesota with Study #1301M26461. Memorial Blood Centers (737 Pelham Boulevard, St. Paul, Minnesota 55114) provided healthy donor blood samples without identifications for research purpose as a service and no consent form was provided per the Memorial Blood Centers policy.
Healthy American blood donors were recruited at the Memorial Blood Center in St. Paul, Minnesota. The age of healthy blood donors ranged from 19 to 84 years-old and >98% of donors in the study were self-declared Caucasians living in the State of Minnesota.
The expression of HNA-2 and the percentage of HNA-2 positive neutrophils were determined with flow cytometry analysis. Leukocytes stained with either FITC-conjugated anti-CD177 mAb (MEM-166, mIgG1, Thermo Scientific) or mIgG1-FITC isotype control were analyzed on a FACS Canto flow cytometer (BD Biosciences). The FlowJo software (Tree Star Inc.) was used to evaluate flow cytometry data. Characteristic light-scatter properties were used to identify neutrophils in flow cytometry. Using the same criteria as in the literature [31], donors had less than 5% of granulocytes positive for MEM-166 staining in flow cytometry analysis were called as HNA-2 deficient.
Peripheral blood leukocytes (2 × 107 cells) were lysed in PBS containing 1% NP-40 and 1× protease inhibitor cocktail (Roche, Indianapolis, ID) for 1 hr on ice. The total proteins (50 μg) from each donor were used for Western blotting analysis under non-reducing condition with mouse anti-CD177 mAbs and rabbit anti-actin mAb (LI-COR Biosciences, Lincoln, NE). IRDye 800CW-labeled goat anti-mouse and IRDye 600-labeled goat anti-rabbit antibodies were used for imaging analysis with the instrument software on an Odyssey Infrared Imager according to vendor’s instructions (LI-COR Biosciences).
Human genomic DNA was isolated from EDTA anti-coagulated peripheral blood using the Puregene DNA isolation kit (Gentra Systems, Minneapolis, MN) by following the vendor’s instruction. Total RNA was purified from peripheral blood leukocytes using TRIzol total RNA isolation reagent (Invitrogen, Carlsbad, CA).
The CNVs of CD177 gene were determined using TaqMan Copy Number Assay kit (the probe location of the assay ID Hs01327659_cn is shown in Fig 1A) (Applied Biosystems, Foster City, CA) and RNase P reference assay (Applied Biosystems, Part# 4403326). Duplex quantitative real-time PCR reactions were carried out on an Applied Biosystems 7500 Real-Time PCR System according to the manufacturer’s instructions. All samples were tested in duplicates, and fluorescence signals were normalized to ROX. TaqMan assay quantitative PCR amplification curves were analyzed using 7500 Software on a plate by plate basis, and the CN was assigned from the raw Cq values using CopyCaller software (version 2.0; Applied Biosystems).
Five μg of total RNA was used for cDNA synthesis with the SuperScript Preamplification System (Invitrogen). The 1411-bp cDNA fragment covering the entire CD177 coding region was amplified with RT-PCR using the sense primer (5’-CTGAAAAAGCAGAAAGAGATTACCAGCCACAG-3’) and anti-sense primer (5’-GTCCAAGGCCATTAGGTTATGAGGTCAGA-3’). The PCR reaction was performed with 2 μl of cDNA, 200 nM of each primer, 200 μM of dNTPs, 2.0 mM of MgSO4, and 1 U of Platinum Taq DNA polymerase High Fidelity (Invitrogen) in a 25 μl reaction volume. Platinum Taq High Fidelity DNA polymerase was used as it allows the amplification of complex cDNA or DNA templates with high accuracy and yield. The ABI Veriti 96-well Thermal Cycler was used for the PCR reaction starting with 94°C for 3 min, 35 cycles of denaturing at 94°C for 30 s, annealing at 56°C for 45 s, extension at 68°C for 1 min and 30 s with a final extension at 72°C for 7 min. All the PCR products, treated with ExoSAP-IT (Affymetrix, Santa Clara, CA), were assessed by direct Sanger sequencing on an ABI 3730xl DNA Analyzer with BigDye v3.1 Sequencing kit (Applied Biosystems). CD177 cDNA was also directly cloned into pCR2.1-TOPO vector (Invitrogen, Carlsbad, CA). Multiple clones containing CD177 cDNA were selected and subsequently sequenced to confirm CD177 SNPs. Two sense primers and two antisense primers were used to sequence the full-length CD177 cDNA coding region (sequencing primers are listed in S3 Table). The electropherogram data, aligned by the DNASTAR software (DNAStar, Madison, WI) were used for the identification of gene polymorphisms.
Since CD177 and its pseudogene contain a highly homologous region between exon 4 and 9 (Fig 1A) [16,17], we used the long-template PCR strategy to obtain the CD177-specific products for sequence analyses. Long-template PCR was carried out to amplify the CD177 genomic DNA containing all 9 exons using the sense primer (5’-CTGAAAAAGCAGAAAGAGATTACCAGCCACAG-3’) and antisense primer (5’-GTCCAAGGCCATTAGGTTATGAGGTCAGA-3’). The PCR reaction was performed with 200 ng DNA, 200 nM of each primer, 200 μM of dNTPs, 2.0 mM of MgSO4, and 2 U of Platinum Taq DNA polymerase High Fidelity (Invitrogen) in a 25 μl reaction volume. The ABI Veriti 96-well Thermal Cycler was used for the PCR reaction starting with 95°C for 3 min; 10 cycles of denaturing at 94°C for 30 s, annealing at 64°C for 30 s, extension at 68°C for 8 min and 30 s; 30 cycles of denaturing at 94°C for 30 s, annealing at 54°C for 30 s, extension at 68°C for 8 min and 30 s; with a final extension at 68°C for 5 min. The CD177 DNA fragment (8728 base pairs) was sequenced with a primer (5’-TCTTTGCCCCACACTAAACA-3’) on an ABI 3730xl DNA Analyzer with BigDye v3.1 Sequencing kit.
The human HNA-2 expression constructs were generated by cloning Hind III/Xba I-flanked RT-PCR products containing full-length CD177 coding region (nucleotide position 25 to 1419, GenBank accession number: NM_020406.2) into the eukaryotic expression vector pcDNA3 (Gibco BRL). The Hind III/Xba I-flanked CD177 cDNA was amplified from the synthesized cDNA of a blood donor using the upper primer 5’-CCCAAGCTTACCAGCCACAGACGGGTCATGAG-3’ and the lower primer 5’-TGCTCTAGAGAGGTCAGAGGGAGGTTGAGTGTG-3’. The changes at nucleotide position 134, 652, 656, 824, 828, 829, 832, 841, 997, and 1084 were generated using QuikChange Site-Directed mutagenesis kit (Stratagene, La Jolla, CA) and primer sets listed in the S3 Table.
The 293 cells (human embryonic kidney cell line) from ATCC (ATCC#CRL-1573, Manassas, VA) were maintained in the DMEM medium supplemented with 10% fetal calf serum and L-glutamine (2 mM) in 5% CO2. Transfection reactions were carried out in the 100 mm cell culture dishes with the plasmid DNA (20 μg) purified with OMEGA Plasmid Maxi Kit (Omega Bio-Tek, Norcross, GA) and 40 μl of Lipofectamine 2000 reagent (Invitrogen). Transfected cells were cultured in DMEM medium supplemented with 10% fetal calf serum for two days before harvesting the cells for HNA-2 expression or the selection of stable cell lines with the supplement of G418 (final concentration: 1 mg/ml). The polyclonal cells surviving the G418 selection were sorted with Stemcelll EasySep Cell Sorter for equivalent HNA-2 expression. The expression of HNA-2 on the transfected 293 cell lines was determined with FITC-conjugated anti-CD177 mAb as described previously. In addition, five defined HNA-2 alloantibodies from the American Red Cross Neutrophil Serology Laboratory were used to evaluate the binding of HNA-2 to the cell lines expressing CD177 variants.
The ANOVA and the nonparametric t-test (Mann-Whitney test) were used to determine whether HNA-2 positive cell population sizes and the HNA-2 deficiency are statistically associated with the nonsense CD177 cSNPs. The χ2 test was used to determine whether the observed genotype frequencies are consistent with Hardy-Weinberg equilibrium.
|
10.1371/journal.pgen.1005674 | The Fanconi Anemia Pathway Protects Genome Integrity from R-loops | Co-transcriptional RNA-DNA hybrids (R loops) cause genome instability. To prevent harmful R loop accumulation, cells have evolved specific eukaryotic factors, one being the BRCA2 double-strand break repair protein. As BRCA2 also protects stalled replication forks and is the FANCD1 member of the Fanconi Anemia (FA) pathway, we investigated the FA role in R loop-dependent genome instability. Using human and murine cells defective in FANCD2 or FANCA and primary bone marrow cells from FANCD2 deficient mice, we show that the FA pathway removes R loops, and that many DNA breaks accumulated in FA cells are R loop-dependent. Importantly, FANCD2 foci in untreated and MMC-treated cells are largely R loop dependent, suggesting that the FA functions at R loop-containing sites. We conclude that co-transcriptional R loops and R loop-mediated DNA damage greatly contribute to genome instability and that one major function of the FA pathway is to protect cells from R loops.
| R loops are co-transcriptional RNA-DNA hybrids that can have a physiological role in transcription and replication, but also may be a major threat to genome stability. To avoid the deleterious effects of R loops, specific factors prevent their formation or facilitate their removal. The double-strand break repair factor BRCA2 is among those that prevent R-loop accumulation. As BRCA2 also protects stalled replication forks and is the FANCD1 member of the Fanconi Anemia (FA) pathway, we studied the role of this pathway in preventing R loop accumulation and R loop-dependent genome instability. Using human and murine cells defective in FANCD2 or FANCA and primary bone marrow cells derived from FANCD2 deficient mice, we show that the FA pathway removes R loops and that many DNA breaks accumulated in FA cells are R loop-dependent. Importantly, FANCD2 foci accumulation is largely R loop-dependent, suggesting that the FA functions at R loop-containing sites. The FA pathway is primarily known as a DNA interstrand crosslinks (ICLs) repair pathway. Our findings reveal a novel function of the FA pathway in preventing R loop-mediated DNA damage, providing new clues to understand the relevance of R-loops as a natural source of genome instability and the way they are processed.
| Genome instability is a cell pathology in which chromosomes undergo alterations in the form of DNA breaks, mutations, rearrangements and loss at a high rate. In many cases, the mechanism responsible for genome instability implies a DNA replication failure. For this reason, genome instability and replication stress are two features tightly linked and are hallmarks of tumor cells [1, 2]. Chromosome duplication emerges thus like the most vulnerable process in the cell, so that events impairing progression of the replication fork (RF) have the potential of compromising genome integrity [3].
Apart from DNA damage generated by reactive oxygen species (ROS) and other natural genotoxic agents such as reactive aldehydes (RA), transcription is a major natural contributor to genome alterations. In the last decade evidence has accumulated that co-transcriptional R loops, structures formed by an RNA-DNA hybrid and a single strand DNA (ssDNA), may have an important role in the origin of genome instability [4–6]. From yeast to human cells, different factors play distinct roles in maintaining low levels of R loops along the genome. Importantly, mutations in such factors not only lead to accumulation of R loops above wild-type (WT) levels but also cause genome instability [7–13]. R loops, however, have been observed at different regions of the eukaryotic genome [14, 15] and have also regulatory roles in transcription [5].
Cells have two ways to limit R loops, those resolving them, such as RNase H or Senataxin, and those preventing their formation such as Topo I, the THO complex or the SRSF splicing factor, among other functions [4]. These functions serve to prevent genome instability by avoiding accumulation of R loops as a putative barrier to RF progression [16, 17], The observations that R loops trigger chromatin condensation and heterochromatin formation [5, 18, 19] suggest the possibility that chromatin compaction may be a major source of R-loop-mediated replication stress and genome instability [20], consistent with previous observations linking premature chromatin condensation and chromosome fragility [21]. Interestingly, factors like the yeast and human FACT chromatin reorganizing complex, which is crucial for RF progression through transcribed regions [22], and of the human BRCA1 and BRCA2 double-strand break repair (DSB) factors [23, 24] are also involved in R loop processing.
The fact that BRCA2/FANCD1 and BRCA1 directly or indirectly participate in the Fanconi Anemia (FA) pathway, involved in the repair of inter-strand crosslinks (ICLs) that block RF progression [25, 26] suggests that R loops may be an important contributor to genome instability in FA cells. To test this hypothesis we investigated the role of the FA pathway in resolving R loops and in protecting cells from R loop-mediated DNA breaks. Using human and murine cells defective in FANCD2 or FANCA and primary bone marrow cells derived from FANCD2 deficient mice, we validated our hypothesis. We propose that R loops accumulate in BRCA/FA- cells due to the incapacity of these cells to replicate R loop-containing regions.
To assay whether the FA pathway has a role in preventing or resolving R loops in human cells, we analyzed R loop accumulation in cells with dysfunctional FANCA or FANCD2 proteins or depleted of either of them (Fig 1). We performed DRIP-qPCR in four human genes, APOE, RPL13A, EGR1 and BTBD19 (S1 Fig) in well-established cell lines derived from Fanconi Anemia patients. We selected these four genes because they were identified as regions prone to form R loops and have been positively validated for the analysis of R loop accumulation [14, 22, 23].
We first determined the levels of R loops in wild-type lymphoblast cell line NV012 used as a reference control as well as in FANCA-/- lymphoblast patient cell line HSC72 and its corrected version [27]. Results clearly show that in all four genes tested, FANCA-/- cell lines accumulated R loops at a statistically significant higher level in APOE, RPL13A and EGR1; in BTBD19 the R loops were also higher but to a lower level (Fig 1A). Importantly, when the samples were treated with RNase H that digests the RNA moiety of RNA-DNA hybrids, the levels of R loops dramatically decreased, confirming that indeed the signal detected was specific for RNA-DNA hybrids. The absolute amount of R-loop signal as a function of input DNA is also provided (S2 Fig). In all analyses, the SNRPN gene was used as negative control because it does not accumulate R-loops, as previously reported [22, 23], and to normalize the values of the four genes analyzed (Fig 1).
Next we assayed whether this result could be extended to another cell line from an FA patient. We used the PD20 human fibroblast cell line from a FANCD2-/- patient [28]. Again, there was a clear increase of R loop accumulation in all four genes analyzed, this accumulation being statistically significant in APOE and RPL13A (Fig 1B). Importantly, the R loop signal was dramatically and significantly reduced when samples were treated with RNase H. Therefore, we can conclude that cell lines of different tissues from patients with two different dysfunctional FA genes accumulate R loops.
Finally, we tested whether this conclusion was also valid for HeLa cells depleted of FA proteins. We depleted cells of FANCD2 by siRNA (S3 Fig) and R loops accumulation was assayed in the same four human genes tested in patient cell lines. R loops clearly increased in siFANCD2 cell lines, up to 3 fold above the siC control levels (Fig 1C). The results confirm that a deficiency in the FA pathway, regardless of whether occurring in cells from human patients or in standard cell lines depleted of an FA factor by siRNA, leads to R loop accumulation. The similarity of results for the FANCD2-/- patient PD20 cell line and siFANCD2 depleted cells enabled us to use FANCD2-depleted HeLa cells as a reliable system to study the role of R loops in FA-deficient cells. In addition to demonstrate the presence of high levels of RNA-DNA hybrids as a consequence of FANCD2 knockdown at the molecular level by DRIP, we also confirmed this fact at the cellular level by immunofluorescence (IF) (Fig 2).
So far we have demonstrated that R loops accumulate in transformed human cells. Next we assayed R loop accumulation in murine embryonic fibroblasts (MEFs) obtained from mice defective in FANCD2. We performed DRIP-qPCR analyses in three different regions of the Acat3 gene (Acat3-1, Acat3-2 and Acat3-3) formerly annotated as AIRN locus (S4 Fig), which have been shown to be reliable for R loop detection in murine cells, as assayed by non-denaturing bisulfite treatment combined to RNase H digestion [15]. DRIP-qPCR in FANCD2-/- MEFs reveals a statistically significant increase of up to 3 fold in R loop accumulation compared to wild-type MEFs (Fig 3A). As expected, the signals decreased when MEFs were treated with RNase H, confirming that the signal detected was specific for R loops.
Next we addressed whether R loops physiologically accumulated in bone marrow cells from FANCD2-/- mice. We analyzed R loop accumulation in the Acat3-1 and Acat3-2 regions of myeloid Gr1+ and lymphoid B220+ committed cells from FA mice by DRIP-qPCR and observed that again R loops were clearly accumulated, at least 5 fold over the levels observed in WT mice (Fig 3B and 3C). As expected, the detected signal was clearly decreased by RNase H treatment.
Our results both at molecular and cellular levels indicate, therefore, that human cells deficient in the FA pathway accumulate R loops, regardless of the cell type analyzed, and the same occurs in bone marrow cells from FANCD2-deficient mice.
Once demonstrated that both human and murine cells defective in FA genes accumulate R loops, we investigated the functional impact of R loops in cells with a defective FA pathway. For this we used the FANCD2-depleted HeLa cells. First we wondered whether the accumulation of double strand breaks (DSBs) in siFANCD2 cells was related to R loop accumulation. We assayed DSBs indirectly by detection of γH2AX foci by immunofluorescence in cells transfected with a control or an RNase H1 overexpressing plasmid. Upon knockdown of FANCD2 a significant increase in γH2AX foci formation was observed by quantifying the number of cells with more than 10 foci (Fig 4A, left panel). Importantly, those high levels of γH2AX foci were strongly reduced by RNase H1 overexpression, confirming that γH2AX formation in FANCD2-depleted cells is R loop-dependent. Since the overexpressed RNase H1 protein localized both in cytoplasm (mitochondria) and nucleus, to prove that the observed effect was specific of the nuclear function of RNase H1 we performed the same experiment using the truncated version of the RNase H1 that lacks the mitochondrial localization signal and localizes only into the nucleus (S5 Fig) [29]. As expected, overexpression of the nuclear form of RNase H1 reduced the γH2AX foci accumulation caused by FANCD2 depletion.
As the hallmark phenotype of FA-deficient cells is sensitivity to inter-strand crosslinking (ICL) agents such as mitomycin C (MMC) we reasoned that MMC might not only generate DNA-DNA ICLs but also RNA-DNA ICLs that would eventually lead to DNA breaks. Therefore, we determined whether MMC-induced γH2AX foci in siFANCD2 HeLa cells were also mediated by R loops. MMC clearly increased γH2AX foci both in siC and siFANCD2 cells as seen by IF and quantification of cells with more than 10 foci, whose value increased 12 to 14 fold above those of the untreated cells (Fig 4A, right panel) reaching >75% in siC MMC-treated cells and >85% siFANCD2 MMC-treated cells. Importantly, however, a large fraction of these MMC-induced foci were reduced in cells transfected with the RNase H1-overexpressing plasmid, confirming that they were R loop-dependent. This conclusion is further supported by the observation that MMC-induced γH2AX foci were also significantly reduced by RNase H1 overexpression in MEFs FANCD2-/- (S6 Fig). Consistent with the partial R-loop dependency of γH2AX foci, DNA-RNA hybrids accumulate at higher levels in cells treated with MMC, as detected by a significant increase in S9.6 signal both in siC and siFANCD2 depleted cells (S7 Fig).
The statistically significant reduction of MMC-induced breaks in both siC and siFANCD2 cells by RNase H1 overexpression suggests that MMC may induce ICLs also at RNA-DNA hybrids as a cause of its DNA damage capacity, and that such RNA-DNA hybrids may be a relevant source of MMC-induced DNA breaks in siFANCD2 cells. Although we cannot formally discard the possibility that ICLs did not form at RNA-DNA hybrids, but instead hybrids could divert limiting-DNA repair factors from ICLs, so that RNase H1 would promote ICL repair indirectly by degrading R-loops and releasing such repair factors, there is no chemical basis to assume that ICLs cannot form between RNA and DNA strands.
If siFANCD2 cells accumulate DNA breaks dependent on co-transcriptional R loops, the breaks should be transcription-dependent. To assay this, we performed single cell electrophoresis or comet assay in cells incubated with cordycepin, a specific inhibitor of adenine incorporation into the nascent RNA. DNA breaks were clearly reduced in the presence of cordycepin in both siC and siFANCD2 cells (Fig 4B), suggesting that a large portion of DNA breaks in cells are mediated by transcription, as expected (reviewed in [30]). Importantly, a significant 5-fold increase in DNA breaks was observed in siFANCD2 cells compared to control cells, that was completely suppressed by cordycepin (Fig 4B). Therefore, DNA breaks accumulated in siFANCD2 cells are transcription-dependent, consistent with them being mediated by co-transcriptional R loops.
As the FA pathway repairs ICLs that impede normal RF progression [31] we reasoned that if ICLs are formed between the RNA and DNA strands, the FA core complex should accumulate at sites containing RNA-DNA hybrids. Therefore, we expected that removal of R loops by RNase H1 overexpression reduced the accumulation of FA foci at the sites of putative RF blockages. To test this possibility we performed IF with anti-FANCD2 antibody in HeLa cells transfected with the plasmid overexpressing RNase H1 as well as the empty plasmid in cells untreated and treated with MMC. As can be seen in Fig 5A, a significant increase of FANCD2 foci was observed after MMC treatment. Importantly, overexpression of RNase H1 drastically reduced FANCD2 foci both in MMC-treated and untreated cells. To exclude the possibility of an indirect effect of RNase H1 overexpression that could slow down proliferation and indirectly the activation of the FA pathway, we determined the effect of RNase H1 overexpression on cell cycle progression by measuring via BrdU incorporation and FACS analysis the percentage of cells in S phase (S8 Fig). We found that 24 hours after plasmid transfection there was no difference in the amount of cells in S phase in RNase H1 overexpressing cells with respect to control cells, which rules out a major impact of RNase H1 on cell proliferation to explain our results.
This result is consistent with the FA core complex locating at RFs blocked at R loop sites and supports that MMC causes RNA-DNA ICLs, indicating that the FA pathway plays a key function assisting the repair of RFs blocked at R loop-containing sites. To prove that the FA pathway acts at the sites where RNA-DNA hybrids are accumulated we performed ChIP of the FA core complex protein FANCA. Using anti-FANCA antibody we found that the FA core complex is indeed recruited to the genes that we had shown to accumulate RNA-DNA hybrids in FA deficient cells (Fig 5B). Finally, to demonstrate the functional link between the site of action of these FA complexes and DNA damage, we assayed whether γH2AX was enriched at genes that accumulate RNA-DNA hybrids in the absence of a functional FA pathway in an R loop-dependent manner. ChIP analyses with anti-γH2AX antibody confirmed that this was the case (Fig 5C). RNase H1-sensitive γH2AX signals were significantly higher in FANCD2-/- cells, confirming a physical link of R-loops with the regions at which DNA damage and the FA proteins are found.
We demonstrated using different cell lines defective in FANCD2 or FANCA from either human patients or HeLa and primary bone marrow murine cells, that FA cells accumulate R loops. Using siFANCD2 HeLa cells, we demonstrated that the increase in DNA breaks is strongly reduced by RNase H1 overexpression and transcription inhibition. The results indicate that the FA pathway plays an important role in protecting cells from naturally formed R loops and that R loops are a major source of DNA breaks in FA cells. This not only occurs in untreated cells, but also in cells treated with the ICL agent MMC (Figs 4 and 5). RNA-DNA hybrids seem to be a major source of RF blockage that requires the FA pathway for replication resumption and repair. Proper replication in FA+ cells would contribute to prevent R loop accumulation. Consistently, a high increase in R loops was observed in highly proliferative bone marrow tissues and MEFs from replication-impaired FANCD2-/- mice (Fig 3).
We have recently shown that BRCA2- and BRCA1- cells accumulate R loops and that an important fraction of the DNA breaks generated in these cells could be suppressed by RNase H1 overexpression [23]. A role of BRCA1 in R loop resolution is supported by its ability to recruit the RNA-DNA helicase SETX to DNA [24, 32]. As BRCA2 binds ssDNA [33] and protects RFs, avoiding their collapse [34, 35], a major role for BRCA2 in preventing R loop accumulation could be mediated by replication, without excluding additional putative roles of BRCA2 and BRCA1 in non-replicating cells. The FA pathway is involved in the repair of ICLs [36], and the fact that BRCA2/FANCD1 is a member of the FA pathway and BRCA1 has an FA-associated function opened the possibility that the FA pathway played a relevant role in removing R loops via the replication of R-loop-containing regions [23]. FANCA belongs to the FA core complex, FANCD2 being the main switcher that activates the pathway after monoubiquitination. Our work with FANCD2-/- and FANCA-/- cells therefore demonstrates the need of the FA pathway to remove R loops or R loop-associated DNA damage, including presumably RNA-DNA ICLs (Figs 1–4). This work also suggests that R loop accumulation might be a potential driver of bone marrow failure and haematopoietic stem cell attrition seen in FA mice deficient in aldehyde catabolism, but further work is needed to verify this hypothesis [37].
Our study suggests that in addition to ribonuclease H and RNA-DNA helicases, R loops might also be resolved during replication/repair. The observation that the FACT chromatin reorganizing complex is involved in RF progression preferentially when transcription is active and that FACT dysfunction leads to R loop accumulation in yeast and human cells, indicate indeed that R loops are a main source of genome instability in cells unable to properly replicate through R loop-containing regions [22]. The role of the FA pathway would be critical for progression of RFs stalled at either R loops or RNA-DNA ICLs. By repairing the R loop-dependent RF block, the R loop would be removed. This is consistent with the observations that FANCD2 foci formed in MMC-treated and untreated cells are strongly reduced by RNase H1 overexpression (Fig 5A), and that FANCA is recruited to R loop-forming genes in an RNA-DNA hybrid-dependent manner (Fig 5B). In addition, DNA damage accumulation in FANCD2-depleted cells specifically takes place at R-loop forming genes (Fig 5C), strengthening the hypothesis that the FA core complex assembles at sites where R loops block the progression of RFs and prevents R loop-dependent DNA damage, as proposed in our model (Fig 6).
R loops may thus constitute a major source of replication stress and genome instability. These are features commonly found in cancer cells and cells lacking a functional FA pathway that will not be able to resume replication through R-loop containing regions. This study, therefore, not only provides evidence that co-transcriptional R loops are major sources of replication stress, but also demonstrates that the FA pathway plays a crucial role in the repair of R loop-mediated damage or RF blockage (Fig 6). We propose that the action of FA during replication allows the removal of the R loop, whereas in FA cells, the block persist and therefore R loops are accumulated and DNA breaks arise. Knowing of the ability of R loops to trigger chromatin condensation [18] it would be certainly interesting to assay in the future the contribution of chromatin condensation to this phenomenon.
HeLa cells were cultured in Dulbecco’s modified Eagle’s medium (DMEM; GIBCO, Thermo Scientific, Waltham, MA) supplemented with 10% heat-inactivated fetal bovine serum at 37°C (5% CO2). Transient transfection of siRNA was performed using DharmaFECT1 (Dharmacon) according to the manufacturer’s instructions. Lipofectamine 2000 (Invitrogen, Carlsbad, CA) was used for plasmid transfection. All assays were performed 48 h after siRNA transfection plus 24 h after plasmid transfection as described previously [38]. The plasmids used were the following: pcDNA3, an empty vector; pcDNA3-RNaseH1, containing the full length RNase H1 cloned into pcDNA3 [39]; pEGFP, a vector expressing GFP, and pEGFP-M27, containing the GFP-fused RNase H1 lacking the first 26 amino acids responsible for its mitochondrial localization cloned into pEGFP [29].
Human NV012 (WT), HSC72 (FA-A) and HSC72+FANCA (FANCA-corrected) EBV-immortalised lymphoblastoid cell lines [27] were cultured in RPMI 1640 (GIBCO) supplemented with 15% heat-inactivated fetal calf serum under standard culturing conditions. Human transformed fibroblasts PD20 (FANCD2-/-) and PD20 corrected (PD20 retrovirally corrected with pMMp-FANCD2 cDNA) [28] were grown in DMEM (GIBCO) supplemented with 15% heat-inactivated fetal calf serum as previously described [40, 41].
Mitomycin C (MMC, M4287, Sigma-Aldrich) was used to a final concentration of 80 ng/ml for 16 h for detection of γH2AX foci, 250 ng/ml for 5 h for S9.6 inmunofluorescence and 40 ng/ml for 16 h and then released for 9 h for FANCD2 foci.
Fancd2-/- mice (Fancd2tm1Hou, MGI code: 2673422, backcrossed into C57BL/6 background for at least 11 generations), obtained from K.J. Patel [42, 43], were maintained in a conventional mouse facility. All animal experiments undertaken in this study were performed under the approval of the EU Directive 2010/63EU, Spanish law RD53/2013 and the Hospital Virgen del Rocio Ethical Review Committee. Timed matings between Fancd2+/− males and females were set up. Females were checked for the presence of a vaginal plug the following morning, and considered to be at day E0.5 of pregnancy. Pregnant females were sacrificed at E13.5, uteruses removed and embryos dissected. Murine embryonic fibroblasts (MEFs) cultures were obtained, genotyped and transformed using a lentiviral vector pLOX-Ttag-iresTK (addgene 12246). Clones were isolated and expanded.
Murine bone marrow cells from femora and tibias were obtained by flushing in 2 mls of PBS+3% fetal bovine serum. Cells were enumerated using trypan blue 0.2% in a TC20 Automated Cell Counter (Bio-Rad). Biotinylated B220 (cloneRA3-6B2), Gr1 (Clone RB6-8C5) were obtained from BDBioscience. Cells were enriched using Streptavidin-bound magnetic particles (BD IMag) according to manufacturer instructions.
For the analysis of DNA damage foci, immunofluorescence was performed as described previously [38]. FANCD2 IF was performed as described previously [44] with minor modifications. In brief, cells were pre-permeabilized with 0.25% Triton X-100 in PBS for 1 minute on ice and then fixed with 2% formaldehyde in PBS. After blocking with 3% BSA in PBS, cells were incubated with the anti-FANCD2 (1:100 dilution) and the anti-RNASEH1 (1:400 dilution) followed by the secondary antibody conjugated with Alexa 488 and Alexa 546. DNA was stained with DAPI. In pre-permeabilized cells the overexpressed RNase H1 stained only nucleus and nucleoli because the rest of the protein had been washed out. S9.6 (hybridoma cell line HB-8730) immunofluorescence was performed as previously described [45] using secondary antibodies conjugated with Alexa 488 and Alexa 647. Images of IF and single-cell electrophoresis were acquired with a Leica DM6000 microscope equipped with a DFC390 camera (Leica). Images of S9.6 immunofluorescence were acquired with a Leica TCS SP5 confocal microscope. Data acquisition was performed with LAS AF (Leica). Images were captured at ×63 (IF) and ×10 (comet assay) magnification. Metamorph v7.5.1.0 software (Molecular Probes) image analysis software was used to quantify foci and nuclear S9.6 signal intensity.
Comet assay was performed as described [22] using a commercial kit (Trevigen, Gaithersburg, MD, USA) following the manufacturer’s protocol. Means and SEM (Standard Error of the Mean) from three independent experiments were obtained and are shown in each case. Comet tail moments were analyzed using Comet-score software (version 1.5).
Anti-γH2AX (clone JBW301; Upstate), Anti-RNASEH1 (15606-1-AP; Proteintech), Anti-FANCD2 (sc-20022; Santa Cruz, Dallas, TX), anti-β-Actin (ab8226, Abcam, Cambridge, UK), anti-Vinculin (V9264; Sigma-Aldrich) and anti-nucleolin (ab50279 Abcam) antibodies were used.
DRIP assays were performed as described [22], with the exception of the DRIP conducted in MEFs, in which double amount of RNase H was used. RNA-DNA hybrids were immunoprecipitated using the S9.6 antibody from gently extracted and enzymatically digested DNA, treated or not with RNase H [15]. Quantitative PCR was performed at the indicated regions (S1 and S4 Figs). The relative abundance of RNA-DNA hybrid immunoprecipitated in each region was normalized to the signal at the negative control region SNRPN gene in human cell lines. All experiments were performed in triplicate; average and SEM of results are provided.
HeLa cells were transfected with the corresponding siRNAs and 48h after siRNA transfection, they were transfected with either the RNase H1-coding plasmid pEGFP-M27 or the control plasmid pEGFP-C1. After 72 h of siRNA transfection, cells were crosslinked and processed for ChIP using standard procedures with minor modifications as previously described [23]. Anti-FANCA (Bethyl Laboratories) or anti-γH2AX (clone JBW301; Upstate) previously conjugated with Dynabeads Protein A (Life Technologies) were used to immunoprecipitate chromatin.
Cells were pulse-labeled with BrdU 10 μM added directly to the growing medium for 20 min, harvested, fixed with 70% ethanol in PBS and incubated on ice for 1 h. Cells were then treated with 2 N HCl 0.5% Triton X-100 for 30 min at room temperature, then with 0.1 M Sodium tetraborate pH 8.5, washed once with washing buffer (1% BSA 0.1% Triton X-100 in PBS), and incubated for 1 h in the same buffer containing 1:25 anti-BrdU antibody conjugated with Alexa Fluor 488 (B35139, Invitrogen) and 0.5 μg/μl RNase A. After one wash with washing buffer, cells were resuspended in PBS containing 100 ng/ml propidium iodide to counterstain DNA for 30 min and examined by flow cytometry (FACSCalibur; BD).
|
10.1371/journal.pgen.1003165 | MCM8 Is Required for a Pathway of Meiotic Double-Strand Break Repair Independent of DMC1 in Arabidopsis thaliana | Mini-chromosome maintenance (MCM) 2–9 proteins are related helicases. The first six, MCM2–7, are essential for DNA replication in all eukaryotes. In contrast, MCM8 is not always conserved in eukaryotes but is present in Arabidopsis thaliana. MCM8 is required for 95% of meiotic crossovers (COs) in Drosophila and is essential for meiosis completion in mouse, prompting us to study this gene in Arabidopsis meiosis. Three allelic Atmcm8 mutants showed a limited level of chromosome fragmentation at meiosis. This defect was dependent on programmed meiotic double-strand break (DSB) formation, revealing a role for AtMCM8 in meiotic DSB repair. In contrast, CO formation was not affected, as shown both genetically and cytologically. The Atmcm8 DSB repair defect was greatly amplified in the absence of the DMC1 recombinase or in mutants affected in DMC1 dynamics (sds, asy1). The Atmcm8 fragmentation defect was also amplified in plants heterozygous for a mutation in either recombinase, DMC1 or RAD51. Finally, in the context of absence of homologous chromosomes (i.e. haploid), mutation of AtMCM8 also provoked a low level of chromosome fragmentation. This fragmentation was amplified by the absence of DMC1 showing that both MCM8 and DMC1 can promote repair on the sister chromatid in Arabidopsis haploids. Altogether, this establishes a role for AtMCM8 in meiotic DSB repair, in parallel to DMC1. We propose that MCM8 is involved with RAD51 in a backup pathway that repairs meiotic DSB without giving CO when the major pathway, which relies on DMC1, fails.
| Species that reproduce sexually have two copies of each chromosome, inherited from their father and mother. During a special cell division called meiosis, these two chromosomes are mixed by homologous recombination to give genetically unique chromosomes that will be transmitted to the next generation. This recombination process is initiated by DNA breaks that must be repaired efficiently to maintain fertility. Using the model plant Arabidopsis thaliana we revealed here that the gene AtMCM8 is required to repair a subset of these DNA breaks. However MCM8 appears to not be required for recombination with the homologous chromosome. Instead MCM8 appears to be involved in a safety system that operates to repair DNA breaks that have not been used for homologous recombination. Interestingly the equivalent gene also has an essential meiotic function in the fly and the mouse. However the three species require MCM8 for different aspects of meiosis.
| Meiosis is a process that occurs in the germlines of sexually reproducing organisms. Two successive rounds of chromosome segregation (meiosis I and II) follow a single round of DNA replication (S phase). The resulting four cells each contain half the genetic content of the pre-meiotic mother cell. The genetic complement of these gametes is a mosaic of the paternal and maternal DNA due to meiotic recombination that occurs between S phase and the first meiotic division [1].
Meiotic recombination begins with programmed DSBs that are dependent on SPO11 and multiple cofactors, including PRD1 in plants [2], [3]. DSBs are subsequently resected to yield 3′ overhangs that invade the homologous chromosome. At this step, two recombinases co-operate to achieve efficient strand exchange with the homolog, RAD51 and DMC1 [4]. RAD51 is a recombinase involved both at mitosis and meiosis while DMC1 is specific to meiosis. Importantly, it has been recently shown in S. cerevisiae that only the strand exchange activity of DMC1, and not of RAD51, is required for meiotic crossover formation [5]. RAD51 appears thus to be an accessory factor of DMC1 for meiotic homologous crossover formation, but may also serve as a backup to repair breaks when DMC1 fails [5]. In Arabidopsis thaliana, RAD51 is indispensable for repair of meiotic DSBs as shown by the extensive meiotic chromosome fragmentation which occurs at meiosis in Atrad51 mutants [6], [7]. AtDMC1 is required for CO formation but not meiotic DSB repair. Indeed, in Atdmc1 mutant, meiotic DSBs are repaired in a AtRAD51-dependent manner which does not promote chromosome pairing and does not yield COs between homologs, likely using the sister chromatid as a template [7], [8]. In addition, consistent with a role of RAD51 in helping DMC1 in wild type, the number of DMC1 foci is severely decreased in a Atrad51 mutant [7], [9], while RAD51 foci are unaffected in Atdmc1 [9]. Thus two meiotic functions of RAD51 emerge, helping DMC1 to promote COs and promoting DSB repair on the sister without DMC1.
Two other Arabidopsis mutants, sds and asy1, have phenotypes reminiscent of Atdmc1, repairing breaks using AtRAD51 but exhibiting major homologous chromosome pairing defects and making no or few COs [10]–[12]. Both sds and asy1 show localization defects of AtDMC1 but not of AtRAD51, suggesting that they work with DMC1 to promote interhomolog recombination [12], [13]. Based on its amino acid sequence, SDS is a cyclin-like protein and ASY1 is a HORMA domain protein making it the likely functional homologue of S. cerevisiae Hop1.
DSB repair events form intermediates that are resolved as either crossovers (COs) or non-crossovers (NCOs) (gene conversion). COs are required for accurate segregation of chromosomes during meiosis I and can arise from at least two independent pathways known as class I and class II COs. These two pathways coexist in budding yeast, mammals and Arabidopsis [1], [14]–[17]. Class I COs are subject to a phenomenon known as interference, whereby the occurrence of a CO significantly reduces the probability of a CO occurring at an adjacent locus, in a distance dependent manner. This pathway is dependent on the ZMM proteins (defined as ZIP1, ZIP2/SHOC1, ZIP3, ZIP4, MSH4, MSH5, MER3) and, in most eukaryotes, is responsible for the majority of COs during meiosis. Class II COs, that do not display interference, require MUS81 [1], [14]–[17].
Here we addressed the meiotic function of MCM8. MCM8 is a member of the eight MCM family proteins (MCM2–9), that all share a well conserved helicase domain. Together MCM2–7, as a hexamer, form a well characterized DNA helicase, which is essential for replication in all eukaryotes [18]. In contrast, MCM8–9 is not present in all eukaryotes [19], being notably missing in S. cerevisiae, S. pombe and C. elegans, but existing in vertebrates and plants. A study in Xenopus showed that MCM8 functions during DNA replication at the elongation stage but it is not required for replication licensing. The Xenopus MCM8 protein is the only MCM8 representative for which helicase activity has been demonstrated in vitro [20]. MCM8 is also involved in, but not essential for the assembly of the pre-replicative complex in human [21]. Very recently, MCM8 and MCM9 has been shown to be involved in homologous recombination-mediated DNA repair in mouse and chicken somatic cells [22], [23]. MCM8 has also been shown to be involved in meiosis. In the fruit fly (Drosophila melanogaster), in which MCM9 has not been identified, MCM8 (also known as REC) is required for 95% of meiotic COs. In contrast to COs, the frequency of NCOs increases in the absence of Dmrec [24]. Finally, a very recent study pointed out a role for MCM8, but not MCM9, in meiotic recombination in mouse [22]. Indeed meiocytes in the mouse mcm8 mutant accumulate DMC1 foci, display synapsis defects and go into apoptosis, consistent with a defect in meiotic DSB repair. The meiotic function of MCM8 has been analyzed only in Drosophila and mouse, with contrasting conclusions. This raises the question of the conservation of this function in eukaryotes. The aim of the present study was to further explore the meiotic function of MCM8 by deciphering its role in the model plant Arabidopsis.
Phylogenetic analyses of the MCM family [19], [24], showed that the Arabidopsis genome contains one clear homolog for each MCM2–9, At3g09660 being the MCM8 homolog. We sequenced the At3g09660 CDS using RT-PCR on mRNA from Arabidopsis inflorescences. Because of some differences in splicing sites, the At3g09660 CDS slightly differed from the predicted sequence found in the genebank (NM_111800), measured 2,406 bp and contained 17 exons (Figure 1) (genebank BankIt1577803 MCM8 KC109786). We nonetheless confirmed by reciprocal BLAST analysis and multiple protein alignment that At3g09660 encodes the Arabidopsis MCM8 homolog (Figure S1 and [24]).
We identified three T-DNA insertions from the public collections within the AtMCM8 gene: Atmcm8-1, Atmcm8-2 and Atmcm8-3 (Figure 1). Plants homozygous for the insertions showed normal vegetative growth but reduced fertility as shown by Alexander staining of pollen (Figure 2). This phenotype (and others described below) was detected only in homozygotes of each mutant. Moreover seed counts showed that Atmcm8-1 has significantly less seeds than wild type (44.8±5.2 (n = 41) compared to 52.4±5.8 (n = 77), Z test p<10−13). Allelism tests showed that the meiotic defects observed (see below) were due to the insertions in Atmcm8.
To investigate if this reduction in fertility was linked to a meiotic defect, we analyzed meiotic progression by DAPI (4′,6-diamidino-2-phenylindole) staining of meiotic chromosome spreads in all three mutant alleles. In wild type meiosis (Figure 3A–3E), chromosomes condense at leptotene. Then, synapsis is initiated at zygotene until its completion in pachytene when the two homologous chromosomes are connected along their entire length by a proteinous structure called the synaptonemal complex [25] (Figure 3A and Figure 4A). Desynapsis occurs at diplotene and further condensation of the chromosomes occurs. Five bivalents continue to condense and become visible at diakinesis. At metaphase I, the five bivalents align on the metaphase I plate (Figure 3B). At anaphase I homologous chromosomes segregate to opposite poles (Figure 3C). At telophase I the two groups of five recombinant chromosomes begin to decondense. At prometaphase II chromosomes recondense and align on the two metaphase II plates (Figure 3D). At anaphase II each of the ten chromosomes segregate their two sister chromatids to opposite poles resulting in four balanced groups of five chromatids (Figure 3E).
In all three Atmcm8 alleles, meiosis appeared to progress normally from leptotene through to pachytene (Figure 3F) where chromosomes condensed, aligned and fully synapsed like wild type. The completion of synapsis in Atmcm8 was confirmed by immunolabelling meiotic chromosomes with antibodies against ASY1 and AtZYP1 (Figure 4B), that are components of the axial elements and of the transverse filament of the synaptonemal complex, respectively [26], [27]. Chromosomes desynapsed normally during diplotene and we observed five bivalents as condensation progressed during diakinesis, revealing the presence of chiasmata (the cytological manifestation of CO). At metaphase I, five bivalents were systematically observed in all mutant alleles, showing that at least one CO is formed per pair of homologous chromosomes (Figure 3G). Anaphase I proceeded, however chromosome fragmentation was observed in all three Atmcm8 alleles (Figure 3H–3K), with 1 to 10 chromosome fragments detected in 60 to 80% of the cells (Figure 5). Chromosomes aligned on the metaphase II plate, with fragments dispersed throughout the cell (Figure 3L). Anaphase II proceeded but additional chromosome fragments appeared (Figure 3M–3O). This fragmentation persists at telophase II. We also observed fragmentation in female meiosis showing that Atmcm8 mutation also affects female meiosis (data not shown).
In Atspo11-2 and Atprd1, no meiotic DSBs are formed and therefore recombination does not occur [3], [28]. Thus at metaphase I, ten univalents are observed and segregate randomly (Figure 6A–6B and 6E–6F). To test whether the chromosome fragmentation seen in Atmcm8 mutants are dependent on DSB formation or not, we introduced the Atspo11-2 and Atprd1 mutations independently into Atmcm8. At meiosis, we observed ten univalents at metaphase I in the Atmcm8/Atspo11-2 or Atmcm8/Atprd1 and, importantly, the chromosome fragmentation was abolished (Figure 6C–6D and 6G–6H, Figure 5). Therefore, the fragmentation defect of Atmcm8 is dependent on AtSPO11-2 and AtPRD1. Thus, AtMCM8 is required for efficient repair of the DSBs that initiate meiotic recombination.
We then tested if the Atmcm8 fragmentation phenotype is dependent on the presence of any of the known pathways of CO formation, using epistasis tests. We used Atmsh4 and Atzip4 that are both required for class I CO formation and Atmus81 that is required for class II CO formation. In the Atmcm8/Atmsh4, Atmcm8/Atzip4, Atmcm8/Atmus81 double mutants and the Atmcm8/Atmsh4/Atmus81 triple mutant, we still observed a chromosome fragmentation defect as in the Atmcm8 single mutant (Figure 5 and Figure 7, data not shown for Atmcm8/Atzip4). Thus the Atmcm8 fragmentation phenotype is independent of MSH4, ZIP4 and MUS81.
In Atmcm8 and Atmcm8/Atmus81 we invariably observed five bivalents at metaphase I, suggesting that the formation of class I COs, which account for most of the CO in wild type, is not grossly affected by the Atmcm8 mutation. This was further supported by counts of AtMLH1 foci, a marker of class I COs at late prophase of meiosis I [29], [30] (Figure S2), that revealed no significant differences between wild type (10.1±1.4 per cell; n = 81) and the Atmcm8 mutant (10.3±1.9; n = 86 (Z p = 0.55)). In Atmcm8/Atmsh4 (Figure 6), the residual number of bivalents at metaphase I was unchanged compared to the single Atmsh4 mutant (1.5±1; n = 91 vs 1.3±1.1; n = 91 (Z p = 0.94)), strongly suggesting that class II CO formation is not affected neither by Atmcm8 mutation. We then measured recombination frequency and crossover interference genetically in Atmcm8. This was achieved using tetrad analysis (Fluorescent-Tagged Lines, FTL) which is a visual pollen assay allowing the measurement of multiple COs simultaneously with access to all four chromatids from the same meiosis [31]. Two different sets of adjacent intervals on chromosome 5 have been analyzed, (I5aI5b and I5cI5d), representing four intervals in total. We did not detect any difference in recombination frequency between the Atmcm8 and wild type for any of these intervals (Table 1, Genetic Distance), consistent with the cytological data. Also, interference, that affects the distribution of crossovers, was unchanged compared to wild type for both sets of adjacent intervals (Table 1, Interference Ratio). Taken together these data suggest that AtMCM8 is not involved in CO formation. This contrasts from the observation that the absence of MCM8 reduces COs frequency by 95% in Drosophila [24].
Mei9/Rad1 is another gene required for the formation of more than 90% of the COs in Drosophila [32]. Given the major difference in MCM8 function between Arabidopsis and Drosophila, we tested the role of AtRAD1 [33]–[35] in crossover formation in Arabidopsis. Cytological analysis showed that the single Atrad1 mutant has no obvious defect in CO formation. We then analyzed if AtRAD1 has a minor effect. To achieve this, we constructed a shoc1/Atrad1 double mutant and a Atmus81/shoc1/Atrad1 triple mutant to be able to detect a weak reduction in CO formation, in a sensitive context where there are no class I and class II COs. However, this triple mutant was not different from Atmus81/shoc1 (0.99±0.84 (n = 74) versus 1.15±1.28 (n = 75), Z p = 0.36) and neither was shoc1/Atrad1 different from shoc1 (1.47±1.07 (n = 51) versus 1.56±0.86 (n = 32), Z p = 0.67). These genes, MCM8 and MEI9/RAD1, are essential for CO formation in Drosophila but not in Arabidopsis showing divergent functions. However, contrary to RAD1, MCM8 has conserved a meiotic function in Arabidopsis.
DMC1 is involved at the strand invasion stage of meiotic recombination and Atdmc1 mutants fail to synapse and to make COs (Figure 8A–8B, 8G–8H). However, DSBs are repaired in Atdmc1, in an AtRAD51-dependent manner, without CO formation, suggesting that the DSBs are repaired on sister chromatids in these mutants [8], [12]. In the Atmcm8/Atdmc1 double mutant, from metaphase I to the end of the meiosis we observed extensive chromosome fragmentation in all cells, which was much more intense than in the single Atmcm8 mutant (compare Figure 8C–8D to Figure 3I–3K and see quantification in Figure 5). Consistently, the Atmcm8/Atdmc1 double mutant was completely sterile whereas Atmcm8 has moderate fertility reduction and Atdmc1 produce some residual seeds [8], [12] (Table 2). Mutating SPO11-2 in this Atmcm8/Atdmc1 double mutant abolished the chromosome fragmentation (Figure 8E–8F, Figure 5), demonstrating that MCM8 and DMC1 act in parallel pathways of meiotic DSB repair.
Furthermore in the Atmcm8 mutant context, we observed a more drastic meiotic chromosome fragmentation in plants heterozygous for DMC1 (Atmcm8−/−AtDMC1+/−) than wild type for DMC1 (Atmcm8−/−AtDMC1+/+) (compare Figure 8I–8J to Figure 3I–3K, quantification on Figure 5), accompanied by a strong reduction of fertility (Table 2). However, the fragmentation observed in Atmcm8−/−AtDMC1+/− was less dramatic than in the double mutant (Atmcm8−/−Atdmc1−/−) (Figure 5), which is also supported by the fertility levels (Table 2). This is despite the AtDMC1 mutation being recessive (in an AtMCM8+/+ or AtMCM8+/− context). Thus, in the absence of Atmcm8, the mutation of one of the two copies of DMC1 was enough to enhance fragmentation, which is even more drastic when both DMC1 alleles are disrupted.
Therefore we tested the relationship of AtMCM8 with ASY1 and SDS, two proteins that are required for normal DMC1 localization [3], [13]. In the sds and asy1 single mutants, COs are greatly reduced (Figure 9E–9F) [10], [11]. In the Atmcm8/asy1 and Atmcm8/sds double mutant, we observed chromosome fragmentation from anaphase I onwards, which was much greater than that seen in the Atmcm8−/− single mutant (compare Figure 9G–9H with Figure 3I–3K, quantification on Figure 5). Thus, mutation of SDS or ASY1 amplified the fragmentation phenotype of Atmcm8. Finally, both the single Atrad51 mutant and the double Atmcm8/Atrad51 mutant show intense chromosome fragmentation (Figure 10). Interestingly, while AtRAD51+/− does not show chromosome fragmentation, Atmcm8−/−/AtRAD51+/− showed more chromosome fragmentation that Atmcm8 (Figure 10, Figure 5). Thus, in the absence of Atmcm8, the mutation of one of the two copies of RAD51 was enough to enhance fragmentation.
Given the relationship between DMC1 functional gene copy number and the degree of Atmcm8-dependent fragmentation, we looked at DMC1 behavior in Atmcm8. No significant difference in DMC1 foci shape or number was observed in Atmcm8−/− compared to wild type (Table 2). Similarly, we did not detect any differences in number or shape of DMC1 foci in Atmcm8−/−AtDMC1+/− or Atmcm8−/−AtRAD51+/− compared to either wild type or Atmcm8−/− (Figure S3, Table 2). In the Atmcm8 Atrad51 double mutant, we observed a marked decrease of DMC1 foci number, which was however similar to what was previously observed in a single Atrad51 mutant [7] (Table 2). It is intriguing that Atmcm8−/−AtDMC1+/− and Atmcm8−/−AtRAD51+/− exhibit a more drastic meiotic defect than Atmcm8−/−AtDMC1+/+, while DMC1 foci number and shape appear similar. However, it is possible that immunolocalization fails to detect subtle differences in DMC1 protein quantity or dynamics.
Next we explored the functional relationship between MCM8 and DMC1, in haploid plants, where homologous chromosomes are not present. Thus, the only template available for meiotic DSB repair is the sister chromatid. Meiotic chromosome spreads, in a wild-type haploid, showed that the five chromosomes were intact and segregated randomly at anaphase I [36] (Figure 11A–11B), suggesting that DSBs are efficiently repaired. The haploid Atmcm8 mutant had a limited fragmentation defect (Figure 11C–11D), similar to the defect in the diploid Atmcm8 mutant (Figure 5 for quantification). The Atdmc1 haploid had no fragmentation (Figure 11C–11F). In clear contrast, in the double Atmcm8/Atdmc1 haploid, we observed extensive meiotic chromosome fragmentation (Figure 11G–11H, see Figure 5 for quantification). This shows that in a haploid context, DSB repair is efficient in wild type and Atdmc1, only slightly affected in Atmcm8, but ineffective in the Atmcm8/Atdmc1 double mutant. This suggests that in the absence of a homologous template, AtMCM8 and AtDMC1 catalyze DSB repair on the sister chromatid in a redundant manner.
Here AtMCM8 was shown to be involved in meiotic DSB repair but not CO formation. This study thus revealed a pathway for DNA DSB repair that does not yield COs. This pathway depends on AtMCM8 and acts in parallel to the AtDMC1 pathway from which COs originate.
Arabidopsis MCM8 is required for effective meiotic DSB repair as all Atmcm8 mutant alleles had a clear, albeit limited, chromosome fragmentation defect at meiosis. The fragmentation is dependent on meiotic DSB formation as it disappears when AtSPO11-2 or AtPRD1 is absent. However, in contrast to Drosophila rec (mcm8) mutants, genetic and cytological data strongly support that CO formation is not affected by AtMCM8 mutation: (1) In the absence of AtMSH4 or AtZIP4 (class I COs) or AtMUS81 (class II COs) fragmentation still occurred and the number of bivalents was unchanged. (2) MLH1 foci numbers, a marker of class I COs, were unchanged in Atmcm8. (3) The genetic analysis using FTLs revealed no difference in terms of genetic distance and the strength of interference. These data showed that AtMCM8 acts in a pathway which repairs a subset of meiotic DSB and does not lead to CO formation.
A striking finding was that AtMCM8 becomes crucial when the DMC1 pathway was affected. Indeed, we observed a drastic amplification of the Atmcm8 mutant chromosome fragmentation defect when one of the two allelic copies of DMC1 was mutated, which was even more drastic when both DMC1 copies were mutated. This extensive fragmentation defect reflects a failure of DSB repair, as it is abolished by SPO11-2 mutation. Further, this extensive fragmentation was consistently confirmed in the absence of AtMCM8 and SDS, or AtMCM8 and ASY1. SDS and ASY1 are essential for AtDMC1 loading/stability [12], [37]. Extensive fragmentation was also observed when one copy of RAD51 was mutated in the Atmcm8 mutant. A function of RAD51 as a cofactor of DMC1 has been recently identified in yeast [5], and consistently DMC1 foci number is drastically reduced in the Arabidopsis rad51 mutant [7], [9]. We thus propose that two pathways of DSB repair coexist, one dependent on AtMCM8 and the other one on AtDMC1. In the absence of AtDMC1, efficient DSB repair occurs without CO formation. This repair depends on AtRAD51 [7], [8], [12] and on AtMCM8 (this study). Such RAD51-mediated, DMC1-independent, repair also exists in S. cerevisiae but is normally inhibited by RAD51 regulators [38]–[42]. Consequently, we suggest that, in the Atdmc1 context, AtMCM8 and AtRAD51 can co-operate to repair DSBs using the sister as a template. In addition to this function, AtRAD51 is required for the AtDMC1-dependent pathway (possibly as an accessory factor for the DMC1 strand-exchange activity as shown in yeast [5]) as repair is completely defective in the single Atrad51 mutant [6], like in the double Atmcm8/Atdmc1 mutant.
The fact that the fragmentation defect is limited in the single Atmcm8 mutant, suggests that the AtMCM8/AtRAD51 pathway would be essential for a limited number of events in wild type, when DMC1 fails. The repair events promoted by AtMCM8 are likely not intended to become a CO, as CO formation was not affected in Atmcm8, leaving sister chromatid repair or NCOs as the only other known possibilities. The absence of synapsis in Atdmc1 [7], [8], in which the AtMCM8/AtRAD51 pathway must be active, favors the hypothesis of sister chromatid repair. In contrast, the DMC1 pathway promotes CO formation. However, DMC1 foci in wild type, outnumber COs by approximately 25 to 1 [7], [43]. This suggests that repair of many DSBs catalyzed by DMC1 do not become CO, but NCO (that involve the homologous chromosome) or sister chromatid exchange (SCE). In Arabidopsis, the genome-wide frequency of NCOs and SCEs is currently unknown. We favor the hypothesis that DMC1 promotes NCOs, as DMC1 promotes synapsis. However, it should be noted that DMC1 is also able to promote SCE, notably in the haploid mcm8 context. Indeed, only the simultaneous mutation of AtDMC1 and AtMCM8 in haploids led to extensive chromosome fragmentation. The capacity of DMC1 to promote inter-sister repair was previously shown in other mutant background in both Arabidopsis [9] and yeast [44].
In summary we suggest that two pathways of DSB repair exist in wild type meiosis: The first pathway relies on the strand exchange activity of DMC1, and is also promoted by ASY1, SDS and RAD51 as a co-factor of DMC1 [5]. This pathway generates the COs, but also NCOs and SCEs in a ratio that remains to be determined. The second pathway of the model, which may be viewed as a backup pathway in case of failure of DMC1, relies on the strand exchange activity of RAD51 and the helicase activity of MCM8, and uses the sister chromatid as a template.
The function of MCM8 appears to differ markedly in Arabidopsis and in Drosophila. Interestingly, DMC1 and MCM8 appear to be partially redundant in Arabidopsis while the Drosophila genome seems devoid of a DMC1 homolog [45]. Thus CO formation in Drosophila appears to rely on a RAD51/MCM8 pathway, which has only a minor role in wild type meiotic DSB repair in Arabidopsis. The CO pathways appear to differ considerably in the two species, mainly using ZMMs in Arabidopsis but not RAD1, and the reverse in Drosophila, i.e. RAD1 but not ZMMs (that are absent from the Drosophila genome). Drosophila appears to be unique, as in distant species like S. cerevisiae, mammals and C. elegans CO formation depends mainly on ZMM. Adding to the complexity, MCM8 exists in mammals but not in S. cerevisiae and C. elegans [19], [24]. In mouse, MCM8 mutation leads to a meiotic arrest, with defects in homologous synapsis and over-accumulation of DMC1 foci before apoptosis, suggestive of defects in DSB repair [22]. We would like to suggest that these defects may compatible with MCM8 being required for a backup pathway in the case of failure of DMC1 to repair breaks, like in Arabidopsis. The lack of the backup pathway may lead to the accumulation of DMC1 foci, and a failure to repair a subset of breaks, triggering apoptosis (it is noteworthy that DSB repair defects do not trigger meiotic arrest or apoptosis in Arabidopsis). This illustrates the variety of mechanisms that arose in the course of evolution to fulfill the conserved outcome of meiotic DSB repair and CO formation.
In conclusion, our data reveals the meiotic function of MCM8 in Arabidopsis. Cytological and genetic analyses showed that AtMCM8 is involved in DSB repair but it is not a determinant for CO formation. This study identified a new pathway of meiotic DSB repair independent of AtDMC1.
A. thaliana accession Columbia (Col-0) was the wild type reference. Atmcm8-1 (Salk_032764, N532764), Atmcm8-2 (Salk_104007, N604007) Atmcm8-3 (Salk_099327, N599327) were obtained from the collection of T-DNA mutants at the Salk Institute Genomic Analysis Laboratory (SIGnAL, http://signal.salk.edu/cgi-bin/tdnaexpress) [46] via NASC (http://nasc.nott.ac.uk/). Other mutants used in this study were Atspo11-2 (Gabi_749C12, N359272) [47], Atprd1 (Salk_024703, N524703) [3], Atdmc1-3 (Sail_170_F08, N871769) [48], Atrad51 (Atrad51-1) [6], asy1-4 (Salk_046272, N546272), sds-2 (Sail_129_F09, N806294) [12], Atzip4-2 (Salk_068052, N568052) [43], Atmsh4 (Salk_136296, N636296) [49], mus81-2 (Salk_107515, N607515), mus81-3 (Salk_002761, N502761) [50], [51], and shoc1-1 (Salk_057589, N557589). rad1-1 (uvh1-1) has a EMS (ethyl methanesulfonate) mutation [33], [34] and was provided by C. White.
Plants were cultivated in greenhouse or growth chamber with a 16 h/day and 8 h/night photoperiod, at 20°C and 70% humidity.
Allelism tests were performed by crossing Atmcm8-1+/− with Atmcm8-2+/− and selecting F1 plants hemizygous for both alleles and likewise for Atmcm8-2+/− with Atmcm8-3+/−. Double mutants were obtained by crossing heterozygous plants for each mutation and selfing the double heterozygous F1 plants. Atmcm8/Atmhs4/Atmus81 triple mutant was identified by crossing Atmcm8/Atmsh4 double heterozygous with Atmus81 single mutant. As Atmsh4 and Atmus81 are linked, a plant heterozygous for Atmcm8/Atmsh4 was self-fertilized and homozygous for Atmus81 to identify the triple mutant in the offspring. Haploid Atmcm8 and Atmcm8/Atdmc1 were obtained by crossing a heterozygous plant for Atmcm8 or Atmcm8/Atdmc1 mutations as male and the GEM line as female [36], [52]. In F1, haploid plants of the desired genotype were selected.
Plants of interest were selected by PCR genotyping using diagnostic primer sets. The three AtMCM8 insertions were genotyped by PCR using following primer combinations to amplify genomic DNA flanking the T-DNA insertions. Atmcm8-1: left borders (LB) with LBsalk2 (5′-GCTTTCTTCCCTTCCTTTCTC-3′)/N532764L (5′-AGCGCCATTAGCAAAATGTC-3′) or with LBsalk2/N532764U (5′-GCAGCTTCATTCTGCAAGTG-3′). Wild type allele with N532764U/N532764L. Atmcm8-2 LB with LBsalk2/N604007L (5′- TCACTACAGCAACGGTGAGC -3′), right border (RB) with RBsalk1 (5′-TCA GAG CAG CCG ATT GTC-3′)/N604007U (5′-GCTGATGGAAGACCTTGTGG-3′). Wild type allele with N604007U/N604007L. Atmcm8-3 LB with LBsalk2/N599327L (5′-TGGTGTGGAATCAGCAGATG-3′) or with Lbsalk2/N599327U (5′-TGTGTCTCTGTTGCAAAGGC-3′). Wild type allele with N599327U/N599327L. T-DNA right and left borders were analyzed by sequencing PCR products. AtSPO11-2 wild type allele was amplified using primers 749C12U (5′-GAGCGAGAATTTTTGGTTGG-3′) and 749C12L (5′- CCACAAGGTCAATTCTTCAAC-3′) and mutant allele using N524703L and LBgabi1 (5′-CCCATTTGGACGTGAATGTAGACAC-3′). AtPRD1 wild type allele was amplified using primers N524703U (5′-AAGTCTGCCCATGGTCACGATTCTCTCTG-3′) and N524703L (5′-GCCTGCTCAAAGGGTCCAGC-3′) and mutant allele using N524703L and LbSalk2. AtDMC1 wild type allele was amplified using primers N871769U (5′- TTTTTAATTGTTTACAGAGGAAATCAG-3′) and N871769L (5′-TCCACTCGGAATAAAGCAATG-3′) and mutant allele using N871769L and Lb3sail (5′-TAGCATCTGAATTTCATAACCAATCTCGATACAC-3′). AtRAD51 wild type allele was amplified using primers RAD51-1U (5′-ATGCCAAGGTTGACAAGATTG-3′) and RAD51-1L (5′- CTCCCCTTCCAGAGAAATCTG -3′) and mutant allele using RAD51-1U and LBgabi1 (5′-CCCATTTGGACGTGAATGTAGACAC-3′). We amplified SDS wild type allele using primers N806294U (5′-CTGCTCCCTGATTACAAGCAG-3′) and N806294L (5′-CTTAACGCATTCAGGCAACTC-3′) and mutant allele using N806294U and Lb3sail. AtMSH4 wild type allele was amplified using primers N636296U (5′-CTTCTTGCAGGTTGTGTTTG-3′) and N636296L (5′-GCCAGCTGTTTTTGTTGTC-3′) and mutant allele using N636296L and LbSalk2. AtMUS81A wild type allele was amplified for Salk_107515 using primers N607515U (5′-CATGCTGACAGTTGAAGGTC-3′) and N607515L (5′-CCTCAAACGTTTCTCCAAAT-3′) and mutant allele using N607515L and LbSalk2. AtMUS81A wild type allele was amplified for Salk_002176 using primers N502176U (5′-CACATACGTTTTTGGTTCCC-3′) and N502176L (5′-AGTGTCCAAGTCCTGCTTTC-3′) and mutant allele using N607515L and LbSalk2. AtZIP4 wild type allele was amplified using primers N568052U (5′-TCCTTCCCACACCTTGACCC-3′) and N568052L (5′-GACTGCTGGAGCAGAAACT-3′) and mutant allele using N568052L and LbSalk2. ASY1 wild type allele was amplified using primers N546272U (5′-TCTATGTTTGTTACGCGTTAATCAG-3′) and N546272L (5′-AGGTGGCTCGTAATCTGGTGGCTGC-3′) and mutant allele using N546272L and LbSalk2. SHOC1 wild type allele was amplified using primers N557589U (5′-TTACCGGAGTTTGAAAACCG-3′) and N557589L (5′-GGCAAAGACTTGAAGGCATC-3′) and mutant allele using N557589L and LbSalk2. AtRAD1 was amplified using primers o629 (5′-CTGGTGAAGAACATTTGGTAG-3′) and o630 (5′-CTCTTATGGCTGCTGCGTCTTC-3′). Polymorphism between wild type and mutant alleles was revealed with Dde1 digestion.
FTL lines were obtained from G.P. Copenhaver. For this study, we used two couple of adjacent intervals: I5aI5b and I5cI5d [31]. The procedure to create plants of interest and to collect data was described in [31], [53]. Statical analysis was performed as described in [31].
Alexander staining for pollen viability was performed as described [54]. The protocol described by [55] was used to observe the female meiosis and the protocol described by [29] for male meiotic spreads. Immunolocalization of AtMLH1 was made as described by [29]. Immunolocalization of AtZYP1 and AtDMC1 was performed according to [56] with the modifications described in [43]. The anti-ASY1 polyclonal [56] and anti-ZYP1 polyclonal [49] antibodies were used at a dilution of 1∶250. The anti-MLH1 antibody [29] was used at a dilution of 1∶200. The anti-DMC1 antibody was described in [43] and the purified serum was used at 1∶20.
For male meiotic spreads, observations were made with a Leica DM RXA2 epifluorescence microscope using an oil PL APO 100X/1.40 objective (Leica). Photographs were taken using a CoolSNAP HQ (Roper Scientific) camera driven by Open-LAB 4.0.4 software (Improvision). For immunocytology and FTLs analyzes, observations were made using a Zeiss Axio Imager2 microscope. We analyzed FTLs using the automatic slide-scanner function of the ZEISS AxioObserver DIC FISH Apotome and its workbench. Photographs were taken using an AxioCam MRm (Zeiss) camera driven by Open-LAB 4.0.4 software AxioVision 4.8. All pictures were processed with AdobePhotoshop 7.0 (Adobe Systems Inc.).
|
10.1371/journal.pbio.1001489 | A Force-Activated Trip Switch Triggers Rapid Dissociation of a Colicin from Its Immunity Protein | Colicins are protein antibiotics synthesised by Escherichia coli strains to target and kill related bacteria. To prevent host suicide, colicins are inactivated by binding to immunity proteins. Despite their high avidity (Kd≈fM, lifetime ≈4 days), immunity protein release is a pre-requisite of colicin intoxication, which occurs on a timescale of minutes. Here, by measuring the dynamic force spectrum of the dissociation of the DNase domain of colicin E9 (E9) and immunity protein 9 (Im9) complex using an atomic force microscope we show that application of low forces (<20 pN) increases the rate of complex dissociation 106-fold, to a timescale (lifetime ≈10 ms) compatible with intoxication. We term this catastrophic force-triggered increase in off-rate a trip bond. Using mutational analysis, we elucidate the mechanism of this switch in affinity. We show that the N-terminal region of E9, which has sparse contacts with the hydrophobic core, is linked to an allosteric activator region in E9 (residues 21–30) whose remodelling triggers immunity protein release. Diversion of the force transduction pathway by the introduction of appropriately positioned disulfide bridges yields a force resistant complex with a lifetime identical to that measured by ensemble techniques. A trip switch within E9 is ideal for its function as it allows bipartite complex affinity, whereby the stable colicin:immunity protein complex required for host protection can be readily converted to a kinetically unstable complex whose dissociation is necessary for cellular invasion and competitor death. More generally, the observation of two force phenotypes for the E9:Im9 complex demonstrates that force can re-sculpt the underlying energy landscape, providing new opportunities to modulate biological reactions in vivo; this rationalises the commonly observed discrepancy between off-rates measured by dynamic force spectroscopy and ensemble methods.
| Many proteins interact with other proteins as part of their function. One method of modulating the activity of protein complexes is to break them apart. Some complexes, however, are extremely kinetically stable and it is unclear how these can dissociate on a biologically relevant timescale. In this study we address this question using protein complexes between colicin E9 (a bacterial toxin) and its immunity protein Im9. These highly avid complexes (with a lifetime of days) must be broken apart for colicin to be activated. By using single-molecule force methods we show that pulling on one end of colicin E9 drastically destabilises the complex so that it dissociates a million-fold faster than its intrinsic rate. We then show that preventing this destabilisation (by the insertion of cross-links that pin the N-terminus of E9 in place) yields a kinetically stable complex. It has previously been postulated that force can destabilise a protein complex by partially unfolding one or more binding partners. Our work provides new experimental evidence that shows this is the case and provides a mechanism for this phenomenon, which we term a trip bond. For the E9:Im9 complex, trip bond behaviour allows a stable complex to be rapidly dissociated by application of a surprisingly small force.
| Protein-protein interactions are integral to diverse cellular processes such as catalysis, transport, and signalling. For complexes of low affinity, changes in the relative concentrations of one (or more) binding partners, or alterations in the environment, are sufficient to trigger complex dissociation, allowing spatial and temporal control of the processes in question. More stable complexes require the input of chemical energy such as that provided by AAA+ proteins for their dissociation [1]. For high affinity complexes without direct access to an energy source, it remains unclear how dissociation can be induced on a biologically relevant timescale. This problem is exemplified by the rapid dissociation (lifetime≈minutes) of highly avid bacterial colicin:immunity protein complexes (Kd≈10−14 M; lifetime≈days) upon binding to the outer membrane of target bacteria [2].
Colicins are protein antibiotics synthesised by E.coli strains to target and kill related bacteria during environmental stress [3]. The E-type colicins of group A, which include E2, E7, E8, and E9, exert their toxicity via nuclease activity. These multi-domain proteins (Figure 1A) contain a receptor domain (R) required for initial binding (to BtuB), a translocation domain (T) used to bring about translocation via interaction with OmpF and a cytoplasmic DNase domain (C, residues 480–582 of full length colicin, termed E9 herein) that results in death of the competing cell subsequent to translocation of this domain to the victim's cytoplasm (Figure 1B). To prevent host suicide, colicins are expressed alongside specific immunity proteins (Im2, Im7, Im8, and Im9) [2], which inactivate colicin enzymatic activity by binding to an exo-site adjacent to the active site (Figure 2A and 2B) [4]. The binding interface is typical for protein-protein complexes, covering a surface area of 1,575 Å2 for E9:Im9, which spans residues 72 to 98 of the nuclease domain [5]. The highly homologous cognate colicin:immunity protein pairs have high affinities (Kd≈10−14 M) [6], while non-cognate pairs bind less tightly (e.g., for E9:Im2, Kd≈10−7 M) [7]. The binding affinity of colicin:immunity protein complexes is determined by two binding “hotspots” on the immunity protein that interact with a distinct binding epitope on E9. Firstly for all colicin:immunity protein complexes, stabilising interactions are formed with residues in helix III of the immunity protein. This helix is identical in sequence in all E-type immunity proteins (apart from one residue in Im7 [6]). Residues within helix III of the immunity protein contact Phe86 and residues in the surrounding hydrophobic pocket of E9. This pocket comprises Tyr 83, Val 98, and the alkyl chains of Lys 89 and Lys 97 of the colicin DNase domain (Figure 2B, highlighted in lilac). Colicin:immunity protein affinity is modulated by stabilising (cognate complexes) or destabilising (non-cognate complexes) interactions between specificity-determining residues of helix II in the immunity protein (which differs significantly in sequence in different proteins; Figure 1C) and the binding interface of the colicin DNase domain (Figure 2) [8],[9]. As the on-rate for cognate and non-cognate colicin:immunity protein complexes is diffusion-limited (kon≈10−8 M−1s−1) [10],[11], the observed differences in affinity, which span almost ten orders of magnitude (Kd = 10−14 (E9:Im9) to 10−4 M (E9:Im7)) [7],[8],[11] are manifested in off-rates that differ by seven orders of magnitude (koff = 10−6 to 101 s−1 for the cognate and non-cognate complexes, respectively).
The large differences (107) in off-rates of different colicin:immunity protein complexes render this system an excellent model for investigating the molecular determinants of molecular recognition and, in particular, for exploring how highly avid complexes can be remodelled in vivo in the absence of an external energy source to allow rapid dissociation when required for biological activity. While the precise mechanism of E-type colicin:immunity protein dissociation is unclear, it is known that colicin invasion uses molecular mimicry to subvert a series of protein-protein interactions that result in linkage of the colicin (bound to the outer membrane) to the TolQRA complex of the energised inner membrane (a translocon; Figure 1B). As TolQRA function and colicin intoxication both require a proton gradient across the inner membrane [12], it has been postulated that the energy of the proton motive force (PMF) may be harnessed to drive colicin:immunity protein dissociation, a necessary prerequisite of translocation of the nuclease to the cytoplasm.
“Inside-out” energy transduction mechanisms are exemplified by the Ton system, which is highly homologous to TolQRA (both require a functioning PMF to carry out their function [13]). In the case of the Ton system, PMF-driven remodelling of the plug domain of the outer membrane protein BtuB allows siderophore import. Remodelling is also thought to play a role in E9 colicin intoxication as cross-linking residues 20 and 66 of the nuclease domain prevents insertion into planar lipid bilayers and protects against cellular toxicity [14]. In accord with a requirement for structural remodelling in the mechanism of colicin invasion, immunity protein release that is usually triggered by formation of the translocon in the presence of the PMF, is prevented by cross-linking of the N- and C- termini of the R-domain (Figure 1B) [15].
Here we use single molecule force methods (using atomic force microscopy) to investigate the requirement for structural remodelling in the dissociation of single E9:immunity protein complexes under defined rates of loading and pulling geometry. The effects of structural re-arrangements in proteins can be investigated by many approaches, but most apply a “peturbant” globally. Single molecule force methods (which use mechanical extension as a peturbant) are ideally suited for such an investigation as force is applied locally to the complex at positions determined by the sites of linker attachment. Using this approach, we show here that a low level of force (<20 pN) commensurate with that applied by protein molecular motors [16]–[18] increases the dissociation rate of the E9:Im9 complex in vitro by a remarkable 106-fold. Using mutagenesis and disulfide cross-linking, we also elucidate the force transduction path through E9, which catalyses complex dissociation, and show that this involves conformational remodelling of E9 triggered by mechanical deformation of its terminal region. The data show that mechanical force can be exploited to enable rapid dissociation of the high affinity colicin:immunity protein interaction by application of force at the N-terminus of E9.
We used atomic force microscopy to measure the dynamic force spectrum of the unbinding of single complexes of the nuclease C-domain of colicin E9 bound to its cognate immunity protein (Im9), together with non-cognate complexes of E9:Im2 and several variants of Im9 containing point mutations in the binding site [7],[8]. The experimental design is depicted in Figures 1D and S1. Briefly, single cysteine residues were introduced into E9 (the wild-type protein lacks cysteine residues) and pseudo–wild-type variants of Im9 and Im2 in which the single naturally occurring cysteine was first mutated to alanine (C23A), to enable immobilisation of each protein specifically to the substrate or cantilever. Sites chosen for mutation to allow immobilisation (S3C, S30C, or S108C in E9 and T38C or S81C in the immunity protein) were solvent accessible and distal to the E9:immunity protein binding interface (Figure 2A and 2B). The immunity protein and E9 were next attached to the atomic force microscope (AFM) tip and substrate, respectively, using hetero-bifunctional polyethylene glycol (PEG) linkers of variable length (Materials and Methods; Figure S1). Gel filtration was used to compare the ensemble off-rates of the wild-type complex and one containing mutated E9 derivatised with PEG linker (kioff = 1.8×10−6 s−1 and 5.8×10−6 s−1 for wild-type [8] and derivatised [E9 S3C:Im9 (S81C)] complexes [Figure 3A], respectively). These data, together with a nuclease assay (Materials and Methods; Figure 3B) showed that neither sequence changes nor PEG derivatisation significantly affected the properties of E9 alone or in complex with Im9. Complexes were repeatedly formed and dissociated by approach-retract cycles of the functionalised AFM tip towards and away from the surface at a defined velocity (Materials and Methods). Unbinding resulted in a single force peak characteristic of a single molecule unbinding event (Figures 4A, bottom, lower plot, and Figure S2) for 99.5% of all force-extension profiles that showed any evidence of interaction between the tip and substrate (typically 10% of all approach-retract cycles). All force-extension data are presented and analysed after accounting for the deflection of the AFM tip (i.e., the distance between apex of the AFM tip and the substrate surface).
Initially, force-extension profiles that displayed a detectable unbinding event greater than 5 nm from the surface (to avoid non-specific tip-surface interactions, see Text S2) were analysed without further filtering. Figure 4B shows a scattergram contour plot and individual frequency histograms for the unbinding force and contour length at rupture for every unbinding event in each force-extension profile of a single dataset. These data show that unbinding events occurred over a narrow range of extensions (mode = 10 nm) suggesting that unbinding occurs by a single pathway (Figure 4A, top). Interestingly, the measured contour length is significantly shorter than that expected (Figure 4A bottom, black solid line) based on the sum of the lengths of each linker (6.62 nm each) and the through space distance between the extension points on E9 and Im9. This distance is 4.72 nm when extending the complex from residue 3 on E9 and residue 81 on Im9. This complex is denoted 3:81 (similar nomenclature is used throughout). To predict the expected contour length more accurately, it is necessary to account for the ability of Im9 to be immobilised anywhere between the apex and the base of the AFM tip and the distribution of end-to-end lengths within the ensemble of the polymeric linkers (Figure S3). Taking these effects into consideration (using a Monte Carlo simulation; Figure S4 and Text S1) yielded a contour length significantly shorter than observed (7.57 nm, Figure 4A bottom, red line). Importantly, these calculations suggest that the complex undergoes deformation or elongation prior to dissociation (see Discussion).
In order to quantify the unbinding forces and loading rates at rupture, force-extension profiles were subsequently analysed using an automated analysis script whereby events (<13% of total force-extension profiles; ) were filtered from non-specific interactions on the basis of their force-extension profiles. To be binned for further analysis (Text S2) a force-extension profile was required to (i) have a rupture force larger than the thermal noise of the experiment (18 pN; Figure S5 and Text S3); (ii) have a distance to the rupture event from the hard tip-surface contact that was between 5 and 32.5 nm or 5 and 40 nm for protein complexes immobilised using (PEG)4 and (PEG)12, respectively. The lower limit avoids the analysis of any non-specific tip-sample interactions and the upper limit is significantly greater than the expected rupture distance so that all events are analysed; and (iii) display only a single unbinding event.
The ability of the experimental setup and data analysis method to recognise only specific E9:Im9 unbinding events was verified using two controls. Firstly, the addition of excess immunity protein to the solution between the AFM tip and substrate resulted in a decrease of the frequency of unbinding events from more than 12% to less than 1% (Figure 4B). Secondly, the addition of EDTA was found to decrease the event frequency 3-fold. EDTA sequesters divalent metal cations from E9, destabilising the protein substantially (Tm = 36°C and 68°C for apo- and zinc-bound E9 [19]) leading to a loss of binding to Im9. Addition of excess Zn2+ restored E9 stability and, consequently, event frequency (unpublished data).
The force and loading rate at unbinding of E9:Im9 were measured for each event from force-extension data (Text S2). Force-frequency distributions (Figure S6) were subsequently calculated for each dataset (typically 100–200 events; Table S1), allowing the extraction of the most probable unbinding force (Figure S6; Text S4) and the loading rate at rupture. The dynamic force spectrum of each complex was then revealed by quantifying how the force at rupture varies as a function of the force loading rate between 700 and 180,000 pNs−1 (Text S5).
The dynamic force spectrum of E9:Im9 dissociation was initially measured by immobilising E9 close to the N-terminus (residue 3) (Figure 2) as this region is immediately adjacent to the R-domain and contiguous with the T-domain, which is translocated during colicin intoxication in vivo (Figure 1B). No force is likely to be applied directly to the immunity protein in vivo. A pulling location was thus selected for Im9 (residue 81) (Figure 2) that is solvent exposed and distal to the binding interface. Sample force-frequency histograms that span the range of loading rates used (700–180,000 pNs−1) and the resultant dynamic force spectrum for this complex (3:81) are shown in Figures S6 and Figure 5A, respectively. Two force regimes are evident. At low loading rates (<5,400 pNs−1), dissociation occurs at low forces with a shallow dependence of unbinding force on the loading rate. This allows rapid dissociation of an avid complex at biologically accessible loading rates [16],[17],[20], or by application of biologically accessible forces (see [16]–[18] and references therein). For example, at a force of 20 pN, the lifetime of E9:Im9 is approximately 12 ms, in contrast to 4.1 d in the absence of force. At higher loading rates (>5,400 pNs−1) the complex is highly force resistant and the unbinding force is strongly dependent on the loading rate. The simplest explanation for these observations is that unbinding occurs by a three-state mechanism: at low forces unbinding rates are limited by a barrier in the energy landscape distal to the bound ground state of the complex (a large xu; Figure 5B, xuo). At higher forces, tilting of the energy landscape results in a previously hidden inner barrier (a small xu) becoming rate limiting (Figure 5B, xui). Such a mechanism is consistent with the dual-recognition (un)binding pathway for E9:Im9 determined using ensemble fluorescence experiments (Figure 5B, top) [8]. In this mechanism, the affinity of the initial encounter complex is determined by interactions between residues of the E9 binding interface (Figure 2, highlighted in lilac) and helix III of the immunity protein (residues S50, D51, I53, and Y55) [5]. Rigid body rotations of the initial encounter complex then allow the formation of stabilising (cognate) or less stabilising (non-cognate) interactions between E9 and specific residues in helix II of the immunity protein [7]. Accordingly, the outer barrier measured by DFS that is rate determining at low rates of forced unbinding is expected to report on the free energy difference between the native state and the barrier for dissociation of E9 from helix III of Im9, while the inner barrier that is rate determining at high rates of forced unbinding is expected to report on the energy gap between the native state and the barrier for dissociation of E9 from helix II (Figure 5B).
To confirm the apparent similarity between the force- and thermally activated unbinding mechanisms of E9:Im9, each linear region of the dynamic force spectrum was fitted to the Bell-Evans equation [21] (Equation 1, where f* is the most probable unbinding force, rf is the force loading rate at rupture, T is the temperature, and kB is Boltzmann's constant). This allows the dissociation rate constants in the absence of force (k0Foff) and the “distance” along the free energy landscape from the bound state to the barrier that is rate limiting for dissociation (xu) to be obtained (Table S1).(1)Obtaining k0Foff values by this method assumes that the outer barrier observed in the dynamic force spectrum remains rate-limiting at lower loading rates that are inaccessible to this technique (see [22],[23] for example). These parameters were found to be k0Foff = 50±17 s−1, xu = 0.9±0.2 Å, and k0Foff = 4.9±1.3 s−1, xu = 5.8±0.4 Å for the inner and outer barriers of E9:Im9 dissociation, respectively. If forced unbinding (at the loading rates applied in these experiments) occurs over the same energy landscape as for the thermally induced pathway, the extrapolated k0Foff for the outermost barrier (the rate determining step at low force), determined by DFS should be identical to that measured by ensemble methods (kioff) [22],[24]–[28]. Remarkably, forced unbinding of E9:Im9 results in a k0Foff that differs from kioff by a striking six orders of magnitude (≈100 and 10−6 s−1, respectively).
To examine whether the rapid rate of force-induced dissociation of E9:Im9 is observed for other E9:immunity protein complexes, the dynamic force spectrum of a non-cognate complex (E9:Im2 (D33A), kioff = 0.054 s−1) [7] was next examined (Figure 6A). This variant was selected since it has a higher affinity for E9 compared with wild-type Im2 (Kd = 1×10−9 M and 1.5×10−7 M, respectively) [7]. Again two force regimes were observed in the dynamic force spectrum of this complex, each of which has a similar xu value to that observed for each barrier of the cognate E9:Im9 complex. This indicates that force-induced unbinding of the cognate and non-cognate complexes occurs by a similar three-state mechanism. At low loading rates that probe the rate determining outer barrier, the unbinding forces (and thus k0Foff) were closely similar for the cognate and non-cognate complexes (Figure 6A; Table S1). Under force, E9:Im2(D33A) thus behaves identically to E9:Im9 (Figure 7, solid dark grey and orange bars, respectively), despite kioff values that differ by four orders of magnitude. Similar to the behaviour of E9:Im9, k0Foff for E9:Im2(D33A) is also faster than its known kioff (7.6 s−1 versus 0.054 s−1) [7], indicating that the underlying energy landscape for immunity protein dissociation from E9 is highly sensitive to the effects of force, regardless of the nature of the bound immunity protein.
By contrast with the dynamic force spectrum of colicin:immunity proteins at low loading rates (governed by the outer barrier), the inner barrier for dissociation of E9:Im2(D33A) is reduced significantly compared with that of the cognate complex and only becomes visible at loading rates >22,000 pNs−1 (Figure 6A). In these kinetic unbinding experiments each force regime is assumed to probe the free energy difference between the bound state and each barrier to unbinding. As the energy difference between the bound and free states is reduced for E9:Im2(D33A) relative to E9:Im9 (Kd is reduced 105-fold [7]), a reduction in unbinding force would be expected for both the inner and outer barriers. Instead, application of force at residue 3 of E9 appears to decouple the dual recognition sites of helices II and III of the immunity protein with E9. The inner barrier measures the “strength” of the specificity residues in helix II of the immunity protein and E9 (which are stabilising for Im9 and less stabilising for Im2(D33A); Figure 1C), while the outer barrier height is determined by the stability of interactions between E9 and immunity protein helix III (identical in sequence across all DNase E-colicins except a Thr substitution for Ser at position 51 of Im7) [6].
As discussed above, the inner and outer barriers appear to be due to the dissociation of immunity protein helices II and III, respectively, from the binding surface of E9. To confirm this assignment, DFS was used to measure E9 unbinding from Im9 variants containing single point mutations that destabilise either the cognate specificity interactions of the inner barrier (V34A in helix II), or interactions that define the outer barrier (D51A in helix III) [8] (Figure 6B and 6C). As predicted, unbinding forces for E9:Im9(V34A) were identical to those for wild-type E9:Im9 at loading rates <5,400 pNs−1, but were reduced by more than 35 pN compared with E9:Im9 at loading rates higher than this (xu remained constant for both barriers). By contrast, at loading rates <3,000 pNs−1, the unbinding forces for E9:Im9(D51A) were reduced to a level below the thermal noise limit of the instrument. However, at higher loading rates, E9:Im9(D51A) behaves similarly to the wild-type E9:Im9 complex. These data are consistent with the proposed dual-site recognition process for colicin:immunity protein (un)binding [8] with force effectively uncoupling the unbinding of helices II and III of Im9 from the E9 binding surface. The ability to assign each regime of the dynamic force spectrum to the unbinding of the two recognition sites of the complex previously identified by ensemble methods [8] renders the presence of an additional hidden barrier unlikely. If present, a hidden barrier would require an xu value of greater than 3.7 nm to obtain a k0Foff value commensurate with kioff. This is larger than the outermost barrier previously observed for the dissociation of biotin from avidin [22]. In contrast to the characteristically flat recognition surface of E9:Im9 that is typical of protein-protein interactions in general, biotin resides in a deep pocket within avidin. We thus consider the presence of an additional barrier unlikely. Overall, therefore, the results indicate that the application of force distal to the E9:Im9 interface enables rapid dissociation of this tight binding complex such that the dissociation rate is enhanced by greater than a million-fold to a timescale commensurate with the kinetics of cell killing by colicins (within minutes) [2],[15],[29].
Force induced conformational changes are known to trigger catalysis [30] or expose “cryptic” binding sites [31] in some proteins. These remodelling events are usually very sensitive to the points of force application as proteins are known to display anisotropic force responses. Thus, when extended in different directions proteins can appear to be mechanically strong or weak [32]–[35]. To investigate whether this effect is the origin of the force-induced lability of E9:immunity protein complexes, the effect of altering the pulling location on the dynamic force spectrum of the E9:Im9 complex was examined. Accordingly, different residues on E9 and Im9 (positions 3, 30, and 108 on E9 and positions 38 and 81 on Im9) were mutated individually to Cys to enable immobilisation to the surface at different points (Figure 2). These experiments showed that k0Foff for the outer (rate-limiting) barrier remains 105- to 107-fold higher than kioff regardless of the pulling location employed (Figures 7 and 8; Table S1). Nonetheless a small, but significant, dependence of the unbinding force (Figure 8A) and k0Foff (Figure 7) on the immobilisation site on E9 was observed, with the highest k0Foff values occurring when E9 was pulled from an N-terminal location (residue 3, k0Foff = 4.9 s−1) and lower values occurring when E9 was immobilised at position 108 or 30 (k0Foff = 1.5 and 0.4 s−1, respectively; Figure 7 and Table S1). By contrast, k0Foff was insensitive to the pulling location on Im9 (Figure 8B). The anisotropy in k0Foff in relation to the E9 pulling location, together with the large disparity between k0Foff and kioff values and the increase in chain length upon dissociation being greater than expected based on linker length (Figure 4A), suggest that remodelling or partial unfolding of E9 takes place under force. This then yields a dissociation pathway with a smaller activation free energy than is accessible in the absence of remodelling. The ability to alter the unbinding kinetics by force-induced substrate remodelling has recently been postulated [36],[37]. The results presented here show a striking example of this phenomenon, with rate enhancements of a million-fold caused by application of only 20 pN force, at most.
The data described above demonstrate that the level of acceleration in the dissociation rate of E9 from Im9 is sensitive to the precise location of force application and that the rate enhancement is greatest when force is applied close to the N-terminus of E9. Examination of the structure of E9 shows that the N-terminal 30 residues (highlighted in red, Figure 2B) do not contact the immunity protein binding interface directly. Leu 23 and Ala 26 of the N-terminal region of E9 do, however, form part of hydrophobic core of E9 formed around Trp 58 that also encompasses residues of the binding interface (Val 79, Pro 85, and Tyr 99) (Figure 2B). The N-terminal region of E9 may thus relay the force trigger to an allosteric site of affinity modulation. To understand the signal transduction pathway in more detail, and to identify the location of the allosteric site that translates the mechanical stimulus to an increase in dissociation rate, a series of mutant E9 domains were produced containing disulfide bonds in different locations of the protein structure (Figure 2A and 2B). Disulfide bond formation in all of these variants was shown to be spontaneous and to proceed to completion using Fourier transform ion cyclotron mass spectrometry (Figure S7). The dynamic force spectrum of E9:Im9 complexes extended from the N-terminal region of E9 (3:81) engineered to contain a disulfide bond that links the N-terminal region of the polypeptide chain to the remainder of the folded globular region of E9 (linking residues 13–117 or 20–66; Figure 2) are shown in Figures 9A and 9B, respectively. Remarkably, both of these cross-linked E9 variants display a simple monotonic dynamic force spectrum over the entire accessible loading rate range with significantly increased unbinding forces (ΔF≈100 pN relative to wild-type complexes). The value for xu, however, is similar to that observed for the outer barrier in the dynamic force spectrum of the uncross-linked, wild-type E9:Im9 complex. Fitting these data to the Bell-Evans model yields k0Foff values for these complexes that are increased by ≈106-fold, resulting in values for k0Foff that are similar to those measured using ensemble techniques (k0Foff = 4.6×10−6 s−1 and 1.4×10−6 s−1 for 313–117:81 and 320–66:81, respectively; Figure 7). Values of kioff measured using gel filtration experiments under identical conditions to those employed for the AFM experiments are 3.0×10−6 and 5.8×10−6 s−1 for E920–66:Im9 and pseudo wild-type E9:Im9 derivatised with methyl-(PEG)12-maleimide, respectively (Materials and Methods; Figure S8; Table S1). Addition of 4 mM DTT reversed this mechanical strengthening, leading to unbinding forces identical to those of wild-type E9:Im9 (Figure S9). To localise the region of E9 involved in force remodelling more precisely, the dynamic force spectrum of a third E9:Im9 (3:81) complex that contains a disulfide cross-link distal to the N-terminal region of E9 (31–122; Figure 2B) was analysed. In this case no force enhancement was observed. Instead, a dynamic force spectrum with a single force regime was obtained, identical to that of the outer barrier of the wild-type uncross-linked complex (Figure 9C). These data localise the allosteric trigger to residues 21–30 or to residues 118–121 in E9. As E9 is extended from the N-terminus in these experiments we consider the latter site to be unlikely as the site of the trigger.
The data presented here reveal that insertion of a disulfide bond is able to modulate how force is propagated through the nuclease domain of colicin E9 preventing remodelling of E9 and thus facile complex dissociation. Extension of the complex in a geometry that propagates the force via a different path would thus be expected to render the cross-link between residues 20 and 66 less effective. In accord with this hypothesis, k0Foff values for E920–66:Im9 complexes were found to be dependent on the position at which force is applied to both E9 and Im9 (k0Foff = 4.0×10−5 s−1, 1.4×10−6 s−1, and 5.3×10−4 s−1 for 320–66:38, 320–66:81, and 10820–66:81, respectively; Figures 7 and 9D; Table S1).
The sensitivity of k0Foff to the pulling location on E9 and to the presence of cross-links in the N-terminal region of this protein demonstrates that force can act as an allosteric trigger for E9:Im9 complex dissociation. We have identified residues 21–30 as the most probable location of this trigger, a region that both links to the N-terminus and contacts residues involved in the binding interface. To be an effective transducer of mechanical signals, the N-terminus of E9 (residues 1–20) would be expected to be mechanically labile. Analysis of the sequences of colicin E2, E7, E8, and E9 reveal that the N-terminal region of all four nuclease domains is highly conserved and has a high content of small aliphatic amino-acids (RNKPGKATGKGKPVGD; Figure 10A). This region of the protein thus docks against the remainder of the globular domain with little side-chain interdigitation, commensurate with the requirements of a trigger activated at low forces [38]. The sequence thus appears ideally suited to transmitting mechanical signals to the binding interface at low force.
The close equivalence of k0Foff and kioff of E9:Im9 upon cross-linking the N-terminal region of E9 to the remainder of this globular protein suggests that unbinding under ambient conditions and that induced by force occur by the same pathway. In this case, co-operativity between binding hot-spots on the immunity protein helices II and III is restored and mutation of residues in either helix should yield changes in the rate-limiting outer barrier that correlate with the change in affinity for that complex. Analysis of 320–66:81 E9:immunity protein complexes that vary in their binding affinity from the tightly bound Im9 (Kd = 1.6×10−14 M), through Im2(D33A) (Kd = 1.0×10−9 M), to the weakly bound Im2 (Kd = 1.5×10−7 M) each yield k0Foff values close to the previously measured ensemble kioff values (Figures 7 and S10). Together, these data provide further evidence that cross-linking switches the force-induced unbinding pathway (which involves remodelling of the E9 subunit within the complex) to a cooperative event that closely matches the thermally induced unbinding mechanism.
Only a single monotonic force regime is observed in the dynamic force spectrum of all complexes that contain a disulfide bridge, irrespective of their mechanical phenotype or Kd value that varies over seven orders of magnitude. This finding may reflect the absence of the second inner barrier (due to the co-operativity between each binding hotspot), or changes in the relative height of each barrier that results in an altered route of force propagation that moves the inner barrier to a loading rate beyond the dynamic range available to AFM. Irrespective of these changes to the energy landscape, the observation of k0Foff values that concur with previously determined kioff values reveals that the rate-limiting step for the forced and thermally activated pathways is similar when the structural pliability of E9 is minimised by bolstering the E9 structure with disulfide cross-links. This observation has important implications for interpreting dynamic force experiments on proteins with mechanically labile structures and helps to explain the differences in off-rates frequently observed obtained by ensemble and DFS methods [39],[40]. The million-fold increase in koff measured for E9:Im9 represents a striking example of this phenomenon.
The evolution of protein sequences has generated a rich repertoire of finely tuned protein-protein interactions whose binding affinities span ≈13 orders of magnitude [41]. Some complexes (barnase-barstar, for example) have evolved to bind tightly and to have a long lifetime (≈1.5 d at pH 8.0) [42]. Other, equally avid, complexes (for example SNARE complexes) need to dissociate more frequently for biological function [43]. Whilst altering protein sequence can modulate the binding affinity and the on- or off-rates of protein complexes, in some cases by many orders of magnitude [8],[44],[45], force-induced substrate remodelling offers further opportunities to tune the energy landscape of complex formation and dissociation. For example, interactions can become stronger (catch bonds) [46], or weaker (slip bonds) [27],[28] upon the application of force, and hidden epitopes required for binding can be exposed by forced unfolding (cryptic motifs) [31]. In the case of cell-cell adhesion mediated by protein-protein interactions, combinations of these have also been identified [47].
Here we have shown a striking example of how force-induced substrate remodelling can modulate complex stability (Figure 10B and 10C). At low loading rates and forces (<20 pN [48],[49]), highly avid cognate E9:immunity protein complexes dissociate in tens of milliseconds. Such lifetimes are ≈106-times shorter than the thermally induced off-rate for the same complex (4.1 d) revealing a remarkable sensitivity of lifetime to force. These force-induced lifetimes are commensurate with the timescale for colicin intoxication of bacteria.
By altering the points of immobilisation and introducing disulfide cross-links at different locations, we identify the N-terminal region of E9 as a force transducer and suggest residues 21–30 as the location of the allosteric effector of force-triggered dissociation. The N-terminal region of E9 is highly conserved with a high content of Ala and Gly residues that render this region of the protein conformationally pliable. Such a sequence provides the ideal circuitry to relay a conformational trigger to the allosteric switch that lies close to the protein complex interface. This conformational rearrangement results in contour lengths at dissociation of all the complexes that are greater than expected (Figure S11; Text S1). This could reflect a degree of local unfolding in one or both proteins involved in the complex, or could result from deformation/elongation of the intact complex under force application prior to its dissociation.
In this study we have used an AFM to apply a stimulus to trigger remodelling of the E9:Im9 interface. In vivo, this triggering force may be driven by conformational re-arrangements caused by changes in the environment or by changes in other domains of the colicin that are transmitted to the DNase domain. Indeed, introduction of a disulfide cross-link across the N- and C-terminal regions of the R-domain of colicin, which is N-terminal to the nuclease domain (Figure 1B) prevents immunity protein release upon translocon formation [15]. In addition to local conformational changes upon formation of the translocon complex, induced conformational changes may drive immunity protein dissociation by (i) differential rates or extent of diffusion of the inner and outer membranes (or protein domains within these) [20] that are linked by the docked colicin:BtuB:OmpF:TolB translocon, or (ii) an energised motor-like function of TolQRA domains on the inner membrane. However, other stimuli may also result in the remodelling of the allosteric trigger. For example, facile dissociation of a colicin:immunity protein complex has also been reported for E3:Im3. In this case, binding of the complex to a strong anion exchange resin was suggested to induce conformational changes in the immunity protein that resulted in colicin release [50]. The responsiveness of E9 to its environment has been further demonstrated by the observation that insertion into a negatively charged membrane (required for colicin intoxication) is prevented by introduction of one the disulfide linkages (20–66) that we show here to prevent E9 remodelling.
In addition to its biological implications for colicin intoxication, our study provides direct experimental evidence that force can induce changes in the energy landscape measured by dynamic force spectroscopy using the AFM and, for E9:Im9, provides a mechanism by which this occurs. The data show that, in addition to tilting of a “zero force” landscape as predicted and quantified by Bell, force can re-sculpt the underlying energy landscape. For E9:Im9 dissociation, we show that these changes allow facile dissociation of an avid complex at low forces. For example at 25 pN the dissociation rates of wild-type and cross-linked E9:Im9 complexes are 163 s−1 and 1.8×10−4 s−1. The ability to re-sculpt the energy landscape by force provides biology the opportunity to break apart highly avid complexes in the absence of a direct source of energy. In support of this, discrepancies between kioff and k0Foff are observed for many complexes [39],[40] but, in contrast to the 106-fold difference in off-rates observed for E9:Im9, these differences are typically relatively small (102 at most). This suggests that colicin sequence and structure have evolved to enable triggered unbinding that is required for their biological function. In rare cases, such as the dissociation of an antigen from a kinetically and mechanically stable single-chain antibody (an immunoglobulin-like domain) excellent agreement between off rates is observed between ensemble and dynamic force spectroscopy methods [28]. This study, together with the identity of k0Foff and kioff for the dissociation of Im9 from disulfide bridged E9 variants (residues 13 and 117 or 20 and 66) adds further support that conformational remodelling can drive dissociation in vivo. The remodelling force could be generated in many ways, such as by energy-dependent remodelling enzymes (AAA+ proteins, for example), by the binding of new ligands leading to changes in the dynamics or conformation of the complex, or by changing the chemical environment.
The force-induced switching between populations of protein complexes with distinct mechanical properties has been observed previously for the nuclear transport complex Ran:importin β [51] and subunits of von Willebrand factor involved in blood clotting [52]. While the mechanism underlying the force switch varies in these two cases, the application of force results in a switch to a more force resistant slip “bond” or, for von Willebrand factor, to a flex “bond.” (Note: these are not single bonds but a series of non-covalent interactions). For E9:Im9 the situation is reversed in that force induces a transition from a high resistance scenario (low koff) to a low force resistance slip bond (high koff). This is akin to a trip wire (a “trip bond”), whereby small forces trigger the remodelling of an interface that is very stable in the absence of force. The identification of a trip bond thus adds to the repertoire of behaviour of biomolecules under force that has emerged over the last decade [30],[46],[47],[52],[53] and provides a mechanism to explain the discrepancy in off-rates often observed between ensemble and DFS measurements. For colicin function, the force response of a trip bond meets the seemingly mutually exclusive requirements to provide long term protection to the host, yet permit the facile dissociation of immunity protein that is required for cell invasion of its competitors.
Triple cysteine variants of E9 were designed using “Disulfide by design” software [54]. All proteins were created and purified as described previously [14].
Silicon substrates were first cleaned by sonication in chloroform for 30 min and silicon nitride AFM cantilevers were cleaned by rinsing with chloroform for a minimum of 5 min. The substrates and AFM probes were then exposed to UV radiation (254 nm) for 30 min. Following this, surfaces to be functionalized were held under vacuum in the presence of 80 µl (3-aminopropyl)triethoxysilane (APTES) and 20 µl of N,N-diisopropylethylamine (DIPEA) for a period of 2 h. After this time the APTES and DIPEA were removed and the treated surfaces were left to cure under a nitrogen atmosphere for 24 h. These aminosilinated surfaces were then reacted with a heterofunctional PEG linker (NHS-(PEG)n-maleimide [n = 4 or 12, Thermo Scientific]) by adding 15 µl of 250 mM PEG linker in DMSO to 1 ml of chloroform in which the surfaces were incubated for 1 h. After functionalization with the PEG linker, the surfaces were washed using chloroform, dried under nitrogen and held under PBS until required. To avoid hydrolysis of the maleimide groups, functionalized surfaces were used within 1 h of their preparation. When required, functionalised surfaces and AFM probes were incubated with protein (at a concentration of 1 mgml−1 in PBS which, is in excess with respect to maleimide groups on the surfaces) for 30 min and then washed with PBS.
All AFM measurements were conducted on an Asylum MFP-3D microscope using Si3N4 cantilevers with nominal spring constants of either 30 or 100 pNnm−1 (Bruker MLCT). For each cantilever used, the spring constant was determined using the thermal method [55],[56] via inbuilt fitting software. Retraction velocities of 200–8,000 nms−1 were employed for dynamic force spectroscopy analysis. For velocities between 200 and 5,000 nms−1 a PEG linker composed of 12 monomers was used, and for retraction velocities of 8,000 nms−1 a shorter PEG linker (four monomers) was used in order to increase the loading rate that could be applied. All experiments were conducted under PBS at 25°C. For retraction velocities less than 5,000 nms−1, 30 pNnm−1 nominal spring constant cantilevers were employed. At retraction velocities greater than 5,000 nms−1, 100 pNnm−1 nominal spring constant cantilevers (which have a smaller cross section) were used in order to reduce the hydrodynamic drag experienced by the cantilever, which becomes significant for the 30 pNnm−1 cantilevers at retraction velocities above 5,000 nms−1. A minimum of three separate functionalized AFM tips and surfaces were used for the collection of each dynamic force spectrum measured.
Size exclusion chromatography (SEC) was used to quantify the release of E9 DNase into solution over time from an E9 DNase:Im9 complex incubated in the presence of excess full length colicin E9. This procedure was used to measure kioff for both E9 (S3C):Im9 (S81C) derivatised with a PEG linker and E920–66 domains (see text) in complex with Im9, under conditions identical to those employed for the DFS experiments. E9 and Im9 were derivatised with methyl-(PEG)12-maleimide ((MM(PEG)12), Thermo Scientific) by incubation of the protein with a 20-fold molar excess of MM(PEG)12 overnight at room temperature in 25 mM Tris.HCl buffer, 1 mM MgCl2 (pH 7.5). Following this, derivatised protein was separated from un-labelled protein and excess MM(PEG)12 by size exclusion chromatography.
E9 DNase domains were first incubated with Im9 at a molar ratio of 1∶2 in order to form the E9 DNase:Im9 complex. The E9 DNase:Im9 complex was then purified from the excess Im9. A 25-µM solution of the E9 DNase:Im9 complex and 125-µM of full length colicin E9 in PBS buffer (pH 7.3), 0.01% (w/v) azide and a protease inhibitor cocktail (set III, EDTA free, Calbiochem) was then incubated for different lengths of time. Samples were removed at various times between 0 and 144 h and analysed via SEC. The intensity of the elution peak that corresponded to the free E9 DNase domain (competed from the complex by the addition of excess of full length colicin E9) was quantified as a function of time to calculate an apparent dissociation rate constant (kioff). An example dataset is shown in Figure S8.
E9 nuclease activity was assessed by monitoring the conversion of supercoiled DNA into other forms upon addition of E9 DNase domain (E9 (S108C) or E9 (S108C) derivatised with MM(PEG)12 as described above. DNA that was predominantly in the supercoiled conformation was isolated using a Hi-speed midi prep kit (Qiagen) at 4°C as described [57]. Nuclease activity was measured by addition of 30 nM E9 (final concentration) to purified DNA at a concentration of 50 µg/ml in 25 mM Tris.HCl buffer containing 1 mM MgCl2 (pH 7.5) in the presence or absence of 50 nM Im9. The reaction was arrested by adding 10 µl of this solution to 5 µl of solution containing 20 mM EDTA and agarose gel electrophoresis loading dye. Multiple time points were taken and the presence of supercoiled, linear, and open circular DNA was assessed by visualisation using agarose gel electrophoresis.
|
10.1371/journal.pcbi.1003271 | Dynamic Change of Global and Local Information Processing in Propofol-Induced Loss and Recovery of Consciousness | Whether unique to humans or not, consciousness is a central aspect of our experience of the world. The neural fingerprint of this experience, however, remains one of the least understood aspects of the human brain. In this paper we employ graph-theoretic measures and support vector machine classification to assess, in 12 healthy volunteers, the dynamic reconfiguration of functional connectivity during wakefulness, propofol-induced sedation and loss of consciousness, and the recovery of wakefulness. Our main findings, based on resting-state fMRI, are three-fold. First, we find that propofol-induced anesthesia does not bear differently on long-range versus short-range connections. Second, our multi-stage design dissociated an initial phase of thalamo-cortical and cortico-cortical hyperconnectivity, present during sedation, from a phase of cortico-cortical hypoconnectivity, apparent during loss of consciousness. Finally, we show that while clustering is increased during loss of consciousness, as recently suggested, it also remains significantly elevated during wakefulness recovery. Conversely, the characteristic path length of brain networks (i.e., the average functional distance between any two regions of the brain) appears significantly increased only during loss of consciousness, marking a decrease of global information-processing efficiency uniquely associated with unconsciousness. These findings suggest that propofol-induced loss of consciousness is mainly tied to cortico-cortical and not thalamo-cortical mechanisms, and that decreased efficiency of information flow is the main feature differentiating the conscious from the unconscious brain.
| One of the most elusive aspects of the human brain is the neural fingerprint of the subjective feeling of consciousness. While a growing body of experimental evidence is starting to address this issue, to date we are still hard pressed to answer even basic questions concerning the nature of consciousness in humans as well as other species. In the present study we follow a recent theoretical construct according to which the crucial factor underlying consciousness is the modality with which information is exchanged across different parts of the brain. In particular, we represent the brain as a network of regions exchanging information (as is typically done in a comparatively young branch of mathematics referred to as graph theory), and assess how different levels of consciousness induced by anesthetic agent affect the quality of information exchange across regions of the network. Overall, our findings show that what makes the state of propofol-induced loss of consciousness different from all other conditions (namely, wakefulness, light sedation, and consciousness recovery) is the fact that all regions of the brain appear to be functionally further apart, reducing the efficiency with which information can be exchanged across different parts of the network.
| Despite the centrality of consciousness to our experience, no agreement has yet emerged on which aspects of brain function underlie its presence, and what changes are connected to its disappearance in the healthy brain (e.g., during sleep) as well as in pathological conditions (e.g., coma). As a consequence, we are currently hard pressed to answer even basic questions concerning the presence, absence, degree and nature of the phenomenon of consciousness in humans and other species [1]. As experimental investigations into this domain have increased, a number of proposals have been put forth to characterize the neural fingerprint of consciousness. According to some views, the crucial feature underlying consciousness is the presence of specific patterns of activations, such as the presence of competing assembly of cells, or ‘neural coalitions’ [2], synchronization of neural activity in specific frequency bands [3], [4], or the level of spontaneous oscillatory activity, at fast frequencies, in the thalamo-cortical system [5]. According to other proposals, consciousness is related to the spread and reverberation of information across the neural system, and in particular within specific regions in parietal and frontal cortices [6], [7] – although the scope of this view is mostly relevant to the idea of conscious availability of content to a neural system, as compared to the more general “state of consciousness” of a neural system [8]. Finally, a recently proposed view [1], [9], stresses the importance of evaluating not the degree of correlation among different (often long-range) regions, but rather the degree of information present and the extent to which information is integrated across the nodes of a system.
In the present work we look at spontaneous low-frequency fluctuations in the functional magnetic resonance imaging (fMRI) signal [10], [11], to assess the relationship between different states of consciousness and basic principles of information processing (as captured by the blood oxygenation level dependent signal; i.e., BOLD). The analysis of spontaneous fluctuations of the BOLD signal has been fruitfully employed to explore consciousness-related changes in clusters of temporally coherent regions during sedation [12], [13], sleep [14]–[16], and in the pathological brain [17], [18]. In particular, associations within specific networks of regions have been found to be monotonically modulated by consciousness [19]–[21], consistent with some theoretical views [3]–[5]. This idea, however, clashes with reports of increased cross-regional correlation concurrent with decrease or loss of consciousness [22], [23], suggesting the importance of characterizing not just the strength but also the quality of information processing within a system [1], [24].
Following this idea, we employ previously collected resting-state fMRI data [21] to assess, in 12 healthy volunteers, the dynamic change of governing principles of brain organization during wakefulness (W), propofol-induced sedation (S) and loss of consciousness (LOC), as well as after consciousness recovery (R), a dynamic approach that has been recently advocated for [25]. In particular, we focus on the change of global and local topological metrics of information processing across conditions [26]–[28], a technique that has been successfully employed to characterize and model dynamics within physical [29], biological [30] and social systems [31], and that has been shown to capture specific aspects of brain organization in the maturing, healthy adult, and pathological brain [32]–[36]. A particularly appealing aspect of this technique in the context of studies of consciousness is the parallel between the measures it offers, focused on characterizing how information is exchanged and propagated through a network, and theories of consciousness that stress the centrality of how information is treated and integrated within the brain [1], [9].
As detailed below, we report three main findings. First, contrary to a recent report [25], we find that long- and short-range connections are not differentially affected by sedation. Second, employing a support vector machine (SVM) classifier, we dissociate the thalamo-cortical and cortico-cortical hyperconnectivity observed during sedation from the cortico-cortical hypoconnectivity observed during loss of consciousness. Finally, contrary to results in other species [37], we find significant global changes in the (functional) topological organization of the brain during sedation. However, we show that normalized clustering, the global metric that was previously reported to be sensitive to the loss of consciousness [25], remains significantly elevated also through post-sedation recovery of wakefulness. Conversely, we find that a strong decrease in efficiency of information distribution (defined as the inverse of the characteristic path length – see Materials and Methods) is the only unambiguous marker of propofol-induced loss of consciousness.
The average connectivity matrices and the frequency distribution of (average) correlations for each condition are shown in Figure 1 and Figure 2a, respectively. According to a two-sample Kolmogorov-Smirnov goodness-of-fit test, the distribution of positive and negative correlations are significantly different for all pairwise comparisons (; ; ; ; all ). In all four conditions about 80% of correlations were between 0 and 0.4. LOC, however, exhibited a leftwards shift of the distribution, as shown by the median correlation value of 0.11, as compared to 0.23, 0.22, and 0.19 for W, S, and R, respectively. Furthermore, 14% of correlations in the LOC condition were negative, as compared to about 2% in all other conditions, while only 6% were above 0.4, versus 17%, 14% and 11% for W, S, and R, respectively. To assess whether correlations between areas at different distances were unequally affected by the level of consciousness, we employed a repeated measures ANCOVA with one within-subjects variable (i.e., condition) with four levels (W, S, LOC, R), and inter-ROI distance as a covariate (with distance defined as the 3-dimensional Euclidean distance between the baricenter of each ROI; see Figure 2b) to predict correlation strength. As expected, we found a significant effect of condition (, ), indicating that correlation strength systematically varied across conditions. Specifically, W consistently exhibited the strongest average correlation level, across all bins, followed by S and R, while LOC consistently exhibited the weakest average correlation across all bins. We also found a significant effect of distance (, ), indicating that, as shown in Figure 2b, the average correlation strength decreased with distance. In addition to the two main effects, we also found a significant interaction between condition and distance (, ), indicating an uneven effect of condition on links of different length. However, when we followed up this significant interaction with a set of separate repeated measures ANOVAs (one per each bin) we found that it was entirely driven by the absence of a significant difference between W and S for the first 3 bins (out of 15; i.e., regions closer than 3.4 cm). With this exception, the effect of propofol was remarkably consistent at all other connection lengths (particularly with respect to the crucial condition – i.e., loss of consciousness – where no difference was found across connection length). Indeed, at all other bins the four conditions were found to be significantly different from each other, based on estimated marginal means and a Sidak correction for multiple comparisons. The observation of a small effect of distance on connection strength across levels` of sedation is also consistent with the extremely low effect size observed for the interaction between condition and distance in the overall ANOVA (), and strengthens the idea that, overall, connection size had a minimal effect on correlation strength – something that is immediately clear from Figure 2b.
Results for the classification of brain networks (i.e., correlation matrices) are reported in Table 1 and Figure 3. At a global level, the SVM algorithm classified successfully states of wakefulness (W & R) versus states of sedation (S & LOC) with high accuracy, sensitivity and specificity (all above 83.33%; ). The same level of classification was also achieved when comparing contiguous brain states (namely, W vs. S; S vs. LOC; and LOC vs. R; see Table 1 for a detailed report of accuracy, specificity, sensitivity and significance for each). Conversely, wakefulness (W) and wakefulness recovery (R) could not be successfully distinguished from each other (). (For completeness the two remaining classifications, namely W vs. LOC and S vs. R, are reported in Figure S1.)
At the local level, accurate classification of each transition relied on different sets of edges within each brain graph (see Tables S1 and S2 for full details). In particular, as depicted in Figure 3b, and more in detail in Figure 4a, the edges mostly contributing to correctly classifying S versus W included positive cortico-cortical (54.8%) and thalamo-cortical (40.9%) connections, as well as a minority of cerebello-cortical (0.5%) and striato-cortical (3.8%) connections. Conversely, as depicted in Figures 3c and 4b, the distribution of connections correctly classifying LOC, as compared to S, mostly included negative cortico-cortical connections (82.5%), as well as a minority of positive cortico-cortical (9.9%), thalamo-cortical (3.5%), cerebello-cortical (2.9%) and thalamo-striatal (1.2%) connections. Notably, when tested statistically, the allocation of classifying edges for these two transitions are significantly different (, ). Finally, as shown in Figures 3d and 4c, as compared to LOC, classification of R was almost entirely based on the re-emergence of positive cortico-cortical connections (98.4%) as well as a small minority of cerebello-cortical connections (1.6%).
In this study we assessed propofol-induced changes in patterns of connectivity, as well as in global and local governing principles of brain organization, during wakefulness, sedation, loss of consciousness, and wakefulness recovery. Our results contribute to a growing literature addressing the topological organization of the human brain [26], [38], the changes in functional architecture accompanying the loss of consciousness [16], [25], [37], as well as a specific hypothesis concerning the role of different subsystems in loss of consciousness [40], [41].
Overall, our main findings are three-fold. First, despite the frequently voiced idea that long-range connections play a key role in anesthesia-induced unconsciousness [40], we fail to find a substantial asymmetric decrease in cross-region correlation as a function of inter-regional distance. Average connectivity strength decreased monotonically with distance in approximately the same manner across conditions (with the sole exception of extremely short connections, below 34 mm, but only during the initial phase of sedation, and not during loss of consciousness). This finding runs counter to a recent report demonstrating an uneven effect of propofol-induced unconsciousness on short-range (i.e., ) versus long-range (i.e., ) connections [25]. The only effect we detected concerned much shorter connections (i.e., ), and was only found for the initial period of sedation, and not for the period of loss of consciousness. Whether the different result is to be attributed to methodological asymmetries (e.g., 2-timepoint versus 4-timepoint paradigms, the binning procedure, the use of different ROIs parcellation schemes) or to un-modelled third factors remains to be determined.
The second central aspect of our results directly addresses the discussion concerning the role of thalamo-cortical versus cortico-cortical circuits in propofol-induced unconsciousness [40], [41]. In particular, our SVM classification isolated increased thalamo-cortical and cortico-cortical synchronization as being maximally informative in the wakefulness versus sedation classification, suggesting a prominent role of this circuit in the initial stages of sedation, before the onset of unconsciousness. Conversely, correct classification of the state of loss of consciousness, as compared to sedation, overwhelmingly relied on negative cortico-cortical correlations. These findings support the view that propofol-induced loss of consciousness is more closely linked to cortico-cortical mechanisms rather than thalamo-cortical ones, as also suggested in a recent EEG effective connectivity study [41]. It is important to point out that our SVM classification is entirely based on the full matrix of ROI-to-ROI correlations and is, therefore, entirely data driven and blind to the existence of particular neural circuits or opposing hypothesis concerning their role in propofol-induced loss of consciousness. The observed major role of negative cortico-cortical connectivity in propofol-induced unconsciousness should be differentiated, however, from studies on pathological loss of consciousness in severe brain injury where post-mortem [42] and in-vivo [43] evidence highlights the role of thalamus in loss and recovery of consciousness [44], [45]. While further studies will have to directly address the issue, our findings are consistent with the suggestion that thalamus may be a necessary but not sufficient component in maintaining consciousness [41] consistent with the view that thalamic lesions might induce unconsciousness after severe brain injury by virtue of disconnecting an otherwise functioning cortex [46], [47].
The third result of our study concerns changes in governing principles of information processing during loss and recovery of consciousness. Contrary to a recent study in other species [37], we do find significant changes in global topological measures across levels of consciousness. Consistent with a previous report [25], we find that loss of consciousness is marked by an increase in normalized clustering (), which measures the ‘cliquishness’ of brain regions, potentially indicating an increase in localized processing and thus a decrease of information integration across the brain. Our multi-stage design, however reveals that clustering remains significantly elevated (as compared to initial wakefulness and sedation) during post-anesthesia wakefulness recovery. This result shows that while it is true that clustering increases once consciousness is lost, it is not a sufficient marker of consciousness, something that the two-point design (i.e., initial wakefulness versus loss of consciousness) in Schröter et al. [25] could not reveal. On the other hand, we find that the normalized characteristic path length () is significantly increased only during loss of consciousness, suggesting that during unconsciousness the efficiency of information distribution within the network is reduced (a finding that is consistent with a very recent study on loss of consciousness in sleep [16]). Whether this state of increased “functional distance” between regions is causal or consequent to propofol-induced loss of consciousness will have to be addressed in future research. As previously reported, the small-world architecture of brain networks () persisted (and in fact increased) in loss of consciousness [25], confirming the robustness of this core principle of organization of biological networks despite profound state changes [32]. Mirroring , however, small-world architecture also remained significantly elevated during wakefulness recovery. Although much weaker, a similar effect of condition was also uncovered for normalized modularity (). Finally, we remark that the presence of different results observed in the two propofol conditions (sedation and loss of consciousness) and, importantly, consciousness recovery, is consistent with the view that changes in global brain topology observed here and elsewhere [25], [37] are not simply due to drug exposure, but rather reflect brain state changes relating to the loss of consciousness, supporting a previously expressed view [25].
Beyond the global reorganization of brain topology, we also observed changes in local network topology. With respect to nodal strength, selected frontal and parietal regions along the midline, as well some lateral and opercular ROIs, appeared to be modulated by changes in the level of consciousness. In particular, regions in medial frontal and parietal cortices, along with occipital and lateral parietal, exhibited less nodal strength during sedation and loss of consciousness. Other regions, on the other hand, in temporal cortex especially, but also in dorsal and ventro-medial prefrontal cortex, exhibited the reverse pattern. Mirroring the result for , local efficiency appeared to be modulated mostly across midline parietal and prefrontal regions. Overall, this pattern of reorganization of local network topology is consistent with the view that propofol affects specific hubs central to normal/wakeful connectivity [48] which are also known to play a critical role in consciousness [49]–[51] and self-consciousness [52].
Taken together, our findings support the idea that (propofol-induced) loss of consciousness correlates with a change in the quality of information processing, and not only a change in the strength of connectivity across regions [1], [24]. In particular, dynamic reconfiguration of thalamo-cortical and cortico-cortical connections, and contemporaneous decrease of efficiency and increased local processing might affect the degree by which information can be effectively integrated across the brain [9].
In terms of theories of consciousness, these findings can be interpreted as making two contributions. First, the significant increase of cortico-cortical decorrelations during loss of consciousness is coherent with views of consciousness stressing the role of coherent reverberation and spread of neural activity [6], [7], particularly within fronto-parietal regions [5]. (We point out that, as shown in Tables S1 and S2, all fronto-parietal connections driving the correct classification of loss of consciousness, compared to sedation, are negative.) Second, our graph theoretic analysis further indicates that, in terms of network information processing, propofol-induced loss of consciousness is marked by a specific change in the quality of information exchange (i.e., decreased efficiency), consistent with the view that the specific modality with which information is exchanged within brain networks is crucial to the maintenance of a state of consciousness [1], [9].
Finally, it is important to stress that many of the methodological limitations expressed elsewhere concerning the interpretation of the blood oxygenation level dependent signal, as well as the current challenges tied to applying graph theory to brain measures previously discussed [25], [28], [32], [34], [37], also apply to our study. In particular, with respect to the implementation of graph-theory measures in neuroscience, several issues are still in search of resolution. Here, we believe it is important to stress five methodological considerations. First, as we note in the Materials and Methods section, most topological measures require thresholding of adjacency matrices, a procedure that presently lacks a defined standard approach (e.g., how many and which thresholds to employ) and might have important effects on the derived metrics [53]. While the real resolution of the issue will likely include measures that can be applied to fully connected matrices [28], we stress that our results were robust to the choice of threshold. Second, in contrast to some previous studies [25], we made use of weighted measures, a difference that might explain the divergence of results. For instance, we note that the observed between-group differences in our study were most pronounced at the lowest density thresholds (corresponding to least sparse networks), in contrast to many binary brain-network studies, in which between-group differences are most pronounced at the highest density thresholds (corresponding to most sparse networks) [54]. Many binary-network studies discard as many as 90–95% of all possible connections to elucidate the observed between-group differences [53] and it is likely that these more radical thresholding approaches are associated with substantial loss of connectivity information [55]. High thresholds are needed in binary studies because when weak and strong links surviving thresholding are equally assigned a value of 1, measures based on path length become susceptible to the creation of spurious long-distance short-cuts, which might obscure the architecture of strong connections and, thereby, important across-group differences [38]. It is therefore possible that the use of binary matrices in previous studies might have obscured the differences in characteristic path length that we have observed. Consistent with our findings, a recent study in the domain of sleep also uncovered loss of efficiency during unconsciousness [16]. Third, as discussed in the Materials and Methods section, because of the known effects of motion on graph theoretic analysis [56], [57], our sample was reduced to 12 volunteers. Although this sample size is within the boundaries of previous work on this same topic (e.g., N = 11 in [25], N = 20 in [37]) it does fall at the low end of the spectrum. Therefore, even though our analyses leverage on a statistically more powerful 4-point repeated measures design (as compared to the more typical two groups across-subjects comparison and two-points within subject design), future studies will have to confirm their generality. Nonetheless, we do stress that the effect-size analysis, which is robust to small samples, shows that our effects are of large magnitude, and that our results are consistent with previous reports [16]. Fourth, while we adopt the presently accepted mainstream interpretation of characteristic path length and global efficiency as measures of functional integration, we acknowledge that these interpretations have not been directly validated and are less trivial to make in networks where edges represent correlations and hence do not necessarily represent causal interactions or information flow [28]. Finally, it is important to stress that a recognized source of variance across results is the choice of ROIs [58], [59]. In particular, we employed more ROIs than in similar previous studies [25], [37], hence it is possible that some of the reported differences are due to the less granular parcellation schemes previously employed. Similarly, it is also possible that, if we had used an even greater number of ROIs, or based our networks on a voxel-wise analysis, results would have differed. However, it has been shown that simple binary decisions concerning the presence of certain network organizational parameters (e.g., small-worldness) are robust across different parcellation granularity [58]–[60]. Consistent with this finding, a recent study evaluating network properties during sleep reported a loss of efficiency during loss of consciousness that paralleles our own findings, despite the fact that their networks featured more than 3,700 nodes [16]. It should be stressed, however, that high granularity parcellations might yield quantitatively very different estimates of network properties, as compared to low granularity parcellations, and might allow topological features to be displayed more prominently [58], [59]. There is, however, an important conceptual difference that separates region-based networks from voxel-based networks [34], [61]. In our report, as in all region-based analyses of brain connectivity, network locality is conceived at a specific scale, determined by the coarseness of the employed parcellation. Hence, when we investigate local network properties, we are investigating topological features calculated over proximal brain regions. Conversely, voxel-wise networks assess locality within regions of the brain, an approach which has the potential advantage of capturing differences across regions of the brain in within- and between-connectivity [34], [61]. In this sense, region-based network analyses might be biased towards highlighting the properties of regions with widely distributed connections at a coarse scale, predominant in heteromodal association areas [62], and blind to local hierarchical connections more predominant in sensory cortical areas [63]. Voxel-based network analysis, instead, allow for examining inter-regional as well as intra-regional connectivity [34]. Nonetheless voxelwise parcellations might however pose conceptual difficulties with respect to computing global network properties because grid-like subdivisions do not generally respect boundaries or sizes of heterogeneous functional areas, an approach that might lead to mischaracterization of brain network function [64]. In conclusion, in interpreting our results (as any region-based network analysis with comparably sized, or larger, ROIs) it is thus important to keep in mind that our statements concerning changes in local topological features are intended as network-local, and do not necessarily reflect local changes at the brain physical level.
In sum, our findings show that changes in the level of consciousness induced by propofol affect basic organization principles and dynamics of information processing across the whole brain as well as within specific regions known to be involved in consciousness. In particular, we find that propofol-induced loss of consciousness is mostly associated with cortico-cortical mechanisms, as opposed to thalamo-cortical ones, and with a substantial decrease in the efficiency of information flow within the network. Future research will have to assess whether different anesthetic agents and pathology (e.g., brain trauma, seizures) induce loss of consciousness via the same mechanisms.
The present report constitutes an entirely novel analysis of data that has been previously described with different methods [21]. Before detailing our analysis approach, based on graph-theoretic measures, we briefly describe the population, manipulation and data acquisition methods.
The study was approved by the Ethics Committee of the Medical School of the University of Liège (University Hospital, Liège, Belgium).
Data analysis was carried out in three stages: initial preprocessing, support vector machine (SVM) matrix classification, and computation of global and local graph-theoretic measures.
|
10.1371/journal.pgen.1004512 | The Translational Regulators GCN-1 and ABCF-3 Act Together to Promote Apoptosis in C. elegans | The proper regulation of apoptosis requires precise spatial and temporal control of gene expression. While the transcriptional and translational activation of pro-apoptotic genes is known to be crucial to triggering apoptosis, how different mechanisms cooperate to drive apoptosis is largely unexplored. Here we report that pro-apoptotic transcriptional and translational regulators act in distinct pathways to promote programmed cell death. We show that the evolutionarily conserved C. elegans translational regulators GCN-1 and ABCF-3 contribute to promoting the deaths of most somatic cells during development. GCN-1 and ABCF-3 are not obviously involved in the physiological germ-cell deaths that occur during oocyte maturation. By striking contrast, these proteins play an essential role in the deaths of germ cells in response to ionizing irradiation. GCN-1 and ABCF-3 are similarly co-expressed in many somatic and germ cells and physically interact in vivo, suggesting that GCN-1 and ABCF-3 function as members of a protein complex. GCN-1 and ABCF-3 are required for the basal level of phosphorylation of eukaryotic initiation factor 2α (eIF2α), an evolutionarily conserved regulator of mRNA translation. The S. cerevisiae homologs of GCN-1 and ABCF-3, which are known to control eIF2α phosphorylation, can substitute for the worm proteins in promoting somatic cell deaths in C. elegans. We conclude that GCN-1 and ABCF-3 likely control translational initiation in C. elegans. GCN-1 and ABCF-3 act independently of the anti-apoptotic BCL-2 homolog CED-9 and of transcriptional regulators that upregulate the pro-apoptotic BH3-only gene egl-1. Our results suggest that GCN-1 and ABCF-3 function in a pathway distinct from the canonical CED-9-regulated cell-death execution pathway. We propose that the translational regulators GCN-1 and ABCF-3 maternally contribute to general apoptosis in C. elegans via a novel pathway and that the function of GCN-1 and ABCF-3 in apoptosis might be evolutionarily conserved.
| Apoptosis, also referred to as programmed cell death, is a crucial cellular process that eliminates unwanted cells during animal development and tissue homeostasis. Abnormal regulation of apoptosis can cause developmental defects and a variety of other human disorders, including cancer, neurodegenerative diseases and autoimmune diseases. Therefore, it is important to identify regulatory mechanisms that control apoptosis. Previous studies have demonstrated that the transcriptional induction of apoptotic genes can be crucial to initiating an apoptotic program. Less is known about translational controls of apoptosis. Here we report that the evolutionarily conserved C. elegans translational regulators GCN-1 and ABCF-3 promote apoptosis generally and act independently of the anti-apoptotic BCL-2 homolog CED-9. GCN-1 and ABCF-3 physically interact and maintain the phosphorylation level of eukaryotic initiation factor 2α, suggesting that GCN-1 and ABCF-3 act together to regulate the initiation of translation. We propose that the translational regulators GCN-1 and ABCF-3 maternally contribute to the proper execution of the apoptotic program.
| Apoptosis is a naturally occurring process that eliminates unwanted cells during development and maintains tissue homeostasis [1], [2]. For example, apoptosis removes most larval tissues of insects during metamorphosis, sculpts the future inner ear in chicks, eliminates the interdigital web in mammals and shapes the endocardial cushion into valves and septa to generate the four-chamber architecture of the mammalian heart [1], [2]. Apoptosis also culls nearly 80% of oocytes prior to birth in humans and eliminates cells that receive insufficient cell-survival signals to maintain homeostasis [1]. The improper regulation of an apoptotic program can result in either too much or too little cell death, leading to developmental abnormalities and a wide variety of human disorders, such as cancer, neurodegenerative diseases, autoimmune diseases and developmental disorders [3], [4]. It is important to identify mechanisms that regulate apoptosis to understand both animal development and human disorders caused by the dysregulation of apoptosis.
The precise spatial and temporal expression of regulators of apoptosis is known to be crucial for initiating the apoptotic cell-killing program during development and in response to environmental stresses, including ionizing radiation, temperature change, nutrient limitation, oxidative stress and viral infection [1], [2]. Many examples of the transcriptional control of apoptosis have been described. For example, in mammals the genes that encode the pro-apoptotic BCL-2 family member BAX, the BH3-only proteins NOXA, PUMA and BID, the apoptotic protease-activating factor-1 APAF-1 and the death receptor 5 DR5 protein are transcriptionally upregulated by the tumor suppressor p53 transcription factor in response to DNA damage or to the induced expression of p53 [5]–[11], resulting in an induction of apoptosis. The Drosophila apoptotic activator gene reaper is upregulated by multiple transcriptional regulators, including Hox transcription factors, nuclear hormone receptors, AP-1, Polycomb, p53, and histone-modifying enzymes, to promote the morphogenesis of segment boundaries, metamorphosis, and DNA damage responses [1]. In C. elegans, the transcription of the pro-apoptotic BH3-only gene egl-1 is directly regulated in a cell-specific manner by transcription factors that include the Hox family proteins MAB-5, CEH-20, LIN-39 and CEH-34, the E2F protein EFL-3, the Snail family zinc finger protein CES-1, the Gli family transcription factor TRA-1, and the basic helix-loop-helix proteins HLH-2 and HLH-3 [12]–[15]. The caspase gene ced-3 is also upregulated by the Hox transcription factor PAL-1 in the tail spike cell before its death [16]. Recently, we showed that the Sp1 transcription factor SPTF-3 directly drives the transcription of both the pro-apoptotic BH3-only gene egl-1, which mediates a caspase-dependent apoptotic pathway, and the AMPK-related gene pig-1, which mediates a caspase-independent apoptotic pathway [17]. The transcriptional regulation of apoptotic genes clearly plays a crucial role in determining whether specific cells live or die during development.
Translational control is also important for the apoptotic process. In mammals, expression of the pro-apoptotic protein APAF-1 and the anti-apoptotic protein X-chromosome-linked inhibitor of apoptosis XIAP are regulated at the translational level by internal ribosome-entry sites (IRES) [18]. Exposure of cultured mammalian cells to etoposide or UV light induces APAF-1 expression via IRES-mediated translation, resulting in the activation of the caspase-dependent apoptotic program [19]. The protein level of XIAP is increased via IRES-mediated translation under stress conditions, such as serum starvation [20]. However, the specific translational regulators involved in IRES-mediated translation of APAF-1 and XIAP are unknown. In C. elegans, the RNA-binding protein GLD-1, which is highly expressed in the transition zone and early pachytene regions of the hermaphrodite gonad, inhibits translation of the mRNA of the p53 homolog cep-1 by directly binding to the cep-1 3′ UTR, thereby preventing cep-1-dependent apoptosis in response to DNA damage [21]. Translational initiation factors have also been reported to be involved in the control of apoptosis in C. elegans. For example, RNAi knockdown of the C. elegans eukaryotic initiation factor-4G IFG-1 induces CED-4 expression in the gonad and increases the frequency of germ-cell death [22], [23]. The eukaryotic initiation factor-3 subunit-k eIF-3.K is partially required for the deaths of somatic cells and acts through the caspase CED-3 to promote those cell deaths [24]. Although many studies have shown that both transcriptional and translational regulation of apoptotic genes is crucial for controlling apoptotic programs, how transcriptional and translational mechanisms are coordinated to promote apoptosis remains elusive.
Here we show that the maternally-contributed translational regulators GCN-1 and ABCF-3 act together to promote the cell deaths of possibly all somatic cells and of germ cells in response to ionizing radiation in a pathway distinct from the BCL-2 homolog CED-9-regulated canonical cell-death execution pathway of C. elegans. GCN-1 and ABCF-3 are required to maintain the basal level of phosphorylation of eukaryotic initiation factor 2 (eIF2α). The functions of GCN-1 and ABCF-3 in the promotion of programmed cell death are evolutionarily conserved between C. elegans and Saccharomyces cerevisiae. We show that GCN-1 and ABCF-3 cooperate with the transcriptional regulators CEH-34, EYA-1 and SPTF-3 and the protein kinase PIG-1 to promote the death of a specific somatic cell, the sister cell of the pharyngeal M4 motor neuron. We propose that the evolutionarily-conserved translational regulators GCN-1 and ABCF-3 contribute to apoptosis in general.
The C. elegans pharyngeal M4 motor neuron is generated during embryonic development and survives to regulate pharyngeal muscle contraction in feeding behavior, whereas the M4 sister cell dies by programmed cell death soon after its generation (Figure 1A) [25], [26]. We created a Pceh-28::gfp reporter transgene that expresses GFP specifically in the M4 neuron of wild-type animals and in both the M4 neuron and the surviving M4 sister of ced-3 caspase mutants defective in programmed cell death (Figure 1B–1C). This reporter allowed us to easily identify mutants with a defect in M4 sister cell death. [15]. Using this reporter, we performed a genetic screen for mutations that cause a defect in M4 sister cell death. Among our isolates were two non-allelic mutations, n4827 and n4927, that caused M4 sister survival in 12% of n4827 mutants and 13% of n4927 mutants (Figure 1D–1F).
We mapped n4827 to a 175 kb interval of chromosome III containing 18 predicted genes (Figure S1A). We used whole-genome sequencing to identify four strain-specific unique homozygous mutations within this interval in n4827 animals (Figure S1A) [27]. Of the four mutations, only one was exonic. This mutation was located in the third exon of gcn-1, which encodes a homolog of the S. cerevisiae Gcn1p protein. The n4827 mutation is predicted to change the tryptophan 164 codon to an opal stop codon, generating a small truncated protein (Figure 1G). A deletion mutation of gcn-1, nc40Δ, phenocopied the n4827 mutation [28]: 11% of gcn-1(nc40Δ) mutants and 12% of n4827 mutants had a surviving M4 sister, respectively (Figure 1F). The cell-death defect of n4827 mutants was partially rescued by a transgene that express gcn-1 cDNA under the control of the gcn-1 promoter (Figure 1F). These results indicate that n4827 is likely a null allele of gcn-1 and that loss of gcn-1 function causes a defect in M4 sister cell death.
We mapped n4927 to a 5.3 Mb interval of chromosome III (Figure S1B). This interval contains the gene abcf-3, which encodes a homolog of the S. cerevisiae Gcn20p protein [29]. Gcn20p physically interacts with Gcn1p, the S. cerevisiae homolog of GCN-1 [30]. We determined the sequence of abcf-3 in n4927 animals and identified a mutation that changes the arginine 206 codon to an opal stop codon (Figure 1G). A deletion mutation of abcf-3, ok2237Δ, that removes most of the abcf-3 coding region phenocopied the n4927 mutation: 13% of abcf-3(ok2237Δ) mutants and 13% of n4927 mutants had a surviving M4 sister (Figure 1F). Furthermore, the cell-death defect of n4927 mutants was completely rescued by a transgene carrying only the abcf-3 genomic locus. We concluded that n4927 is likely a null allele of abcf-3 and that loss of abcf-3 function causes a defect in M4 sister cell death.
abcf-3 encodes an AAA ATPase protein with two AAA domains (Figure 1G). In many proteins AAA domains have ATPase activity. To determine whether ATPase activity is important for ABCF-3 to promote M4 sister cell death, we generated abcf-3 transgenes carrying mutations that presumably inactivate the ATPase activity of each AAA domain by altering the lysine residues known to be catalytically essential for other AAA ATPases [31]. A wild-type abcf-3 transgene as well as mutant abcf-3 transgenes that changed lysine 217 of the first AAA domain to methionine [abcf-3 (K217M)], lysine 536 of the second AAA domain to methionine [abcf-3 (K536M)] or both lysine residues [abcf-3 (K217M, K536M)] completely rescued the defect in M4 sister cell death of abcf-3(n4927) mutants (Figure 1F). These results support the idea that the ATPase activity of ABCF-3 is dispensable for M4 sister cell death. This result is consistent with studies of S. cerevisiae Gcn20p, the homolog of C. elegans ABCF-3. Gcn20p that lacks the ATPase activities of both AAA domains because of mutations in conserved glycine residues (Gly371 and Gly654) or because of the deletion of two AAA domains still retains Gcn20p function comparable to that of wild-type Gcn20p [32].
GCN-1 and ABCF-3 are evolutionarily conserved among S. cerevisiae, C. elegans and humans (Figure 1H, Figure S2 and S3). Expression of S. cerevisiae GCN1, the homolog of C. elegans gcn-1, and GCN20, the homolog of C. elegans abcf-3, under the control of the abcf-3 promoter rescued the defect in M4 sister cell death of C. elegans gcn-1 and abcf-3 mutants, respectively, indicating that S. cerevisiae GCN1 and GCN20 are functional homologs of C. elegans gcn-1 and abcf-3, respectively (Figure 1F).
S. cerevisiae Gcn1p has a domain (amino acids 1350–2152) similar to that of translation elongation factor 3 (EF3). The EF3-like domain is highly conserved among species (Figure 1H and Figure S2) and is necessary and sufficient for binding to Gcn20p [32]. We therefore tested whether C. elegans GCN-1 can physically interact with ABCF-3 using the yeast two-hybrid assay (Figure 2A). Full-length GCN-1 (1–2634) interacted with full-length ABCF-3 (1–712). To identify the protein domains important for GCN-1 to bind to ABCF-3, we generated a series of deletion constructs of GCN-1 and assayed each for ABCF-3-binding activity using the yeast two-hybrid assay. GCN-1 fragments not containing entire the EF3 domain (1–1760, 1–880, 880–1760 and 1760–2634) or containing only the EF3 domain (1350–2150) failed to bind ABCF-3, whereas GCN-1 fragments containing the EF-3 domain and surrounding regions (880–2634) bound ABCF-3. These results suggest that GCN-1 physically interacts with ABCF-3 but that unlike in yeast the EF3-like domain is not sufficient for GCN-1 to bind to ABCF-3.
We also defined the domains of ABCF-3 important for ABCF-3 to bind to GCN-1 (Figure 2A). ABCF-3 fragments lacking the N-terminal region (202–712, 512–712 and 202–512) failed to bind GCN-1, whereas ABCF-3 fragments containing the N-terminal region (1–712, 1–512 and 1–202) bound GCN-1, suggesting that the N-terminal portion of ABCF-3 (which does not include the first AAA domain) is necessary and sufficient for binding to GCN-1, just as the N-terminal region of S. cerevisiae Gcn20p is necessary and sufficient for binding to Gcn1p, the S. cerevisiae homolog of GCN-1.
To determine whether GCN-1 and ABCF-3 interact in vivo, we generated antibodies against GCN-1 and ABCF-3 and performed co-immunoprecipitation experiments. We first tested whether these antibodies specifically recognize GCN-1 or ABCF-3 protein using western blot analysis. The antibodies against GCN-1 or ABCF-3 recognized proteins of the sizes predicted for the GCN-1 or ABCF-3 proteins in wild-type animals but not in gcn-1(n4827) or abcf-3(n4927) animals, respectively, confirming the specificity of these antibodies (Figure 2C). Then we tested whether GCN-1 could be co-immunoprecipitated with ABCF-3. Whole-protein extracts from wild-type animals were subjected to immunoprecipitation using an anti-ABCF-3 antibody (or normal IgG as a control), and then immunocomplexes were analyzed by western blotting using antibodies against ABCF-3 or GCN-1. Both ABCF-3 and GCN-1 were recovered in an immunocomplex purified with the anti-ABCF-3 antibody, whereas neither ABCF-3 nor GCN-1 was recovered in an immunocomplex purified with normal IgG (Figure 2B). We conclude that GCN-1 and ABCF-3 are present in the same protein complex in vivo.
Since GCN-1 and ABCF-3 form a complex in vivo, we suspected that deletion of either protein might affect the stability of the other protein [29], [33]. To test this hypothesis, we examined the levels of GCN-1 and ABCF-3 proteins by western blot analyses of whole-protein extracts prepared from wild-type, gcn-1(n4827) and abcf-3(n4927) animals using antibodies against ABCF-3 or GCN-1. The steady-state level of ABCF-3 protein was decreased in gcn-1(n4827) animals by 3.6 fold compared to that of wild-type animals. Similarly, the steady-state level of GCN-1 protein was decreased in abcf-3(n4927) animals by 4.4 fold (Figure 2C). These results suggest that a lack of ABCF-3 or GCN-1 protein affects the stability of the other protein and support our conclusion that GCN-1 and ABCF-3 are in a protein complex together in vivo.
If GCN-1 and ABCF-3 physically interact in vivo to promote M4 sister cell death, GCN-1 and ABCF-3 should act together in the same pathway. Since gcn-1(n4827) and abcf-3(n4927) are likely null mutations, the gcn-1(n4827) mutation would not enhance the M4 sister cell-death defect of abcf-3(n4927) mutants if gcn-1 and abcf-3 function in the same process or pathway. Indeed, we observed no enhancement of the M4 sister cell-death defect of gcn-1(n4827) abcf-3(n4927) double mutants compared to that of either single mutant: there was 13% M4 sister survival in gcn-1(n4827) abcf-3(n4927) double mutants, 12% M4 sister survival in gcn-1(n4827) mutants and 13% M4 sister survival in abcf-3(n4927) mutants (Figure 1F). We conclude that gcn-1 and abcf-3 function together in the same process of pathway to promote M4 sister cell death, consistent with our finding that GCN-1 and ABCF-3 physically interact in vivo.
In S. cerevisiae, Gcn1p and Gcn20p are required for the efficient phosphorylation of eukaryotic initiation factor 2 (eIF2α) under both normal conditions and conditions of amino-acid starvation [29], [30]. Gcn1p and Gcn20p form a protein complex that activates the serine-threonine protein kinase Gcn2p, which then phosphorylates an evolutionarily conserved serine residue of eIF2α. The amino acid sequences surrounding the eIF2α phosphorylation site are identical in S. cerevisiae, C. elegans and humans, suggesting a conserved regulatory mechanism of eIF2α [28]. We tested whether GCN-1 and ABCF-3 promote the phosphorylation of eIF2α in C. elegans using an antibody that specifically recognizes eIF2α that is phosphorylated at serine 49 (P-eIF2α). From wild-type animals cultivated under normal physiological conditions, a single band of eIF2α was detected in western blotting analyses using either the anti-P-eIF2α antibody or an antibody that recognized total eIF2α (Figure 2D). In gcn-1(n4827) and abcf-3(n4927) mutants, the phosphorylation levels of eIF2α in physiological conditions were 52% and 54% of the levels in wild-type animals, respectively (Figure 2D and 2E). We conclude that gcn-1 and abcf-3 are required to maintain the steady-state level of the phosphorylation of eIF2α.
The regulation of phosphorylation of eIF2α plays an essential role in the initiation of translation. We therefore directly tested whether gcn-1 and abcf-3 affect gene expression at the translational level. Since gcn-1 and abcf-3 are highly expressed in the gonads at the fourth larval stage, maternally contribute to the death of the M4 sister and affect most programmed cell deaths (see below, Figure 3D and 3I, Table 1 and Table S5), we isolated both wild-type animals and gcn-1 and abcf-3 mutants at the fourth larval stage and performed mRNA-seq and ribosome profiling (Ribo-seq) to generate quantitative genome-wide information concerning mRNA abundance and the locations of mRNAs occupied by ribosomes [34]. Parallel analyses of data from Ribo-seq and mRNA-seq studies allowed us to distinguish differences in mRNA abundance from differences in translational control and to generate a quantitative and comprehensive list of genes the expression of which is likely regulated by gcn-1 and abcf-3 at the translational level. Loss of gcn-1 or abcf-3 function affected the expression of a large number of genes at either the transcriptional or translational level or at both (Figure S4A and S4B and Table S1). Since GCN-1 and ABCF-3 very likely function in translational control, their effects on transcript levels are likely indirect. Changes in gene expression compared to wild-type animals were similar between gcn-1 and abcf-3 mutants, supporting our conclusion that gcn-1 and abcf-3 act together (Figure S4C and S4D). The expression of 464 genes or 217 genes changed in both gcn-1 and abcf-3 mutants compared to wild-type animals at least two-fold (p<0.1) in mRNA-seq or Ribo-seq analyses, respectively (Figure S4E and S4F). Of the 217 genes altered in translational expression, 98 genes showed no alterations in mRNA levels using our standards of a two-fold change and p<0.1 (Table S2 and Table S3). These genes are candidates for being directly regulated by both gcn-1 and abcf-3 translationally. These results suggest that gcn-1 and abcf-3 function together in the translational control of many genes.
In mammals, eIF2α phosphorylation is mediated by at least four different protein kinases: PKR-like endoplasmic reticulum kinase (PERK), general control non-derepresessible-2 (GCN2), double-stranded RNA-activated protein kinase (PKR) and heme-regulated inhibitor kinase (HRI); each of these kinases is activated by a distinct stress signal [18]. These kinases share homology in their kinase catalytic domains, but their effector domains are distinct and are subject to different regulatory mechanisms. Homologs of genes encoding two of these protein kinases exist in the C. elegans genome: the PERK homolog PEK-1 and the GCN2 homolog GCN-2. Y38E10A.8 has a kinase domain similar to that of mammalian eIF2α kinases but does not have an obvious homolog. We tested whether these three protein kinases are required for the programmed cell death of the M4 sister. Neither single mutants of each kinase gene nor the triple mutant was defective in M4 sister cell death (Table S4), suggesting that one or more unidentified protein kinase(s) regulated by GCN-1 and ABCF-3 are responsible for phosphorylating eIF2α in the regulation of M4 sister cell death. Alternatively, it is possible GCN-1 and ABCF-3 promote M4 sister cell death through one or more targets other than eIF2α.
To determine the expression patterns of gcn-1 and abcf-3, we generated transgenes expressing a reporter GFP under the control of the endogenous gcn-1 or abcf-3 promoter. Both gcn-1 and abcf-3 were expressed in most cells during all stages of development. We observed gcn-1 and abcf-3 expression in head neurons, hypodermal cells, intestinal cells, body wall muscles, and pharyngeal neurons, including the M4 neuron (Figure 3A–3C and 3F–3H). We also used the technique of fluorescence in situ hybridization (FISH) with a level of sensitivity sufficient to detect single mRNA molecules [35] to observe endogenous gcn-1 and abcf-3 transcripts. Consistent with the expression of the GFP reporter transgenes, gcn-1 and abcf-3 mRNAs were observed in most somatic cells. In addition, gcn-1 and abcf-3 mRNAs were abundant in the germ cells in the hermaphrodite gonad (Figure 3D, 3E, 3I and 3J). The similar expression patterns of gcn-1 and abcf-3 are consistent with our observations that GCN-1 and ABCF-3 physically interact and act together to promote the death of the M4 sister.
Since gcn-1 and abcf-3 are ubiquitously expressed and required to broadly maintain the basal level of phosphorylation of eIF2α, we tested whether gcn-1 and abcf-3 might be involved in other biological processes. We did not observe abnormalities in the morphologies of the hermaphrodite vulva, the male tail or the neurite processes of the M4, I2 and PVQ neurons. However, the growth rate of gcn-1 and abcf-3 mutants from embryogenesis to the fourth larval stage was around 24 hours longer than that of wild-type animals, and the mitotic pachytene region of the hermaphrodite gonad was expanded over the loop regions of the gonads. (data not shown). These observations suggest that gcn-1 and abcf-3 affect biological processes in addition to programmed cell death.
Given the ubiquitous expression patterns of gcn-1 and abcf-3, we tested whether gcn-1 and abcf-3 promote programmed cell deaths in addition to that of the M4 sister. We examined gcn-1(n4827) and abcf-3(n4927) mutants for defects in the deaths of the NSM sisters, the PVQ sisters, the g1A sisters, the RIM and RIC sisters and multiple cells in the anterior pharynx. gcn-1(n4827) and abcf-3(n4927) single mutants did not exhibit defects in the deaths of these cells (Table 1). However, when either the gcn-1(n4827) or the abcf-3(n4927) mutation was combined with the partial loss-of-function ced-3(n2427) mutation, which sensitizes strains to weak defects in cell death [36], we observed significant cell-death defects for all cell types tested (Table 1). For example, the gcn-1(n4827) and abcf-3(n4927) mutations enhanced the ced-3(n2427) defect from 16% to 38% and 34%, respectively, for the NSM sister and from 13% to 36% and 45%, respectively, for the g1A sister. We conclude that gcn-1 and abcf-3 promote programmed cell death generally rather than specifically affecting the M4 sister cell death.
We next tested whether gcn-1 and abcf-3 are involved in the deaths of germ cells in the gonad of the adult hermaphrodite. More than half of germ cells stochastically undergo programmed cell death under normal conditions during oocyte differentiation [37]. We scored the number of apoptotic germ cells using the vital dye acridine orange (AO), which stains nucleic acids within apoptotic cells in living animals [38]. gcn-1(n4827) and abcf-3(n4927) mutants had 9.1 and 9.5 apoptotic germ cells per gonadal arm on average, respectively, similar to wild-type animals, which had 8.6 apoptotic germ cells per gonadal arm (Figure 4I). We also scored the number of apoptotic germ cells by direct observation of the gonads of engulfment-defective ced-1(e1735) mutants, in which cell corpses accumulate because of a defect in cell-corpse engulfment, facilitating a sensitive assay for the deaths of germ cells [37]. ced-1(e1735) mutants had an average of 14.4 cell corpses per gonadal arm (Figure S5). ced-1(e1735) double mutants with gcn-1(n4827) or abcf-3(n4927) had nearly identical numbers of cell corpses per gonadal arm, 13.9 and 14.0, respectively (Figure S5). These results indicate that gcn-1 and abcf-3 are dispensable for germ-cell death under physiological conditions.
Since many germ cells undergo apoptosis in response to genotoxic stresses such as ionizing radiation [39], we tested whether gcn-1 and abcf-3 mediate ionizing radiation damage-induced germ cell death. As assayed with AO, wild-type animals normally contained an average of 8.6 apoptotic germ cells per gonadal arm, while wild-type animals exposed to ionizing radiation contained on average 27.1 apoptotic germ cells (Figure 4A, 4E and 4I). This germ-cell death was completely blocked by a mutation in the caspase gene ced-3 in wild-type, gcn-1(n4837) and abcf-3(n4927) animals (Figure 4B, 4F and 4I). Strikingly, ionizing radiation failed to increase the number of apoptotic germ cells in gcn-1(n4827) and abcf-3(n4927) mutants (10.9 and 10.4 apoptotic germ cells per gonadal arm in gcn-1 and abcf-3 mutants, respectively, 24 hours after gamma ray irradiation) (Figure 4C, 4D, 4G, 4H and 4I). These results indicate that gcn-1 and abcf-3 are required for ionizing radiation-induced germ cell death but not for the stochastic germ cell death that occurs in physiological conditions.
The gcn-1(n4827) and abcf-3(n4927) mutations partially blocked both the programmed cell deaths of somatic cells (Table 1) and the deaths of germ cells in response to ionizing radiation (Figure 4). Both somatic and ionizing radiation-induced germ cell deaths involve the canonical cell-death execution pathway consisting of the BH3-only gene egl-1, the BCL-2 homolog ced-9, the pro-apoptotic APAF-1 homolog ced-4, and the caspase gene ced-3 [40]. Interestingly, animals doubly heterozygous for gcn-1 and ced-3, ced-4 or egl-1 had a defect in M4 sister cell death (gcn-1/+; ced-3/+ 18%, gcn-1/+; ced-4/+ 12% or gcn-1/+; egl-1/+ 17%, respectively) significantly higher than that of singly heterozygous animals (gcn-1/+ 4%, ced-3/+ 0%, ced-4/+ 0% or egl-1/+ 1%, respectively) (Table S5). These results indicate that the simultaneous reduction by half of the dosage of gcn-1 and of genes in the canonical cell-death execution pathway causes a significant defect in M4 sister cell death. We observed a similar genetic interaction in animals heterozygous for abcf-3 and ced-3, ced-4 or egl-1 (Table S5).
We observed that maternal gcn-1 and abcf-3 contribute to zygotic programmed cell death. While gcn-1(−) animals generated by gcn-1(−) hermaphrodites and gcn-1(−) males exhibited a defect in M4 sister cell death (12% of M4 sister survival), gcn-1(−) animals produced from gcn-1/+ hermaphrodites and gcn-1(−) males did not (0% of M4 sister survival) (Table S5), indicating that maternal gcn-1 is sufficient to promote programmed cell death. gcn-1(−) animals generated by gcn-1(−) hermaphrodites and gcn-1/+ males exhibited a defect in M4 sister cell death (13% of M4 sister survival). gcn-1/+ animals generated by gcn-1(−) hermaphrodites and gcn-1(+) males exhibited a very weak defect in M4 sister cell death (4% of M4 sister survival) compared to 12% of M4 sister survival in gcn-1(−) self-progeny of gcn-1(−) hermaphrodites. By contrast, gcn-1/+ animals produced from gcn-1(+) hermaphrodites and gcn-1(−) males exhibited no defect in M4 sister cell death (0% of M4 sister survival) (Table S5). These results indicate that maternal gcn-1 is partially required for the M4 sister to undergo programmed cell death. We observed a similar maternal requirement and sufficiency for abcf-3 (Table S5). We conclude that maternal gcn-1 and abcf-3 are sufficient and partially required for the M4 sister to undergo programmed cell death.
To examine interactions between gcn-1 and abcf-3 and the canonical cell-death execution pathway, we performed epistasis analyses between gcn-1 or abcf-3 and ced-9, which functions downstream of egl-1 and upstream of ced-4 and ced-3 in the cell-death execution pathway [40]. Because the ced-9(n2812) null mutation causes ectopic cell deaths and organismic inviability, we used the ced-3 partial loss-of-function mutation n2446 to suppress ced-9(n2812) lethality [41]. We observed that 50% of ced-9(n2812) animals had a surviving M4 sister in the ced-3(n2446) mutant background. This increase over the 5% frequency of M4 sister survival in ced-3(n2446) mutants is consistent with the proposal that ced-9 has a cell-killing activity [42]. We observed that gcn-1 ced-9 and abcf-3 ced-9 double mutants were more highly penetrant for M4 sister survival (90% and 89%, respectively) than either single mutant in the ced-3(n2446) mutant background: gcn-1 (45%), abcf-3 (40%) and ced-9 (50%), respectively (Table 2). These results indicate that ced-9 is not required for gcn-1 and abcf-3 to promote programmed cell death. Thus, gcn-1 and abcf-3 function downstream of or in parallel to ced-9 in the regulation of programmed cell death.
We next tested whether the activity of gcn-1 and abcf-3 can act cell-autonomously to promote programmed cell death. Previous studies showed that expression of a ced-3, ced-4 or egl-1 cDNA under the control of the mec-7 promoter can act cell-autonomously to cause the deaths of a set of touch neurons, including the PLML and PLMR cells. We expressed gcn-1 and abcf-3 cDNAs in the PLM neurons under the control of the mec-7 promoter. We observed that 100% of the PLM neurons survived in wild-type animals, whereas only 50% or 89% of the PLM neurons survived in animals expressing gcn-1 or abcf-3, respectively, under the control of the mec-7 promoter (Figure 5A). Expression of both gcn-1 and abcf-3 also reduced a survival of the PLM neurons: 61% of the PLM neurons survived. These results indicate that expression of gcn-1 and abcf-3 are sufficient to induce cell death and suggest that gcn-1 and abcf-3 acts cell-autonomously to promote programmed cell death.
Our genetic screen for mutants defective in M4 sister cell death identified other genes in addition to gcn-1 and abcf-3: ceh-34, eya-1, sptf-3 and pig-1 [15], [17]. We previously showed that the Six family homeodomain protein CEH-34 and the Eyes absent homolog EYA-1 directly drive the transcription of the BH3-only gene egl-1 in the M4 sister to promote M4 sister cell-type specific death [15] and that the SP1 family transcription factor SPTF-3 directly drives the transcription of both egl-1 and the AMPK-related protein kinase gene pig-1, which also promotes M4 sister cell death [17]. We determined how gcn-1 and abcf-3 interact with these genes by examining double mutants. The partial loss-of-function alleles ceh-34(n4796) and sptf-3(n4850) and the null allele pig-1(gm344Δ) enhanced the M4 sister-cell death defect of gcn-1(n4827) and abcf-3(n4927) null mutants (Table 3). These results indicate that ceh-34, sptf-3 and pig-1 function in pathways distinct from that of gcn-1 and abcf-3 to promote M4 sister cell death.
We demonstrated that the translational regulators GCN-1 and ABCF-3 are pro-apoptotic factors that maternally contribute to the programmed cell death of the M4 sister in C. elegans. GCN-1 and ABCF-3 promote the deaths of all somatic cells tested. Essentially all somatic cell deaths are mediated by an evolutionarily conserved cell-death execution pathway consisting of the BH3-only gene egl-1, the BCL-2 homolog ced-9, the APAF-1 homolog ced-4 and the caspase gene ced-3 [40]. How do gcn-1 and abcf-3 interact with this pathway to regulate apoptosis? We propose that gcn-1 and abcf-3 likely act in a novel pathway distinct from the canonical cell-death execution pathway. First, gcn-1 and abcf-3 promote apoptosis in the absence of ced-9 activity, indicating that gcn-1 and abcf-3 function independently of ced-9 in the regulation of apoptosis and hence do not regulate either ced-9 or egl-1. Second since ced-3 and ced-4 function downstream of ced-9 in the cell-death execution pathway, gcn-1 and abcf-3 could act through these genes to promote cell death. However, our mRNA-seq and Ribo-seq results indicate that gcn-1 and abcf-3 do not have major effects on mRNA abundance or the ribosome footprint density of ced-3 and ced-4 (Figure S6). Our preferred model is that GCN-1 and ABCF-3 function in a pathway that acts in parallel to the canonical cell-death execution pathway, although we cannot preclude the possibility that GCN-1 and ABCF-3 translationally regulate unidentified factors that act through ced-3 or ced-4 without changing the transcriptional and translational levels of the products of these genes. Also, since we used whole animals for our mRNA-seq and Ribo-seq analyses, we would not have detected alterations in CED-3 or CED-4 levels that were specific to a small subset of cells, including the M4 sister.
Our genetic analyses revealed that gcn-1 and abcf-3 maternally contribute to the death of the M4 sister, which undergoes programmed cell death during embryogenesis. Maternally-contributed factors might act to ensure the rapid deaths of cells during embryogenesis; perhaps zygotic expression of apoptotic genes would be too slow. Also, the maternal effects of gcn-1 and abcf-3 might explain why we discovered new general cell-death genes, despite the fact that many genetic screens have been performed in search of C. elegans mutants defective in somatic cell deaths. Most such genetic screens have examined F2 animals after mutagenesis, and would have missed maternally-contributed genes that affect general cell death. Perhaps, additional maternal-effect genes with functions in apoptosis exist in C. elegans. Such genes might be efficiently identified by screening in the third generation after mutagenesis.
gcn-1(n4827) and abcf-3(n4927) single mutations appeared to cause a defect in only M4 sister cell death, and these mutations both affected other cell deaths to differing extents in strains sensitized to weak defects in cell death. For example, the death of the M4 sister was most sensitive and the deaths of the PVQ sisters were least sensitive to the gcn-1(n4827) and abcf-3(n4927) mutations among the cells we tested (Table 1). We speculate that sensitivity to perturbation of cell-death genes is different among different cell types. This hypothesis is supported by the observations that penetrance of cell-death defects varies among different cell types in partial loss-of-function ced-3(n2427) mutants and that the extent of the cell-death defect of ced-3(n2427) mutants is well correlated with that of gcn-1(n4827) and abcf-3(n4927) mutants.
Our genetic and biochemical data strongly suggest that GCN-1 and ABCF-3 physically interact in a complex in vivo to promote apoptosis. First, GCN-1 and ABCF-3 interacted in the yeast two-hybrid system. Second, GCN-1 co-immunoprecipitated with ABCF-3 from a total protein extract from C. elegans. Third, the absence of either the GCN-1 or ABCF-3 protein decreased the steady-state level of the other protein (ABCF-3 or GCN-1, respectively), indicating that an interaction between GCN-1 and ABCF-3 is likely important for the stability of both proteins. Fourth, gcn-1(n4827) abcf-3(n4927) double mutants were not enhanced in the defect in apoptosis compared to each single mutant.
Although GCN-1 and ABCF-3 promote the deaths of most somatic cells, in gcn-1 or abcf-3 mutants only 12% or 13% of animals are defective in M4 sister cell death, respectively, and the cell-death defect of most other cells was observed only in a partial loss-of-function ced-3 mutant background, which is sensitized to weak defects in cell death. Furthermore, loss-of-function of gcn-1 and abcf-3 did not affect the deaths of germ cells under physiological conditions. By striking contrast, we found that GCN-1 and ABCF-3 play an essential role in germ-cell deaths induced by ionizing radiation. These results suggest that translational control by GCN-1 and ABCF-3 plays a more important role in germ-cell deaths induced by ionizing radiation than in somatic cell deaths. The hypothesis that translational control is particularly important for cell deaths induced by ionizing radiation is supported by a recent report that a mutation in RNA polymerase I (rpoa-2), which synthesizes ribosomal RNAs, causes a defect in germ-cell deaths induced by ionizing radiation [43].
Ionizing radiation causes DNA double-strand breaks, which lead to the progressive accumulation of mutations and chromosomal aberrations as damaged cells undergo division, resulting in apoptosis and the demise of genetically damaged cells. In C. elegans, ionizing radiation causes massive deaths of the germ cells during the late pachytene stage of oocyte development in adult gonads, resulting in the elimination of the damaged oocytes [39]. Germ-cell deaths induced by ionizing radiation specifically involve activation of the p53 homolog CEP-1 by the DNA damage response pathway and subsequent CEP-1- dependent transcriptional induction of the BH3-only gene egl-1, which activates the cell-death execution pathway regulated by CED-9 [44], [45]. How might GCN-1 and ABCF-3 interact with the known DNA-damage response and cell-death execution pathways in the regulation of germline cell deaths induced by ionizing radiation? Our genetic results suggest that GCN-1 and ABCF-3 function independently of CED-9, at least in the regulation of the death of the M4 sister cell. We suggest that as is the case for somatic cell deaths, GCN-1 and ABCF-3 function in a novel pathway independently of CED-9 in regulating the germ-cell deaths induced by ionizing radiation. Alternatively, if GCN-1 and ABCF-3 regulate germ-cell deaths induced by ionizing radiation via a mechanism different from that of somatic cell deaths, it is possible that GCN-1 and ABCF-3 act through egl-1 and its target ced-9, since egl-1 is involved in somatic programmed cell deaths and germ-cell deaths induced by ionizing radiation but not in the stochastic germ-cell deaths that occur under physiological conditions.
GCN-1 and ABCF-3 are conserved proteins from yeast to humans. The C. elegans GCN-1 protein has 43% and 53% similarities (23% and 32% identities) to the homologs of S. cerevisiae and humans, respectively, and the C. elegans ABCF-3 proteins has 57% and 69% similarities (40% and 49% identities) to the homologs of S. cerevisiae and humans, respectively (Figure 1H). The yeast GCN-1 homolog Gcn1p and ABCF-3 homolog Gcn20p are required to maintain the basal level of the phosphorylation of eukaryotic initiation factor 2α (eIF2α) in the physiological condition and to increase the phosphorylation of eIF2α in response to amino-acid starvation. Gcn1p and Gcn20p activate the serine-threonine protein kinase Gcn2p, which phosphorylates an evolutionarily conserved serine residue of eIF2α. The phosphorylation of eIF2α results in both the inhibition of global translation and the translational activation of the GCN4 mRNA, which encodes a basic leucine zipper transcription factor. Translation of GCN4 mRNA is regulated by four short upstream open reading frames (uORFs) in the 5′ UTR with start codons that are out-of-frame with the main coding sequence and which generally reduce translation from the main reading frame [46].
We speculate that the mechanistic roles of C. elegans GCN-1 and ABCF-3 in translational control are conserved between yeast and C. elegans. First, the amino-acid sequences of GCN-1 and ABCF-3 proteins are conserved between yeast and C. elegans, particularly in functionally important domains (Figure 1H). Second, the functions of GCN-1 and ABCF-3 can be substituted with those of S. cerevisiae GCN1 and GCN20, respectively, for the promotion of M4 sister cell death. Third, like their yeast counterparts, C. elegans GCN-1 and ABCF-3 are required to maintain the basal level of phosphorylation of eIF2α and physically interact through an EF3-like domain-containing region of GCN-1 and an N-terminal ABCF-3 domain [43]. Fourth, like Gcn20p, the AAA domain ATPase activity of ABCF-3 is not required for its function [32]. Fifth, the atf-5 gene, the C. elegans homolog of S. cerevisiae GCN4, has two upstream ORFs that have been shown to inhibit the translation of the atf-5 mRNA [47].
We have shown that in addition to gcn-1 and abcf-3, ceh-34, eya-1, sptf-3 and pig-1 function in M4 sister cell death [15], [17]. We previously reported that the Six family homeodomain protein CEH-34 and the Eyes absent homolog EYA-1 physically interact to directly drive expression of the pro-apoptotic BH3-only gene egl-1 in the M4 sister, leading to the death of the M4 sister (Figure 5B) [15]. We found that the SP1 family transcription factor SPTF-3 directly drives the transcription of the gene egl-1, which encodes a BH3-only protein that promotes apoptosis via the CED-3 caspase-mediated canonical cell-death execution pathway [17]. SPTF-3 also directly drives the transcription of the AMPK-related gene pig-1, which encodes a protein kinase that functions in a pathway in parallel to the CED-3-mediated canonical cell-death execution pathway. These interactions are shown in Figure 5B.
Our analyses indicate that gcn-1 and abcf-3 likely function in a pathway that acts in parallel to those of pig-1, ceh-34 and sptf-3. These results are consistent with a model in which GCN-1 and ABCF-3 act independently of CED-9 to promote M4 sister cell death. In short, we propose that the regulatory network for the death of the M4 sister includes at least three different pathways involving translation, transcription and protein phosphorylation (Figure 5B). Each gene in this network (gcn-1, abcf-3, sptf-3, pig-1, egl-1, ceh-34 and eya-1) has a human counterpart, some of which are implicated in human diseases, including developmental disorders and cancer. We anticipate that further analyses of this regulatory network will both reveal an evolutionarily conserved mechanism of apoptosis shared between C. elegans and humans and provide insights concerning how abnormalities in this apoptotic network can lead to human disease.
C. elegans strains were cultured at 20°C as described [48]. The N2 strain was used as the wild type. The following mutations, integrations and extrachromosomal arrays were used.
LGI: sptf-3(n4850), eya-1(ok654Δ), nIs177[Pceh-28::gfp, lin-15AB(+)], nIs180[Ptdc-1::gfp, lin-15AB(+)], zdIs5[Pmec-4::gfp, lin-15AB(+)].
LGII: rol-1(e91), gcn-2(ok871Δ), Y38E10A.8(tm4094Δ).
LGIII: ced-4(n1162), ced-9(n2812), gcn-1(n4827, nc40Δ), abcf-3(n4927, ok2237Δ), unc-45(r450), dpy-18(e364), nIs176[Pceh-28::gfp, lin-15AB(+)].
LGIV: ced-3(n717, n2427, n2446), pig-1(gm344Δ), nIs175[Pceh-28::gfp, lin-15AB(+)].
LGV: egl-1(n1084 n3082), ceh-34(n4796), oyIs14[sra-6::gfp].
LGX: lin-15(n765), pek-1(ok275Δ), nIs106[Plin-11::gfp, lin-15AB(+)], nIs429[Pphat-5::gfp, lin-15AB(+)], bcIs24[Ptph-1::gfp, lin-15AB(+)].
Unmapped: nIs460[Pgcn-1::gfp], nIs488[Pabcf-3::gfp], nIs645 and nIs646[Pmec-7::gcn-1 cDNA, Pmec-7::abcf-3 cDNA, Pmec-3::mCherry, rol-6(su1006)], nIs648 and nIs649[Pmec-7::gcn-1 cDNA, Pmec-3::mCherry, rol-6(su1006)], nIs651 and nIs652[Pmec-7::abcf-3 cDNA, Pmec-3::mCherry, rol-6(su1006)]. Extrachromosomal arrays: nEx1817 and nEx1818[Pgcn-1::gcn-1 cDNA::gcn-1 3′ UTR, Plin-44::gfp], nEx1925 and nEx1926[abcf-3(+), Plin-44::gfp], nEx1928 and nEx1929[abcf-3 K217M, Plin-44::gfp], nEx1931 and nEx1932[abcf-3 K536M, Plin-44::gfp], nEx1934 and nEx1935[abcf-3 K217M K536M, Plin-44::gfp], nEx2223 and nEx2224[Pceh-34::eIF2α S49A, Plin-44::gfp]
gcn-1(n4827) and abcf-3(n4927) were isolated from a genetic screen for mutations that cause an extra GFP-positive M4-like cell in animals carrying the Pceh-28::gfp transgene [15]. Mutagenesis was performed as described [48]. Mutagenized P0 animals were allowed to lay eggs, and 144,000 synchronized F2 animals were screened with a fluorescence-equipped dissecting microscope. Single nucleotide polymorphisms were used to map gcn-1(n4827) and abcf-3(n4927) to a 175 kb interval (III: 2,044,521–2,220,200) and a 5.3 Mb interval (III: 5,346,407–10,613,191), respectively [49]. Whole-genome sequencing of gcn-1(n4827) mutants was performed using an Illumina/Solexa GAII, according to the instructions of the manufacture. DNA sequencing of the abcf-3 locus of abcf-3(n4927) mutants was performed using an Applied Biosystems 3130×.
The programmed cell deaths of specific cells were scored at the indicated stages using the following strains, which express GFP in specific cells. A fluorescence-equipped compound microscope was used to score the programmed cell deaths. M4 sister cell death, nIs175, nIs176 or nIs177 at the L1 stage. g1A sister cell death, nIs429 at the L1 stage. PVQ sister cell death, oyIs14 at the L4 stage [50]. NSM sister cell death, bcIs24 at the L1 stage [51]. RIM and RIC sister cell death, nIs180 at the L1 stage. Extra cells in the anterior pharynx were scored using a compound microscope equipped with Nomarski differential interference contrast optics. For physiological germ-cell deaths, germ-cell corpses in gonads of animals 24 hours after the fourth-larval stage were counted by direct observation using Nomarski optics. For ionizing radiation-induced germ-cell deaths, fourth-larval stage animals were exposed to 120 Gy of ionizing radiation, and germ-cell deaths were scored using acridine orange at 24 hours post-irradiation as described [38].
The transgenes Pceh-28::gfp, Ptph-1::gfp and sra-6::gfp are described [15], [50], [51]. The phat-5 promoter sequence in pGD48 was cloned in pPD122.56 to generate the Pphat-5::gfp transgene [52]. The Pflp-15::gfp transgene contained 2.4 kbp of the 5′ promoter of flp-15 in pPD122.56. The Pgcy-37::gfp transgene contained 1.1 kbp of 5′ promoter of gcy-37 in pPD122.56. The Ptdc-1::gfp transgene contained 4.5 kbp of 5′ promoter of tdc-1 in pPD121.83. The Pgcn-1::gcn-1 cDNA::gcn-1 3′UTR transgene (pTH gcn-1 cDNA) contained 4.2 kbp of 5′ promoter of gcn-1, a full-length gcn-1 cDNA and 1.0 kbp 3′ of the stop codon of gcn-1. The 5′ promoter of gcn-1, a full-length gcn-1 cDNA and the 3′ promoter of gcn-1 were generated by PCR and fused in pBluescript II using the In-Fusion cloning system (Clontech). The abcf-3(+) transgene contained 1.6 kbp of 5′ promoter, the coding region and 0.8 kbp 3′ of the stop codon of abcf-3 in pBluescript II. The QuickChange II XL Site-Directed Mutagenesis Kit (Stratagene) was used to generate transgenes of abcf-3 K217M, abcf-3 K536M and abcf-3 K217M K536M. gcn-1 cDNA corresponding to amino acids 1–2634, 1–1760, 1–880, 880–2634, 880–1760 or 1760–2634 of GCN-1 was cloned in pGBKT7. abcf-3 cDNA corresponding to amino acids 1–712, 1–512, 1–202, 202–712, 512–712 or 202–512 was cloned in pGADT7. The Pgcn-1::gfp transgene contained 4.2 kbp of 5′ promoter of gcn-1 in pPD122.56. The Pabcf-3::gfp transgene contained 1.6 kbp of 5′ promoter of abcf-3 in pPD122.56. The Pceh-34::eIF2α S49A transgene contained 3.8 kbp of 5′ promoter of ceh-34 and eIF2α with a replacement of serine 49 with alanine in pPD49.26. For the Pmec-7::gcn-1 cDNA and Pmec-7::abcf-3 cDNA transgenes, full-length cDNA of gcn-1 and abcf-3 were cloned in pPD96.41. Primer sequences used are available from the authors.
Germline transformation was performed as described [53]. The gfp reporter transgene was injected at 50 µg/ml into lin-15(n765ts) animals with 50 µg/ml of pL15EK as a coinjection marker [54]. To rescue the defect in M4 sister cell death, the transgenes pTH gcn-1 cDNA, abcf-3(+), abcf-3 K217M, abcf-3 K536M and abcf-3 K217M K536M described above were injected at 20 µg/ml into gcn-1(n4827) or abcf-3(n4927) animals with 50 µg/ml of Plin-44::gfp as a coinjection marker [55]. To establish transgenic lines carrying the Pceh-34::eIF2α S49A transgene, the Pceh-34::eIF2α S49A transgene was injected at 50 µg/ml into nIs175 animals with 50 µg/ml of Plin-44::gfp as a coinjection marker. The Pmec-7::gcn-1 cDNA and Pmec-7::abcf-3 cDNA transgenes were injected at 50 µg/ml, respectively, into zdIs5 animals with 50 µg/ml of pRF4[rol-6(su1006)] and 20 µg/ml of the Pmec-3::mCherry transgene as coinjection markers.
GAL4 fusion constructs were introduced into yeast strain PJ649A as described [56]. Single colonies were streaked and cultured for two days at 30°C on SD plates containing minimal supplements without tryptophan and leucine. Then yeast strains were streaked and cultured for three days at 30°C on SD plates containing minimal supplements without tryptophan, leucine and histidine to test yeast growth.
Protein fragments corresponding to amino acids 753–857 of GCN-1 and 74–185 of ABCF-3 fused to glutathione S- transferase (GST) were expressed, purified using glutathione Sepharose 4B (Amersham Biosciences) and used to raise rabbit anti-GCN-1 or anti-ABCF-3 antibodies, respectively. Antisera were generated by Pocono Rabbit Farm and Laboratory. Specific antibodies were affinity-purified using identical GCN-1 or ABCF-3 protein fragments fused to maltose-binding protein (MBP) and coupled to Affigel 10 (Bio-Rad).
Protein extracts were prepared from nIs175, gcn-1(n4827); nIs175 and abcf-3(n4927); nIs175 animals synchronized at the fourth larval stage as described [57]. 10 µg of total protein was loaded onto a 7.5% SDS PAGE gel and then transferred to nitrocellulose membranes. The membranes were probed with anti-GCN-1 or anti-ABCF-3 antibody. Immunocomplexes were detected using HRP-conjugated anti-rabbit IgG secondary antibodies (Invitrogen) followed by chemiluminescence (Western Lightning ECL, PerkinElmer). To determine the level of phosphorylated eIF2α, protein extracts were prepared from nIs175, gcn-1(n4827); nIs175 and abcf-3(n4927); nIs175 animals synchronized at the fourth larval stage as described [57]. 15 µg of total protein was loaded on a 10% SDS PAGE gel and then transferred to nitrocellulose membranes. The membranes were probed with anti-phospho-eIF2α (Cell Signaling Technology) and anti-eIF2α antibodies [28]. Immunocomplexes were detected as described above.
For immunoprecipitation experiments, protein extracts were prepared from mixed-staged wild-type animals in TNE buffer containing 50 mM Tris-HCl (pH 8.0), 150 mM NaCl, 1 mM EDTA, 1% NP-40, 5 mM β-mercaptoethanol, 10% glycerol. Protein extracts were mixed with either an affinity-purified anti-ABCF-3 antibody or a control IgG at 4°C for 2 hours. Immunocomplexes were recovered using Protein A Sepharose 4 Fast Flow (GE Healthcare Life Sciences) and washed with TNE buffer four times. The recovered immunocomplexes were subjected to western blot analysis using anti-GCN-1 or anti-ABCF-3 antibody.
Fluorescence in situ hybridization was performed as described [58]. The gcn-1 and abcf-3 probes (Biosearch Technologies, Inc) were conjugated to the fluorophore Cy5 using the Amersham Cy5 Mono-reactive Dye pack (GE Healthcare). DNA was visualized using 4′,6-diamidino-2-phenylindole (DAPI). The probe sequences used are shown in Tables S3 and S4. Figures 4D and H are maximum intensity projections of a Z-stack of images processed with the FFT Bandpass Filter operations in the image processing program Fiji. Oligonucleotides used for gcn-1 and abcf-3 FISH probe were described in Table S6 and S7).
For mRNA-seq [59], total RNA was purified using an RNAeasy Mini kit (Qiagen) from synchronized L4 animals of wild-type animals and gcn-1 and abcf-3 mutants. The purified RNA was subjected to oligo (dT) selection, fragmentation and first- and double-strand synthesis with an Illumina Tru-Seq kit according to the manufacturer's instructions. DNA fragments longer than 30 bp were purified using SPRI-TE beads (Beckmann Coulter) according to the manufacturer's instructions. The purified DNA was end-repaired and single A bases were added for adaptor ligations. The adaptor-ligated DNA was then subjected to double SPRI-TE purification to select for 200 bp fragments. These fragments were enriched and barcoded by PCR for multiplexing. A final SPRI-TE purification was performed to purify the barcoded RNA-Seq libraries for Illumina DNA sequencing using HiSeq 2000. RNA-seq data were aligned against the C. elegans reference genome (ce10) using the Burrows-Wheeler Aligner (BWA) and Tophat.
Ribosome profiling was performed as described [60] with modifications. Synchronized L4 wild type animals and gcn-1 and abcf-3 mutants were collected and washed with M9 buffer three times. Animals were homogenized using a dounce homogenizer in lysis buffer containing 20 mM Tris (pH 7.5), 150 mM NaCl, 5 mM MgSO4, 1 mM DTT, 100 µg/ml cycloheximide, 1% Triton X-100 and 25 U/ml Turbo DNase (Invitrogen) and centrifuged at 20,000 g for 20 min at 4°C. The absorbance of the extract was measured at 260 nm. 40 absorbance units of extract were incubated with 300 units of RNase I at 25°C for an hour, and then 200 units of SUPERase In RNase Inhibitor (Invitrogen) were added. Digested extracts were loaded on 10–50% linear sucrose gradients containing 20 mM Tris (pH 7.5), 150 mM NaCl, 5 mM MgSO4, 1 mM DTT and 100 µg/ml cycloheximide and centrifuged for three hours at 35, 000 rpm at 4°C using a SW-40 rotor to isolate a monosome fraction. RNA from the monosome fraction was purified by phenol-chloroform extraction followed by miRNeasy Mini Kit (Qiagen) and separated using a 15% TBE-Urea gel (BioRad) to isolate ribosome-protected fragments (RPFs). The RPFs were eluted from gels by incubating in RNA elution buffer containing 300 mM sodium acetate (pH 5.5), 1 mM EDTA and 0.25% SDS. RPFs were 3′ dephosphorylated with T4 polynucleotide kinase (New England Labs) and ligated to Universal miRNA Cloning Linker (New England Labs) using T4 RNA ligase 2, truncated (New England Labs) according to the manufacturer's instructions. RPFs ligated with the linker were separated from an unligated linker using a 15% TBE-Urea gel (BioRad) and eluted from gels using RNA extraction buffer followed by phenol-chloroform extraction. RFPs were reverse-transcribed by Superscript III (Invitrogen) with a reverse transcription primer according to the manufacturer's instructions. The products of reverse transcripts (RT) were purified using a 15% TBE-Urea gel (BioRad) and eluted from a gel by incubating in DNA elution buffer containing 300 mM NaCl, 10 mM Tris (pH 8.0) and 1 mM EDTA followed by phenol-chloroform extraction. The RT products were circularized by CircLigase (Epicentre) according to the manufacturer's instructions. About a quarter of the RT products were used in PCR reactions containing 1× Phusion HF buffer, 0.2 mM dNTP, 0.5 µm forward library primer, 0.5 µm reverse indexed primer and 0.02 units/µl Phusion polymerase (New England Labs), and PCR was performed with a 30 second initial denaturation at 98°C, followed by 6, 8, 10, 12 and 14 cycles of 98°C for 10 second, 65°C for 10 second and 72°C for 5 second. PCR products were separated using a 8% TBE gel (BioRad) and eluted from gels by incubating in DNA elution buffer followed by phenol-chloroform extraction. PCR products were suspended in 20 µl of 10 mM Tris (pH 8.0) and sequenced by HiSeq 2000. The adaptor sequences (CTGTAGGCACCATC) from 3′ end of the ribosome footprint reads were removed, then trimmed reads were mapped using BWA to distinguish the reads from ribosomal RNAs. About 60% of the reads were filtered out, and the remaining reads (non-ribosomal) were aligned to the C. elegans reference genome (ce10) using BWA and Tophat. Because translational initiation is thought to be blocked rapidly by the stress animals encounter during harvesting, many ribosomes are stalled at the beginning of each transcript in the presence of cycloheximide, which prevents translation elongation [34]. Hence, high frequencies of reads at the beginning of each transcript might not correspond to high rates of translation. For this reason, the reads that mapped to the first 25 nucleotides of each transcript were not counted in evaluating gene expression in the Ribo-seq analyses.
|
10.1371/journal.pgen.1002356 | SOX9 Governs Differentiation Stage-Specific Gene Expression in Growth Plate Chondrocytes via Direct Concomitant Transactivation and Repression | Cartilage and endochondral bone development require SOX9 activity to regulate chondrogenesis, chondrocyte proliferation, and transition to a non-mitotic hypertrophic state. The restricted and reciprocal expression of the collagen X gene, Col10a1, in hypertrophic chondrocytes and Sox9 in immature chondrocytes epitomise the precise spatiotemporal control of gene expression as chondrocytes progress through phases of differentiation, but how this is achieved is not clear. Here, we have identified a regulatory element upstream of Col10a1 that enhances its expression in hypertrophic chondrocytes in vivo. In immature chondrocytes, where Col10a1 is not expressed, SOX9 interacts with a conserved sequence within this element that is analogous to that within the intronic enhancer of the collagen II gene Col2a1, the known transactivation target of SOX9. By analysing a series of Col10a1 reporter genes in transgenic mice, we show that the SOX9 binding consensus in this element is required to repress expression of the transgene in non-hypertrophic chondrocytes. Forced ectopic Sox9 expression in hypertrophic chondrocytes in vitro and in mice resulted in down-regulation of Col10a1. Mutation of a binding consensus motif for GLI transcription factors, which are the effectors of Indian hedgehog signaling, close to the SOX9 site in the Col10a1 regulatory element, also derepressed transgene expression in non-hypertrophic chondrocytes. GLI2 and GLI3 bound to the Col10a1 regulatory element but not to the enhancer of Col2a1. In addition to Col10a1, paired SOX9–GLI binding motifs are present in the conserved non-coding regions of several genes that are preferentially expressed in hypertrophic chondrocytes and the occurrence of pairing is unlikely to be by chance. We propose a regulatory paradigm whereby direct concomitant positive and negative transcriptional control by SOX9 ensures differentiation phase-specific gene expression in chondrocytes. Discrimination between these opposing modes of transcriptional control by SOX9 may be mediated by cooperation with different partners such as GLI factors.
| Chondrogenic differentiation is a key process in the formation of endochondral bone. Despite the wealth of information about gene expression patterns and signaling pathways important for this process, it is not clear how differentiation state-specificity of transcription is controlled. The transcription factor SOX9 regulates chondrocyte differentiation, proliferation, and entry into hypertrophy and is highly expressed in immature/proliferating chondrocytes. It directly transactivates Col2a1, enhancing this gene's expression in immature/proliferating chondrocytes. The Col10a1 gene is specifically expressed in hypertrophic chondrocytes in which Sox9 is downregulated. How is differentiation phase-specific transcription of genes controlled in chondrocytes, particularly during hypertrophy? We found that SOX9 directly represses Col10a1 expression in immature/proliferating chondrocytes of the growth plate, so that its expression is restricted to hypertrophic chondrocytes. Discrimination of this concomitant opposing transcriptional control may involve cooperation between SOX9 and different partners such as GLI factors (effectors of hedgehog signaling). SOX9 control of chondrocyte maturation therefore may be integrated with hedgehog signaling. Mutations in human SOX9 cause the skeletal malformation syndrome campomelic dysplasia, which is attributed to the disruption of the chondrogenic differentiation program because of failure to express SOX9 target genes. This interpretation should be revised to include inappropriate expression of genes normally repressed by SOX9.
| Chondrogenesis and the formation of bone by endochondral ossification depend on progressive steps of cell differentiation. Mesenchymal cells condense and differentiate into chondrocytes in a pattern that will define the eventual shape of the different skeletal elements. These chondrocytes proliferate, mature, exit the cell cycle and become prehypertrophic. The differentiation program culminates in the terminal differentiation and apoptosis of post-mitotic hypertrophic chondrocytes [1]. This differentiation program is controlled by members of the SOX and RUNX families of transcription factors and the integration of multiple signaling pathways mediated by Indian hedgehog (Ihh), parathyroid hormone-related protein (PTHrP), Wnts, BMPs, and Notch (reviewed in [2]). PTHrP and Ihh are two important players which interact to form a feedback loop that controls the pace of the differentiation program [3].
Sox9 is essential for chondrogenesis and chondrocyte differentiation [4]–[6]. It is essential for mesenchymal condensation prior to chondrogenesis, and in its absence chondrocyte differentiation fails. Inactivation of Sox9 in chondrocytes at different stages of differentiation suggests that its expression is essential for the survival of chondrocytes so that they can progress to hypertrophy [5]–[7]. Mutations in SOX9 are associated with the human skeletal malformation syndrome, campomelic dysplasia, in which skeletal abnormalities can be attributed to the disruption of the chondrogenic differentiation program due to failure to express SOX9 target genes. Upon hypertrophy, chondrocytes down-regulate Sox9 expression [8], [9], which is believed to mark the end of SOX9 control in the growth plate.
Despite the wealth of information about spatial and temporal gene expression patterns in the developing growth plate, it is not clear how transcriptional controls achieve appropriate and specific gene expression during chondrocyte differentiation. SOX9 activates many genes expressed in proliferating chondrocytes, including the extracellular matrix (ECM) genes Col2a1, Col9a1, Col11a2, Acan (aggrecan) and Cd-rap/Mia1 [10]–[15]. For the Col2a1 gene, which is expressed most strongly in proliferating chondrocytes, SOX9 directly transactivates the gene in vivo via a conserved enhancer sequence within the first intron [10], [11].
The collagen X gene, Col10a1, is a hypertrophic chondrocyte specific marker. The specificity and reciprocity of Sox9 and Col10a1 expression epitomise the strict control of temporal and differentiation phase-specific gene expression in the growth plate. Col10a1 is ideal for studying transcriptional regulation because as well as its highly specific expression pattern, over-expression or loss-of-function does not disrupt chondrocyte differentiation. These properties simplify interpretation of changes in gene expression resulting from perturbing transcriptional control [16]–[18]. Here, we examined the transcriptional controls that restrict Col10a1 expression to hypertrophic chondrocytes. We found that SOX9 coordinates gene expression during chondrocyte differentiation through both transcriptional activation and repression. Discrimination between these opposing actions is probably achieved by cooperation between SOX9 and different partners such as GLI factors.
Previous cell transfection studies identified an enhancer element upstream of human COL10A1 [19]. This element is highly conserved in mammals and corresponds to a 640 bp region between −4.3 and −3.6 kb of the mouse Col10a1 gene (designated element A) (Figure 1A and 1B). We used DNase I footprinting assays to test the configurations in which the element A sequences could be directly bound by nuclear factors derived from chondrocytes at different differentiation states (Figure 2A). Extracts from hypertrophic chondrocytes MCTs, but not fibroblasts COS-1 or osteoblasts MC3T3-E1, protected six blocks of sequence (H1–H6). Noticeably, four different blocks (P1–P4) that partially overlap H1–H4 were protected by extracts from the proliferating chondrocyte/chondrosarcoma cell line CCL (Figure 1C and Figure 2A). Since proliferating chondrocyte/chondrosarcoma cells do not express Col10a1 (Figure S1A), these results suggest that in these cells, the proteins that bind to element A may contribute to the repression of Col10a1.
We and others previously showed that SOX9 regulates COL2A1/Col2a1 gene via a functional in vivo binding site in the intron 1 enhancer element [10], [11]. We identified the same SOX9-binding sequence within the Col10a1 element A (Figure 1C and Figure 2B). This site, COL2C1, lies on a region in block P3 that is not protected in hypertrophic chondrocytes, and is adjacent to a stretch of thymidine/guanine-rich (TG-rich) sequence. Electromobility shift assays revealed that SOX9 bound to this SOX9/TG-rich motif with a similar affinity as to the COL2A1 enhancer element (Figure S1C, S1E, S1F) [10] and the interaction involved dimeric binding (Figure 2C, cf. lane 3–4). SOX9 also interacted with the TG-rich motif but with a lower affinity than with the consensus SOX9 site. Mutation of the TG-rich motif reduced the overall SOX9 binding to the P3 element (Figure 2C, cf. lanes 6–10). The TG-rich motif resembles a RUNX binding consensus sequence, but we found that RUNX2 did not interact with this motif effectively compared with its binding to the RUNX site in the Bglap (osteocalcin) enhancer (Figure 2D). Chromatin immunoprecipitation (ChIP) assays using extracts from E13.5 mouse limb, a stage at which the cartilage anlagen is largely composed of immature chondrocytes, confirmed specific SOX9 binding to the Col10a1 element A and the Col2a1 enhancer in vivo (Figure 2E).
The paired SOX9 binding sequences in element A are separated by 4 bp, a distance similar to that between the paired SOX-like consensus sequences in Col2a1, Col9a1, and Acan that mediate transactivation of expression [15]. We tested the in vivo role of element A and the effects of SOX9/TG-rich motif mutations on the expression of Col10a1 mini-genes (Figure 3A) in transgenic mice. We have previously shown that a Flag-tagged Col10a1 vector Col10Flag (formerly known as FColX) is expressed in P10 hypertrophic chondrocytes [18]. Here, we show that in E15.5 humeri, the Col10Flag transgene was expressed in islands in prehypertrophic and hypertrophic chondrocytes in the upper hypertrophic zone (Figure 3B, e). In two independent mouse lines, a transgene comprising element A fused to the Col10Flag (Col10Flag-E) was expressed in a similar pattern as Col10Flag, but was significantly more strongly expressed than Col10Flag in all hypertrophic chondrocytes (Figure 3B, i), reflecting the enhancer activity of element A (see also Figure S2). However, mutation of the SOX9 site in element A (Col10Flag-EΔ1) resulted in marked expansion of the expression domain of the transgene, extending from the hypertrophic zone to the proliferating zone in the majority of transgenic fetuses (71.4%) (compare Figure 4, a with Figure 3B, i) and in almost all the Sox9-expressing chondrocytes in the rest (Figure 4, e). Expansion of transgene expression in proliferating chondrocytes was also noted but was less marked when the TG-rich motif in element A was mutated (Col10Flag-EΔ2) (compare Figure 4, m and i with a and e). Mutation of either SOX9 or the TG-rich motif did not abrogate transgene expression in hypertrophic chondrocytes. Together, these observations suggest that element A contains both positive and negative regulatory sequences, and that mutations in the SOX9/TG-rich motif in element A might disrupt SOX9-mediated repression in immature chondrocytes.
To test whether SOX9 negatively regulates Col10a1, we established a cell line from hypertrophic chondrocytes MCTs which expressed the Col10Flag-E transgene at the non-permissive (growth-arrest) temperature (Figure 5A). Similar to previous findings for dedifferentiated chondrocytic cells MC615, over-expression of SOX9 in MCTs did not transactivate endogenous Col2a1 [20]. However, SOX9 over-expression significantly down-regulated expression of both endogenous Col10a1 and the exogenous Col10Flag-E reporter (Figure 5B) supporting the notion that SOX9 is a negative regulator of Col10a1. To study the regulation in vivo, SOX9 was ectopically expressed in hypertrophic chondrocytes in mice. Mice expressing Cre recombinase inserted into the endogenous Col10a1 gene (Col10a1-Cre) [21] were crossed with transgenic mice carrying a single copy of a Cre-inducible Sox9-IRES-EGFP expression construct (Z/Sox9) [22] (Figure 5C). In Col10a1-Cre;Z/Sox9 mice, Sox9 and the linked Egfp reporter gene were activated in the hypertrophic chondrocytes at E17.5 in the anterior ribs (Figure 5D, h, i). These mice displayed an expanded hypertrophic zone and reduced Col10a1 expression (Figure 5D, e, k). Interestingly, we also found that transcription of Cre from the Col10a1 locus was reduced in Col10a1-Cre;Z/Sox9 mice (Figure 5D, a, g). Col2a1 expression in Col10a1-Cre;Z/Sox9 mice hypertrophic chondrocytes was not up-regulated suggesting the reduction of Col10a1 expression in these cells was not because of a reversion to a more immature state (Figure 5D, d, j). In the ribs Runx2 was expressed predominantly in osteoblasts, the perichondrium flanking the hypertrophic chondrocytes, and in prehypertrophic chondrocytes (Figure 5D, f). Expression was low in hypertrophic chondrocytes. There was no significant change in Runx2 expression in prehypertrophic and hypertrophic chondrocytes in Col10a1-Cre;Z/Sox9 mice (Figure 5D, l). Collectively, these data suggest that SOX9 negatively regulates Col10a1 gene expression independent of Runx2.
The specificity of SOX protein action is known to be achieved through interaction with cell-specific partners [23], [24]. We questioned whether concomitant transactivation of Col2a1 and repression of Col10a1 by SOX9 in proliferating chondrocytes could be mediated by different combinations of cofactors. ChIP assays in E13.5 mouse limb chondrocytes or CCL cells revealed similar interactions of TRAP230/MED12, a mediator of SOX9 activity [25], and of TRPS1, a GLI3-interacting repressor [26], [27], with both the Col10a1 element A and the Col2a1 enhancer (Figure 6A, upper panel). On the other hand, the transcriptional co-repressor, histone deacetylase HDAC4 [28] immunoprecipitated neither element. GLI1, GLI2 and GLI3 are effectors of Hh signaling which controls chondrocyte proliferation and maturation [29]. GLI1 is a transactivator expressed in proliferating chondrocytes and perichondrial tissue flanking the prehypertrophic and hypertrophic zones [30] whereas GLI2 and GLI3 can act as repressors and are predominantly expressed in non-hypertrophic chondrocytes and are down-regulated in hypertrophic chondrocytes [29], [31]. Since there is a conserved GLI-binding site near the SOX9/TG-rich motif in the same footprint block P3 (Figure 1C), we examined whether GLI1, GLI2 and GLI3 can interact with the element A. Strikingly, while SOX9 bound to both the Col10a1 element A and Col2a1 enhancer, GLI2 and GLI3 associated with only Col10a1 element A (Figure 6A, lower panel). GLI3 interacted the most with element A, while GLI1 interaction was much less. Quantitative ChIP assays confirmed the preferential interaction (Figure 6B). From these results we hypothesized that GLI proteins may repress Col10a1 expression. To test this in vivo, we examined the impact on transgene expression of mutating the GLI-binding site in element A (Col10Flag-EΔ3). Consistent with our hypothesis, the majority of fetuses (7 out of 10) expressing Col10Flag-EΔ3 showed distinct islands of transgene misexpression in non-hypertrophic chondrocytes (Figure 6C, e, f). Mutating all three sites (GLI, SOX9, TG-rich) in the transgene (Col10Flag-EΔ4) did not restrict the expansion of the expression domain to proliferating chondrocytes in all the expressing transgenic fetuses obtained (Figure 6C, i, j). Indeed in the majority of these expressing fetuses (3 out of 5), transgene expression extended throughout the entire cartilage zones. Thus mutation of the GLI site alone had a similar derepressing effect as mutating the SOX9/TG-rich motif and mutating all the motifs did not restrict expression but resulted in more extensive mis-expression. This is consistent with a model whereby SOX9 and GLI act cooperatively to repress Col10a1 transcription.
To assess whether the cooperation of SOX9 and GLI2/3 is a potential common mechanism for restricted or preferential gene expression in hypertrophic chondrocytes, we searched in silico for this configuration of binding sites in genes, other than Col10a1, that have strong and specific up-regulation in hypertrophic chondrocytes (HC genes) in the growth plate. Six of 11 HC genes analyzed, namely Col10a1, Bmp2, Hdac4, Mef2c, Runx2, and Sox4, possess the linked SOX9 and GLI sites (<100 nt spacing) in the inter- or intragenic conserved non-coding regions (Figure 7A and Figure S3). In contrast, these sites were absent from most of the genes tested (12 out of 14) that were expressed in proliferating but not (or down-regulated) in hypertrophic chondrocytes (PC genes). These include known SOX9 targets: Col2a1, Col9a1, Col11a2, Acan, and Mia1 (see Figure 7A legend for all negative genes). The exceptions were Sox5 and Sox6 (Figure 7A and Figure S3). To investigate whether the over-representation of linked SOX9-GLI sites in the HC genes but not the PC genes occurs by chance, we performed a hypergeometric test to calculate the probability of finding 6 or more SOX9-GLI site-containing genes out of 11 genes randomly sampled from the mouse genome. For the HC genes, the results showed that the occurrence of 6 or more genes with associated conserved SOX9-GLI sites is unlikely to occur by chance (p = 0.0000201) (Figure 7B). For the PC genes, the p-value was 0.17, which is comparable to random occurrence. Furthermore the frequency of the presence of SOX9-GLI sites for HC genes (6/11) was significantly higher than that for PC genes (2/14) (Fisher's test p = 0.043, one tailed). This suggests that the linked SOX9-GLI sites are preferentially associated with the HC genes.
The positive and negative mechanisms mediating the stage-specific transcription of genes within the growth plate are not well defined, partly because of the difficulty in distinguishing direct effects on transcription from the consequences of abnormal differentiation. In this study we have exploited the specificity of Col10a1 expression in hypertrophic chondrocytes and the fact that manipulating its expression in vivo has no overt effect on differentiation, to dissect these transcriptional controls. We provide new insight into how differentiation stage-specific gene expression is achieved in the growth plate, presenting in vitro and in vivo evidence that SOX9, in addition to its known role as a transactivator of many genes preferentially expressed in non-hypertrophic chondrocytes, such as Col2a1, directly represses expression of Col10a1 at a stage prior to the onset of hypertrophy and subsequently in proliferating chondrocytes. This discovery extends our understanding of the mechanisms by which SOX9 controls chondrocyte differentiation phase-specific gene expression.
We have identified a conserved regulatory sequence, element A, that acts as an enhancer of Col10a1 expression in both cultured cells and in vivo. This element contains a SOX9 binding sequence that, when bound by SOX9, represses Col10a1 expression in immature and proliferating chondrocytes. Since Sox9 is expressed in non-hypertrophic chondrocytes but not in hypertrophic chondrocytes, this repressive action of SOX9 restricts Col10a1 expression to hypertrophic chondrocytes.
SOX9 has been proposed to direct chondrogenic fate in osteo-chondroprogenitor cells in part by interacting with RUNX2 [32], [33]. SOX9 may inhibit chondrocyte hypertrophy in part via activation of Bapx1 which represses Runx2 [34], [35]. Previous in vitro and in vivo studies suggest that Col10a1 expression is regulated positively by Mef2c, Runx2/Cbfa1, and AP-1 members, which are expressed in hypertrophic chondrocytes [19], [36]–[38]. RUNX2 has been shown to directly regulate the expression of Col10a1 [37]. The element A that we identified contains no conserved consensus RUNX site. The RUNX2 site revealed by Zheng et al. [37] is located within a poorly conserved region outside the element. Our data showed that the ectopic expression of Col10a1 transgene in non-hypertrophic chondrocytes does not require co-expression of Runx2. In addition, RUNX2 is not expressed in the costal hypertrophic chondrocytes and cultured hypertrophic chondrocytes MCTs (which is derived from costal cartilage), where Col10a1 expression is strong. Although real-time PCR showed levels of Col10a1 was markedly reduced in P1 Runx2-null mice [37], hypertrophic chondrocytes with strong Col10a1 expression do develop in many cartilages in Runx2 null fetuses [39], [40]. Collectively existing data suggest that RUNX2 together with other factors regulate Col10a1 in vivo via promoting chondrocyte hypertrophy or otherwise functions to initiate a cascade of regulatory pathways that sustain Col10a1 expression in hypertrophic chondrocytes.
Previous in vitro studies in chicken have suggested that a combined action of positive and negative DNA elements may contribute to the hypertrophic chondrocyte-specific expression of Col10a1 [41], [42]; however, these chick Col10a1 elements are not conserved in mammals. The enhancer element we identified is highly conserved in mammals, but not in chicken, which agrees with previous data [43]. This suggests that in both mammals and chicken, Col10a1 transcription is restricted to hypertrophic chondrocytes by repression, though by different cis-acting elements. In the chicken, this repression may extend to non-chondrogenic cell types [41]. We found no evidence to support such a mechanism in the mouse since when we abolished the interaction of SOX9 with the repressive element, we observed no ectopic Col10a1 expression in non-chondrogenic cells.
Consistent with a role for SOX9 in repressing Col10a1 in vivo, we have shown in Col10a1-Cre;Z-Sox9 mice, that activation of Sox9 expression in hypertrophic chondrocytes in costal cartilage caused down-regulation of Col10a1. It is also notable that in a recent report where Sox9 was over-expressed in hypertrophic chondrocytes in transgenic mice, expression of Col10a1 and the BAC-Col10-Sox9 transgene appeared reduced in the hypertrophic chondrocytes [44]. In the same report Sox9 knockdown in cultured chondrocytes did not affect Col10a1 expression [44]. By contrast Yamashita et al. showed that shRNA knockdown of Sox9 can up-regulate Col10a1 expression in primary costal chondrocyte culture, and that over-expression of Sox9 can down-regulate it [35]. These contradictory results may be related to the incomplete elimination of SOX9 protein and the known dosage dependent requirement for SOX9 action. A role for SOX9 as transcriptional repressor of Col10a1 in non-hypertrophic chondrocytes is consistent with the observation that COL10A1 expression was up-regulated in cartilage isolated from SOX9 haploinsufficient campomelic dysplasia patients [33].
How may SOX9 act both as a transactivator and a repressor in non-hypertrophic chondrocytes? It has been proposed that interactions between specific partner factors stabilize SOX protein binding to DNA and hence regulate target selection [24], thereby determining cell specification, as exemplified by SOX2 in embryonic stem cells and other systems [45]. Such selective cooperation of protein binding partners may mediate the concomitant positive and negative regulation of SOX9 target genes in the same cell that contributes to specification of the differentiation state (Figure 8). As illustrated in the schematic (Figure 8), we propose that SOX9 mediates repression of Col10a1 in proliferating chondrocytes by selective cooperation with GLI factors. Together with its role in activating Col2a1 and other matrix genes, SOX9 therefore plays an important role in maintaining chondrocytes in an immature non-hypertrophic state. RUNX2 by contrast promotes chondrocyte hypertrophy.
SOX9 cannot regulate chondrocyte differentiation appropriately without sonic hedgehog (Shh), which mediates the generation of chondrogenic precursor cells [46], and Indian hedgehog (Ihh), which regulates their proliferation and maturation [47]. GLI proteins are the effectors of Hh signaling. Double knockout mutants indicate that GLI2 has overlapping functions with GLI1 and GLI3 in skeletal and CNS development [48], [49]. Binding of Hh to its receptor, Patched, blocks the proteolytic processing of the GLI transcription factors from active (GLIA) to repressive (GLIR) forms, and the balance between these forms modulates hedgehog target gene expression [50]. In the growth plate, GLI2A can positively regulate chondrocyte hypertrophy and control vascularization of the hypertrophic cartilage in endochondral ossification [29], [51]. GLI3, which acts mainly as a repressor, has been suggested to inhibit chondrocyte hypertrophy [29], [52] and it is interesting that the highest interaction of element A was with GLI3. Mau et al. reported that Gli2/3 null mutations altered the expression domain of collagen X, but it was not possible to distinguish whether this was due to a direct effect on Col10a1 transcription or more general perturbation of hypertrophy [29]. How the GLI factors interact with other regulatory factors or genes in the chondrocyte differentiation program is not clear. Our results are consistent with cooperation between SOX9 and the Hh signaling pathway and suggest that SOX9 acts in synergy with GLI2 and GLI3, probably their repressive forms GLIR, to repress transcription in chondrocytes. Thus, reduced synergy between GLIR and SOX9 may explain the accelerated chondrocyte hypertrophy seen when Ptch1 is inactivated [53]. Over-representation of the SOX9-GLI paired consensus in a number of genes that are preferentially expressed in hypertrophic chondrocytes and not in proliferating chondrocytes, suggests that SOX9 may use this partnership to repress transcription of several genes in other chondrocyte types. However this partnership may not be the exclusive mechanism by which SOX9 acts to repress expression in chondrocytes.
Hattori et al. have recently shown that SOX9 directly represses Vegfa in cultured primary chondrocytes [44] by interacting with the 5′ untranslated region of the gene. This agrees with our findings of a repressive role of SOX9, however, we found no linked SOX9-GLI binding sites near the Vegfa gene and a recent SOX9 ChIP-on-chip study reported no in vivo interaction in Vegfa exon 1 [54].
A different mode by which SOX9 may repress gene expression in chondrocytes has been proposed by Huang et al [55]. In their model, SOX9 negatively regulates Ccn2 expression in non-hypertrophic chondrocytes via binding to overlapping binding sites for SOX and TCF/LEF, thereby interfering with binding of a TCF/LEF/β-catenin transactivation complex. Reduction of SOX9 upon hypertrophy allows this TCF/LEF/β-catenin complex to activate Ccn2 expression [55]. However, Ccn2 is also expressed in resting zone chondrocytes in the epiphyses of the growth plate and it is not clear why SOX9 does not repress the gene in these cells. We identified a conserved TCF consensus site in Col10a1 element A, but this is unlikely to interfere with SOX9 binding since it is located 59 bp downstream of the functional SOX9-GLI motif, unlike in Ccn2 where the SOX and TCF/LEF sites overlap (Figure 1C). This suggests that the model proposed by Huang et al. does not apply to Col10a1 element A-mediated repression of transcription.
It is also possible that SOX9 and GLI cooperate to activate or repress transcription depending on context. While our data implicate a cooperation of SOX9 with GLI factors in transcriptional repression, this association between SOX9 and GLI may not be restricted to negative regulation. Amano et al. have recently reported that GLI2 cooperates with SOX9 to transactivate the Pthlh gene (also known as PTHrP) in chondrocyte culture without direct binding to the gene [56]. However the expression patterns of Sox9 and Pthlh are mutually exclusive in the developing growth plate, Pthlh being expressed mainly in the perichondrium and only at extremely low levels in proliferating chondrocytes [57], [58]. This contradiction may reflect differences between in vitro assays and regulation in vivo, and it is also unclear whether the expressed GLI2 was processed to a repressor form or not in these cells. The observed stimulation of Pthlh promoter activity in the cultured chondrocytes could therefore be attributable to over-expression of GLI2 which persisted largely as the activated form GLI2A. However this report does raise the possibility of a context dependent SOX9-GLI partnership that mediates either transactivation or repression.
Sox9 has been suggested to act upstream of Sox5 and Sox6 in chondrogenesis [6], [46]. The presence of conserved SOX9–GLI sites in the Sox5 and Sox6 genes suggests their expression in proliferating chondrocytes may be positively controlled via cooperation of SOX9 with GLIA or GLI1, an activator that reinforces GLIA function. Hence, the roles played by SOX9 in transcriptional regulation may be determined by context—partnering with GLIA/GLI1 favours transactivation, with GLIR favours repression. Alternatively, as discussed above, the mode of regulation might depend on whether intermediate factors are present to interfere with the SOX9-GLI interaction. Interestingly, while there is a linked SOX9-GLI motif in Sox6, a conserved TCF site occurs between the SOX9 and GLI sites (Figure S3). Cooperation between SOX9 and TCF/LEF/β-catenin might therefore abrogate cooperative repression by SOX9 and GLI and transactivate Sox6 in proliferating chondrocytes.
Validation of these different modes of cooperative regulation by SOX9 and GLI factors in vivo would require the generation and analyses of compound null or conditional knockout mutants; however, the consequent dysregulation of chondrogenesis and impact on cell survival would make it impossible to distinguish changes in transcriptional control from effects on differentiation. For example, Sox9 is essential for chondrogenesis and Sox9 conditional null chondrocytes undergo apoptosis and as a consequence, hypertrophy with the characteristic activation of Col10a1 expression, fails to occur [4], [7], [59]. Because inactivation of Col10a1 does not disrupt the chondrogenic program, it provides an ideal system and tools to interrogate the transcriptional controls governing specificity of gene expression within the growth plate, independent of changes in chondrocyte differentiation. Important questions to be addressed in future are the identities and diversity of SOX9 partnerships and how the activity of SOX9 and its partners is modulated. In early myogenic differentiation, SOX9 and Smad3 have been reported to prevent premature expression of α-sarcoglycan gene expression by synergistic repression in a transforming growth factor β dependent manner [60] suggesting negative transcriptional regulation by SOX9 and its partners may be more common. The transactivation of Matn1 in chondrocytes by SOX9 can be modulated by combined action of L-SOX5, SOX6 and NFI factors [61]. Whether control of transcription by the SOX9-GLI partnership can be modulated by additional factors is an important question to be addressed in future.
In summary, our study implicates a complex regulatory function for SOX9 whereby it acts with different partners to orchestrate activation and repression of transcription in the chondrogenic differentiation pathway. Mutations in human SOX9 cause the skeletal malformation syndrome campomelic dysplasia which is attributed to the disruption of the chondrogenic differentiation program because of failure to express SOX9 target genes. This interpretation may need to be revised to include inappropriate expression of genes normally repressed by SOX9.
COL10A1 sequences were aligned using LAGAN and PipMaker. The conserved SOX9 binding sites (COL2C1 and COL2C2) [10] were identified by rVISTA. The UCSC Mouse 30-Way Multiz Alignment data and a custom Perl script were used to identify human-mouse perfectly conserved SOX9, GLI, and TCF sites with a maximum spacing of 100 bp in the inter- and intra-genic non-coding regions. Conservation percentage of sequence spanning the SOX9 and GLI sites was calculated based on the number of perfectly matched nucleotides among all the aligned species (mouse, human, chimpanzee, canine, bovine, and opossum). From the 32,120 genes in the mouse genome, the number of genes containing conserved linked SOX9-GLI sites was found and used as the reference frequency of such genes in the genome. Whether the frequency of the presence of conserved SOX9-GLI sites in HC or PC genes exceeded this reference frequency was assessed by the hypergeometric distribution. The difference in the frequency of the presence of SOX9-GLI sites in the HC and PC genes was assessed by the Fisher's exact test.
Hypertrophic chondrocyte cell line MCTs (gift of Véronique Lefebvre [62]) was transfected with pCol10Flag-E, pSG-Sox9 (gift of Peter Koopman), or pSG5 expression vector using Fugene 6 (Roche). Expression of the exogenous collagen X from pCol10Flag-E in MCTs cells was examined by immunohistochemistry using anti-Flag M2 antibody (Sigma). CCL (gift of James Kimura [63]), MCT3T3-E1, and COS-1 cells were cultured in DMEM (Invitrogen) containing 10% FCS (Wisent) at 37°C at 5% CO2. MCTs cells were normally cultured at 32°C at 5% CO2 for expansion. Prior to assays, MCTs cells were cultured at 37°C for 1 day to induce growth arrest [62].
A 300 bp DNA fragment within element A, corresponding to −4240 to −3935 bp of the mouse Col10a1, was used as probe. [γ-32P]ATP-labeled probes were incubated with nuclear extracts from CCL, MCTs, MCT3T3-E1, or COS-1 cells in the presence of poly(dA⋅dT) at room temperature, followed by DNase I digestion and denaturing PAGE.
COS-1 nuclear extracts over-expressing SOX9 and RUNX2 were pre-incubated with poly(dI⋅dC) or poly(dG⋅dC) at room temperature, followed by reaction with [γ-32P]ATP-labeled probes with or without the presence of nucleotide competitors, or antibodies for SOX9 (gift of Peter Koopman [64]) and OSF2/RUNX2 (gift of Gerard Karsenty [65]), then subjected to non-denaturing PAGE at room temperature. The sequences of oligonucleotide COL2C1 and OSE2 were as previously described [10], [65].
The mouse Col10a1 element A, a 640 bp-fragment located between −4.2 and −3.6 kb, was cloned at the 5′ end of the pCol10Flag, previously known as FColX [18] consisting of −2070 to +7176 bp mouse Col10a1 genomic sequence, to generate pCol10Flag-E. The SOX9, TG-rich motif, and GLI binding sites in pCol10Flag-E were mutated to generate the single site mutants – respectively pCol10Flag-EΔ1, pCol10Flag-EΔ2, and pCol10Flag-EΔ3. All of these 3 motifs in pCol10Flag-E were mutated to generate pCol10Flag-EΔ4.
A 4.8 kb fragment of mouse genomic DNA (from 82 bp upstream of the start of transcription of Sox9 to 1119 bp downstream of the polyadenylation sequences), including the Sox9 coding region, its two introns and 1.1 kb of 3′ flanking DNA, together with an IRES2-EGFP (Clontech) sequence inserted between the Sox9 stop codon and the polyadenylation site (at +3237 bp), was cloned downstream of the loxP-flanked βgeo/3xpA of the pCall2 vector (gift of Andras Nagy [66]) to create the pZ/Sox9 expression vector. pZ/Sox9 was transfected into 129/SvEv-derived L4 embryonic stem (ES) cells by electroporation and ES clones containing a single copy of the transgene were injected into blastocysts followed by crossing of chimeras with C57BL/6N mice to generate the mouse line Z/Sox9. A mouse line carrying a single copy of pZ/Sox9 was generated which was then crossed with Col10a1-Cre mice [21] to obtain compound mutants.
CCL cell lysates or 13.5 dpc mouse limb tissue lysates were cross-linked followed by lysis and sonication to yield 200–500 bp DNA fragments then immunoprecipitation with antibodies against acetylated histone H3/H4 (Ac-H3, Ac-H4) (Upstate), HDAC4 (Abcam), GLI1, GLI2, GLI3 (all from Santa Cruz), SOX9 (gift of Robin Lovell-Badge [67]), TRAP230/MED12 (gift of Robert Roeder [25]), or TRPS1 (gift of Yasuteru Muragaki [26]). The target elements in Col10a1, Col2a1, or Crygb genes were amplified by real-time or standard PCR.
In-situ hybridization was performed as previously described [9]. The probes used were pRK26 for Col10a1 [68], pWF21 for the Flag sequence [18], p88 for full-length Sox9 (gift of Peter Koopman), pBS-Cbfa1-Sh for full-length Runx2 (gift of Gerard Karsenty), pWF98 for full-length Egfp, pBSII-Cre-frag for Cre (gift of Andrew Groves), and pNJ61 for Col2a1 [9]. Gene expression in cell culture was analyzed by RT-PCR.
For additional details of all experiments, see the Text S1.
|
10.1371/journal.ppat.1004899 | A Ribonucleoprotein Complex Protects the Interleukin-6 mRNA from Degradation by Distinct Herpesviral Endonucleases | During lytic Kaposi’s sarcoma-associated herpesvirus (KSHV) infection, the viral endonuclease SOX promotes widespread degradation of cytoplasmic messenger RNA (mRNA). However, select mRNAs escape SOX-induced cleavage and remain robustly expressed. Prominent among these is interleukin-6 (IL-6), a growth factor important for survival of KSHV infected B cells. IL-6 escape is notable because it contains a sequence within its 3’ untranslated region (UTR) that can confer protection when transferred to a SOX-targeted mRNA, and thus overrides the endonuclease targeting mechanism. Here, we pursued how this protective RNA element functions to maintain mRNA stability. Using affinity purification and mass spectrometry, we identified a set of proteins that associate specifically with the protective element. Although multiple proteins contributed to the escape mechanism, depletion of nucleolin (NCL) most severely impacted protection. NCL was re-localized out of the nucleolus during lytic KSHV infection, and its presence in the cytoplasm was required for protection. After loading onto the IL-6 3’ UTR, NCL differentially bound to the translation initiation factor eIF4H. Disrupting this interaction, or depleting eIF4H, reinstated SOX targeting of the RNA, suggesting that interactions between proteins bound to distant regions of the mRNA are important for escape. Finally, we found that the IL-6 3’ UTR was also protected against mRNA degradation by the vhs endonuclease encoded by herpes simplex virus, despite the fact that its mechanism of mRNA targeting is distinct from SOX. These findings highlight how a multitude of RNA-protein interactions can impact endonuclease targeting, and identify new features underlying the regulation of the IL-6 mRNA.
| During replication of Kaposi’s sarcoma-associated herpesvirus (KSHV), the vast majority of mRNAs in the cytoplasm are cleaved and degraded by the viral nuclease SOX. However, some mRNAs escape this fate, including the transcript encoding the immunoregulatory cytokine IL-6. Here, we discover that this escape is mediated by a group of proteins that associates with a sequence element on the IL-6 mRNA. One of these proteins is nucleolin (NCL), a factor with diverse roles in RNA processing that is frequently co-opted during viral infection. During KSHV replication, a proportion of NCL is redirected from the nucleolar subcompartment of the nucleus into the cytoplasm, where it binds both the IL-6 3’ UTR and a complex of cellular proteins including the translation initiation factor eIF4H. This network of interactions is required for escape from virus-induced degradation. Collectively, these findings reveal novel interplay between the SOX escapees and the cellular mRNA stabilization machinery, and shed light on the complex crosstalk between viruses and hosts over the control of gene expression.
| The posttranscriptional fate of mRNA, including translation, subcellular localization, and stability, is tightly controlled through complex networks of RNA-protein interactions. Many mRNA regulatory elements are located in the 3’ untranslated region (UTR), where they recruit factors that control the levels of the mRNA and its encoded protein both during homeostasis and in response to changes in the cellular environment [1]. In many cases the mechanisms by which these RNA-protein complexes assemble to direct a particular outcome remain unknown, although the best characterized elements are those that promote rapid degradation of mRNAs through recruitment of specific decay enzymes [2–4]. In this regard, mRNA stability is a key point of regulation that is readily engaged during pathogenesis.
Viruses have evolved ways to both circumvent and hijack cellular mRNA decay pathways [5,6]. In particular, gamma-herpesviruses (HVs), including Kaposi’s sarcoma-associated herpesvirus (KSHV) and Epstein-Barr virus (EBV), use RNA degradation as a means to broadly control both cellular and viral gene expression [7–10]. During their lytic replication cycle, gamma-HVs promote widespread acceleration of mRNA decay through the activity of the virally-encoded mRNA-specific endonuclease SOX. SOX internally cleaves cytoplasmic mRNAs in a site-specific manner and promotes their subsequent degradation by the cellular 5’-3’ exonuclease Xrn1 [11]. The importance of SOX-induced mRNA degradation has been demonstrated in vivo using the model virus murine gamma-HV 68 (MHV68), which displays defects in viral trafficking, cell type specific replication, and latency establishment upon introduction of a point mutation in SOX that selectively inhibits its mRNA degradation activity [7,9].
Despite this widespread mRNA degradation, approximately one-third of mRNAs appear to escape SOX-induced cleavage. Although in many instances ‘escape’ is likely a reflection of a secondary transcriptional compensation rather than a failure of SOX to cleave the mRNA, a subset of escapees are truly refractory to SOX targeting [12,13]. These are thought to escape SOX cleavage either because they lack a functional SOX targeting sequence or because they possess specific protective features that render them inaccessible to the viral nuclease. This latter class of escapees is of particular interest, as their characterization could reveal pathways of mRNA regulation that are inaccessible to viral or cellular endonucleases.
Interleukin-6 (IL-6) is an immunomodulatory cytokine important for survival of KSHV-infected B cells [14–16], and its mRNA is directly refractory to SOX-induced decay [17,18]. IL-6 expression spikes during KSHV infection both as a consequence of transcriptional and post-transcriptional control by the virus [17,18]. The ability of the IL-6 mRNA to escape SOX cleavage has been mapped to a specific protective sequence that resides within its 3’ UTR [18]. Fusion of the IL-6 3’ UTR to an mRNA that is normally targeted by SOX renders the mRNA protected, indicating that this RNA element somehow overrides the SOX targeting mechanism. This element recruits a largely undefined complex of cellular proteins, although two components have been identified as HuR and AUF1 and shown to participate in the protective phenotype [18].
Here, we sought to gain a more detailed understanding of how the IL-6 3’ UTR promotes escape from viral endonuclease targeting. Using a ribonucleoprotein (RNP) purification strategy coupled with mass spectrometry, we identified a set of proteins that specifically associate with this protective sequence. Depletion of at least five of these proteins adversely impacts protection, suggesting that the complex as a whole impacts SOX targeting. Among these, nucleolin (NCL) emerged as having the most robust contribution to IL-6 mRNA escape. We found that its re-localization during infection, coupled with specific long-range protein interactions formed only in the context of RNA binding, are prominent components of the protective phenotype. Finally, we demonstrate that the IL-6 3’ UTR also blocks mRNA degradation by the unrelated herpes simplex virus endonuclease vhs, suggesting a protective mechanism that operates across distinct endonuclease targeting strategies.
The majority of cellular mRNAs, as well as reporter mRNAs such as GFP, are endonucleolytically cleaved by the KSHV SOX protein and subsequently degraded. However, the 3’ UTR of the IL-6 mRNA contains a sequence element that protects it against SOX cleavage [18]. Fusion of the IL-6 3’ UTR to a GFP reporter mRNA (GFP-3’IL-6) prevents SOX-induced cleavage, indicating that protection is transferrable. A 100 nucleotide (nt) region of the IL-6 3’ UTR (nt 790–890) is known to be involved in protection [18]. However, deletion of this 100 nt sequence does not eliminate protection from SOX, suggesting that additional flanking sequences might also contribute to escape. To more precisely define the region involved in the escape mechanism, we deleted larger fragments in IL-6 3’UTR, and identified a 200 nt-long sequence encompassing the original element (nt 689–890) that was both necessary and sufficient to confer resistance of the GFP-3’IL-6 fusion to cleavage by SOX (Fig 1A). We refer to this domain as the SOX-resistant element (SRE). RT-qPCR measurements of GFP mRNA levels showed that deletion of the SRE (GFP-IL-6SRE) eliminated protection from SOX-induced decay, whereas fusion of just the 200 nt SRE to GFP (GFP-IL-6 SRE) was sufficient to confer protection against SOX in transfected 293T cells (Fig 1B). These results were confirmed by measuring the half-life of GFP 3’ IL-6, SRE and ΔSRE in the presence or absence of SOX (S1 Fig). As observed previously, removing the SRE from the reporter results in stabilization of the transcript, due to the deletion of portions of AU-rich destabilization elements present in the IL-6 3’ UTR [18].
Sequence elements that impact mRNA stability generally function through the specific recruitment of RNA binding proteins that control message fate. To identify the set of factors specifically associated with the SRE, we applied a recently developed ribonucleoprotein (RNP) purification tool based on the conditional activity of the Csy4 ribonuclease from the bacterial CRISPR antiviral system (Fig 1C) [19]. Briefly, the Csy4 variant H29A/S50C binds extremely tightly (50 pm KD) to a 28 nt CRISPR RNA hairpin, and can be activated to cleave at a precise position in the hairpin in the presence of imidazole [19]. A hairpin-fused RNA segment and its associated RNP complex can therefore be purified by incubation over beads bound by recombinant Csy4 H29A/S50C, released in the presence of imidazole, and subjected to mass spectrometry (MS) to identify each of the bound proteins.
The CRISPR hairpin sequence was thus fused to the IL-6 SRE or, as a control, to an unrelated sequence corresponding to IL-6 coding region similar in size (nt 251 to 450), and the in vitro transcribed RNAs were bound to Csy4-coupled beads. Each fragment was then incubated with lysates from B cells stably infected with KSHV (TREX-BCBL-1) or from 293T cells, as the latter were used for the initial IL-6 SRE characterization experiments and thus contain the necessary cohort of factors required for SRE-mediated protection. After stringent washing, the RNP complexes were released by imidazole treatment and subjected to MS (Fig 1C). Of the 450 proteins identified by MS (S1 Table), 23 were specifically associated with or strongly enriched (at least 7-fold over the control) on the SRE-containing IL-6 RNA from both TREX-BCBL-1 and 293T lysates (Table 1). Each of these 23 proteins had a minimum of 3 peptide hits from the IL-6 SRE RNA purification, and a maximum of 1 peptide hit from the control RNA purification. Both AUF1 and HuR, the two known components of the SRE RNP [18], were re-identified as specific SRE-binding proteins in this manner, indicating that this is a robust methodology for revealing functionally relevant RNA-protein interactions. GO term analysis of this set of SRE-binding proteins revealed seven functional groupings, with a clear enrichment of RNA binding proteins and proteins involved in RNA regulation, as would be expected for factors that control the post-transcriptional fate of an mRNA (Fig 1D and S2 Table).
To determine whether the complex of SRE-binding proteins was involved in the IL-6 escape mechanism, we selected 10 candidates for further analysis based on the robustness of their interaction and their putative or characterized roles in the regulation of RNA stability. These included nucleolin (NCL; the interaction with the most peptide hits), as well as STAU1, hnRNP U, DHX57, and DHX36, IGF2BP1, YTHDC2, NPM1, HNRNPAB and ZC3HAV1. Each factor was individually depleted from 293T cells using specific siRNAs, and the abundance of the GFP-3’IL-6 mRNA in the presence and absence of SOX was measured by RT-qPCR (Fig 2A and 2B). The SRE-containing GFP-3’IL-6 mRNA was protected against SOX-induced degradation in 293T cells transfected with a control nonspecific siRNA (Fig 2A). However, siRNA-mediated depletion of five out of the ten SRE binding proteins significantly decreased the protective effect of the IL-6 3’ UTR (Fig 2A). Out of these five proteins, only NCL knock down resulted in a reduced steady state level of the reporter independently of SOX (S2 Fig), which is not surprising given its known role as a regulator of RNA maturation [20].
It is possible that the effect of these SRE binding protein on IL-6 escape are underestimations of the contribution of each factor towards SRE-mediated protection, as the siRNA treatments resulted in only partial depletion of each endogenous transcript (Fig 2B). However, these data indicate that at least a subset of the SRE-binding proteins we identified by MS are functionally linked to IL-6 escape from degradation by SOX.
The strongest decrease in SRE-mediated protection was observed in cells depleted of NCL, and thus we decided to pursue this interaction in more detail. To confirm the interaction of NCL with the IL-6 SRE in vivo, we immunoprecipitated (IP) endogenous NCL from 293T cells transfected with either GFP-3’IL-6 or GFP-ΔSRE, and performed qRT-PCR to measure the level of co-precipitating RNA. We observed a ~10-fold enrichment of GFP-3’IL-6 over the mock (IgG) IP, but detected no enrichment of the GFP-ΔSRE construct or the negative control RNA ARF1 (an a priori non-NCL target) (Fig 3A and 3B). Thus, NCL exhibits an SRE-dependent interaction with the IL-6 3’ UTR in vivo.
NCL contains four RNA binding domains (RBD) that, when mutated, have been shown to compromise the ability of the protein to bind target RNA [21]. We therefore hypothesized that the RBD should be required for the ability of NCL to potentiate the protective effect of the SRE in complementation assays. To evaluate the importance of this domain in conferring protection from SOX, we first engineered 293T cells to stably express two doxycycline-inducible short hairpin (sh) RNAs targeting nucleolin (293TΔNCL). Doxycycline treatment of these cells resulted in an ~80% reduction of endogenous NCL protein (Fig 3C) and, in agreement with the NCL siRNA-based depletion data, rendered the GFP-3’IL-6 mRNA susceptible to SOX-induced degradation (Fig 3D). We first confirmed that the alterations in RNA abundance upon NCL depletion were due to changes in mRNA stability by measuring the half-life of GFP 3’ IL-6, SRE and ΔSRE in this cell line in the presence or absence of SOX (S3 Fig). We then constructed a mutant version of NCL in which key residues within RBD1 (F347/Y349) and RBD2 (I429/Y431) required for RNA binding were mutated to aspartic acid (NCLmutRBD) [21]. Although transfection of WT NCL into doxycycline-treated 293TΔNCL cells rescued protection of the GFP-3’IL-6 mRNA in the presence of SOX, no protective effect was conferred by transfection of the NCLmutRBD (Fig 3E). Ectopic expression of WT NCL not only rescued the protection phenotype, but also increased the basal levels of GFP expression. As we observed in S2 Fig, NCL depletion decreased GFP mRNA steady state levels, but ectopic expression of NCL rescued this decrease (S4 Fig), likely explaining the increase observed in Fig 3E. We confirmed by Western blotting that both proteins were expressed equivalently (Fig 3F). These observations demonstrate that NCL must bind to the SRE to confer protection against SOX.
NCL is enriched in the nucleolus, but can also be present to a lesser extent in the nucleoplasm, cytoplasm, and at the plasma membrane [22,23]. Cleavage of mRNA by SOX takes place in the cytoplasm [24,25], and thus presumably sufficient cytoplasmic NCL must be present to ensure IL-6 protection during lytic KHSV infection. We monitored endogenous NCL localization in cells latently and lytically infected with KHSV by immunofluorescence assay (IFA) and by subcellular fractionation. First, we performed IFA for NCL in KSHV-positive TREX-BCBL-1 cells that were either latently infected or treated with doxycycline to induce lytic replication. In latently infected TREX-BCBL-1 cells, NCL expression was predominantly nucleolar, in agreement with previous reports [26] (Fig 4A and S1 Video). However, upon lytic reactivation, NCL localization shifted dramatically to the nucleoplasm and to punctate granules within the cytoplasm (Fig 4A and S2 Video).
We also used subcellular fractionation to monitor NCL localization in a second cell type, KSHV-positive iSLK.219 cells [27]. Similar to the TREX-BCBL-1 cells, iSLK.219 cells contain a doxycycline-inducible version of the major KSHV lytic transactivator RTA that enables lytic reactivation. In latently infected iSLK.219 cells, NCL remained almost exclusively nuclear (Fig 4B). However, in lytically reactivated iSLK.219 cells, a proportion of NCL was redistributed into the cytoplasm (Fig 4B). These results indicate that lytic KSHV infection induces re-localization of NCL, including into the cytoplasm where SOX-induced mRNA cleavage takes place.
NCL is a shuttling protein and contains in its N-terminal region a bipartite nuclear localization signal (NLS) [28]. To determine which population of NCL is important for SRE-mediated protection from SOX, we generated an NCL NLS mutant (NCLΔNLS) that was restricted to the cytoplasm, as well as a version of NCL fused to a nuclear retention signal (NRS-NCL) that was restricted to the nucleus (Fig 4C). We verified by Western blot (WB) that these constructs were expressed at similar levels (Fig 4C). It should be noted that although the intensity of the nuclear staining of WT NCL made it difficult to detect the cytoplasmic population by IFA, subcellular fractionation experiments confirmed that in 293T cells both endogenous and transfected WT NCL could be detected in both compartments (S5 Fig). We next evaluated the ability of each protein to rescue SRE-mediated escape from SOX degradation in the 293TΔNCL cell line. Both WT NCL and NCLΔNLS rescued levels of the GFP-3’IL-6 mRNA in the presence of SOX (Fig 4D). However, NRS-NCL was unable to rescue GFP3’IL-6 mRNA from SOX degradation (Fig 4D) Taken together, these results demonstrate that cytoplasmic NCL is involved in SRE-mediated protection.
Finally, we analyzed whether depletion of NCL from iSLK.219 cells by siRNA treatment impacted IL-6 mRNA levels and/or the lytic KSHV lifecycle. Indeed, NCL knockdown reduced the abundance of IL-6 mRNA during the KSHV lytic cycle compared to cells treated with control siRNAs (S6A Fig). This effect is not as robust as the effects of NCL depletion in the context of SOX transfection, likely because during infection KSHV has several other mechanisms to transcriptionally and post-transcriptionally increase IL-6 abundance [29,30]. We also detected a robust impairment of expression of KSHV late gene expression as measured by K8.1 levels, as well as a corresponding failure of the infected cells to produce progeny virions in supernatant transfer assays (S6B and S6C Fig). These results are supportive of a role for NCL in IL-6 protection in the context of KSHV infection, although NCL clearly plays additional crucial roles in the KSHV lifecycle.
To determine whether the location of the SRE might impact protection, we tested the effect of moving the SRE from the 3’ UTR to the 5’ UTR on the GFP reporter (GFP-5’ SRE). Unlike the GFP-3’ SRE mRNA, the GFP-5’ SRE mRNA was degraded in SOX-expressing cells, indicating that SRE positioning is important (Fig 5A). This could be explained if ribosome scanning through the 5’ UTR disrupted NCL binding to the SRE, and/or if NCL positioning on an mRNA impacted its interactions with other mRNA-bound proteins to potentiate protection from SOX. We tested the first part of this hypothesis by measuring the efficiency with which NCL associated with the GFP-5’ SRE compared to GFP 3’ SRE. NCL displayed significantly reduced binding to the GFP-5’ SRE mRNA in RNA IPs, suggesting that the SRE RNP does not assemble efficiently if located in the 5’ UTR (Fig 5B).
We next pursued the idea that once recruited to the SRE, NCL-induced protection from SOX may involve its interaction with other cellular proteins. Previously described protein interactions of NCL largely occur through its C-terminal arginine-glycine repeat (RGG) region [23,31–35]. NCLΔRGG failed to protect the GFP-3’IL-6 mRNA from SOX in the doxycycline-treated 293TΔNCL cell line (Fig 5C and 5D), indicating a role for protein interactions in the SRE escape function of NCL. It should be noted that although the ΔRGG mutant is expressed at lower levels, increasing the amount transfected to produce levels matching those of WT did not rescue the protection phenotype (S7 Fig). Because the 5’ cap and 3’ poly(A) tail are defining mRNA features and positionally fixed, we hypothesized that interactions with one or more factors bound to these elements might impact the SRE-related function of NCL. We further reasoned that protein-protein interactions related to SOX resistance might be enhanced specifically during lytic KSHV infection, when NCL relocalization occurs. Using a targeted approach, we therefore searched for mRNA-associated factors that displayed selective or enhanced binding to NCL during lytic KSHV infection using co-immunoprecipitation (co-IP) assays. We found the cap-associated translation initiation factor eIF4H to selectively immunoprecipitate NCL from lytically but not latently infected iSLK219 cells (Fig 6A). This enrichment appeared specific to eIF4H, as we detected no differential interaction profile for NCL with additional mRNA cap- or tail-bound proteins including eIF4G, eIF4E, eIF4B and PABPC (S8 Fig). The interaction between NCL and eIF4H was disrupted when the lysates were treated with RNase (Fig 6A), in agreement with the idea that these proteins are not normally stably associated, but are brought together in the context of mRNA-bound NCL via a long-range interaction. Furthermore, the NCLΔRGG mutant failed to bind eIF4H in co-IP assays (Fig 6B), while still being able to bind specifically to a SRE containing reporter (Fig 6C), suggesting that the failure of this mutant to protect SRE-containing mRNAs from SOX may be due, at least in part, to its inability to bind eIF4H.
We reasoned that if the NCL-eIF4H interaction played a role in the escape of SRE-containing mRNAs from SOX cleavage, then depletion of eIF4H should decrease the efficiency of escape. Indeed, similar to our results with NCL, siRNA-mediated depletion of eIF4H rendered the GFP-3’IL-6 susceptible to degradation by SOX (Fig 6D). Depletion of eIF4H did not affect the expression of SOX, NCL, or XrnI, arguing against a generalized impediment of protein translation in these experiments (S9 Fig). This was not unexpected, given that an increasing number of translation factors previously thought to have generalized roles in translation, including the eIF4F complex, have instead been shown to be selectively required for only specific types of host mRNAs [36,37].
Viruses that promote widespread degradation of mRNA generally do so by encoding endonucleases or endonuclease-activating proteins [38–40]. We therefore explored the possibility that the IL-6 SRE might also confer protection against additional viral endonucleases. Herpes simplex virus (HSV) encodes an mRNA-targeting endonuclease (vhs) that, while not homologous to KSHV SOX, degrades most mRNAs during infection [41–44]. To test whether the SRE conferred protection against HSV-1 vhs, we measured by RT-qPCR the ability of vhs to degrade the GFP reporter mRNA fused to the IL-6 3’ UTR versus the control IL-6 5’ UTR. Intriguingly, the IL-6 3’ UTR as well as IL-6 SRE conferred complete protection from vhs, while the GFP mRNA containing the IL-6 5’UTR or ΔSRE was readily degraded (Fig 7A). To determine whether protection from vhs-mediated cleavage required NCL, we co-expressed vhs and GFP-3’IL-6 in the 293TΔNCL cell line. Upon Dox treatment to deplete NCL, GFP-3’IL-6 was no longer protected from vhs (Fig 7B). Thus, the SRE-containing IL-6 3’UTR can block mRNA cleavage by at least two non-homologous endonucleases via an NCL-dependent mechanism.
RNA degradation rates are heavily impacted by the cohort of proteins associated with each transcript, and here we define an RNP complex that inhibits viral endonuclease targeting. Unlike the majority of mRNAs in the cytoplasm that are degraded by the SOX endonuclease during lytic KSHV infection, the IL-6 mRNA is strongly induced and directly refractory to cleavage by SOX [8,17]. Although other mRNAs can also escape cleavage, IL-6 is the only mRNA known to escape via a dominant protective mechanism. We found that the 200 nt IL-6 protective element directs assembly of a large RNP complex, of which five components associated with the regulation of mRNA stability have now been shown to contribute to escape from SOX [18]. Notably, this escape element also functions to guard mRNAs against the HSV-1 vhs nuclease, despite the fact that vhs and SOX are unrelated and cleave mRNAs at distinct sites [11,45]. This key observation suggests that the underlying mechanism of escape does not involve a SOX-specific feature but instead must involve the general accessibility of the mRNA to these (and perhaps other) cytoplasmic endonucleases.
Although not homologous, SOX and vhs do share certain features in their RNA targeting strategies. Both proteins selectively cleave mRNA but not RNAs transcribed by RNA Polymerase I or III [44,46,47]. Furthermore, both proteins can target mRNAs prior to recruitment of the 40S ribosomal subunit, suggesting that ongoing translation of the target mRNA is not required for cleavage [46]. However, translation may nonetheless play some role in targeting, as vhs cleavage sites can be altered by mutating the target mRNA start codon or Kozak consensus context, and SOX can cleave mRNAs in polysomes [11,40,48]. While the factor(s) involved in recruiting SOX to its mRNA targets remain unknown, vhs recruitment involves interactions with the translation initiation factors eIF4H and eIF4AI/II [49,50]. Once brought to the mRNA, vhs tends to cleave in a cap-proximal manner in the 5’ UTR or near the start codon [48,51], whereas SOX requires a specific recognition sequence that can be located at sites far downstream from the cap [11,40,48]. The observation that SOX and vhs cleave mRNAs at distinct sites suggests that there must be differences in their mechanisms of targeting. In this regard, SRE-mediated protection could occur by blocking a factor required for both SOX and vhs recruitment to mRNAs. It is notable that eIF4H has been shown to bind vhs and help direct it to mRNAs [48,49,52,53]. However, unlike with vhs, no interaction between SOX and eIF4H have been reported in the literature, arguing against eIF4H accessibility as the feature underlying endonuclease escape for both vhs and SOX. The escape mechanism for these viral proteins is therefore expected to be different, although it is possible that an additional factor required for both SOX and vhs recruitment is occluded or displaced by the SRE RNP. Alternatively, the SRE may direct localization of the IL-6 mRNA into SOX- and vhs-inaccessible sites in the cytoplasm.
Although the fate of IL-6 during HSV-1 infection remains unknown, other specific mRNAs have been shown to escape degradation by vhs, some of which contain AU-rich elements in their 3’ UTR [54,55]. This has been studied most intensively for the IEX-1 mRNA, however, at present, reports differ as to what form of the IEX-1 mRNA is stabilized during infection and the precise role of vhs in altering IEX-1 mRNA decay [55–59]. Both GADD45β and the ARE-containing TTP mRNAs have also been shown to be directly refractory to vhs-mediated decay and are up-regulated at the protein level during HSV-1 infection [54,55,57]. Like IEX-1 and TTP, the IL-6 3’ UTR contains an ARE, which overlaps with the SRE [17]. While it is perhaps notable that the best-studied herpesviral escapees contain AREs, ARE-bearing mRNAs are not enriched in the overall pool of SOX escapees and the majority of ARE mRNAs are susceptible to degradation by SOX [13], arguing against this being the feature driving the IL-6 protective mechanism.
Among the proteins identified to selectively bind the SRE, NCL was the most potent modulator of escape. NCL has diverse roles in RNA biogenesis and has been previously shown to affect mRNA turnover [20]. Although the effects of NCL differ depending on the target transcript, it has been reported to interact with the 3′UTR of numerous mRNAs and enhance their stability. Known targets include amyloid precursor protein (APP), β-globin, Bcl-2, Bcl-xL, interleukin 2 (IL-2), and the growth arrest- and DNA damage-inducible 45 (GADD45A) [60–63]. NCL stabilization of mRNAs has also been linked to its ability to bind AREs [61,64,65]. One notable example is NCL-mediated stabilization of the GADD45A mRNA which, similar to IL-6, occurs via the binding of NCL at its 3’UTR [63]. GADD45A is one of the transcripts identified by RNAseq as being refractory to SOX cleavage [13] and to be up-regulated during HSV-1 infection [66]. Similar to what we observed during KSHV infection, stabilization of GADD45A is associated with the re-localization of NCL from dense nuclear foci to the nucleoplasm and cytoplasm upon arsenic-induced stress [63]. Thus, redistribution of NCL may contribute to stabilization of multiple stress-responsive mRNAs upon chemical or viral insults.
Although NCL clearly contributes to SRE function, the composition of the 200 nt SRE complicates the ability to make a direct and selective link between NCL and SOX escape. For example, a portion of the SRE contains AU-rich sequences, which are elements with established roles in the destabilization of many labile mRNAs [67]. Furthermore, the facts that NCL has been implicated in numerous aspects of mRNA biology and can impact the abundance of non-SRE mRNAs (such as GFP) highlight the broad effects this protein has in cells. Thus, it is possible that depletion of NCL has secondary effects on mRNA accumulation that indirectly influence the stability of IL-6 in SOX-expressing cells. Nonetheless, our observations that NCL binds specifically to the SRE and that this binding in the cytosol is required for protection against SOX suggest that at least some aspects of SRE-mediated escape from SOX are connected to the presence of NCL in the SRE-bound protein complex. In this regard, NCL-induced mRNA stabilization often involves its interaction with other RBPs, including HuR [63] and AUF1 [68], both of which are important for IL-6 escape from SOX degradation [18]. Given that NCL is a highly connected protein [23,31–35], it could act as a hub to assemble larger protein complexes, perhaps explaining its potent role in the escape mechanism. For example, it interacts with several of the identified SRE-binding proteins, including NPM1 [69,70] and STAU1 [71,72].
Here, we describe a novel interaction between NCL and the helicase accessory factor eIF4H. The fact that the NCL-eIF4H interaction selectively occurs during lytic but not latent KSHV infection suggests that this interaction is facilitated by NCL relocalization to the cytoplasm, although infection could also alter the translational requirements for eIF4H. The RNA-dependent nature of the NCL-eIF4H interaction indicates that these proteins associate in the context of mRNA-bound NCL, rather than freely in the cytoplasm, and demonstrates the importance of long-range interactions in mediating protection from SOX. Differential interactions between NCL and other translation-linked proteins were not observed, suggesting that NCL and eIF4H may selectively associate with specific mRNAs. In this regard, it is possible that eIF4H does not play a widespread role in translation but is instead recruited to a subset of mRNAs including IL-6. While not yet explored for eIF4H, it has recently been shown that the cap binding complex eIF4F is preferentially required for the translation of mRNAs that contain 5’ pyrimidine-rich elements [36,37]. Furthermore, eIF4H has a closely related homolog in mammalian cells, eIF4B, which might play a role redundant to that of eIF4H on other mRNAs [73,74]. It will be of interest to determine whether additional NCL-bound mRNAs recruit eIF4H and, additionally, whether other NCL and eIF4H-bound mRNAs escape SOX and vhs. Furthermore, it would be interesting to explore whether KSHV infection favors differential expression of the translation initiation complex components and whether this influences viral gene expression and/or escape from viral induced host shutoff.
Cytoplasmic NCL has been shown to be co-opted by a diverse set of viruses, including to help mediate the human respiratory syncytial virus entry, HIV gag complex assembly, and poliovirus virion formation [75–78]. NCL also plays a positive role in HSV-1 infection [75], where, similar to KSHV infection, it is relocalized to the nucleoplasm and cytoplasm [79,80]. Thus, although NCL is directly involved in mediating protection of IL-6, it is likely that its cytoplasmic relocalization during KSHV reactivation plays additional roles in the viral lifecycle, as its depletion also causes strong defects in K8.1 late gene expression and virion production. At present it is difficult to distinguish its role in IL-6 mRNA accumulation during KSHV infection from its additional crucial roles in the viral life cycle. NCL localization has been linked to its phosphorylation state [81,82]. Thus, one possibility is that NCL is phosphorylated by one of the herpesviral protein kinases, although infection may also activate NCL-targeting cellular kinase cascades. Future studies are anticipated to reveal whether the cytoplasmic population of NCL is posttranslationally modified during infection, and whether this facilitates its interactions that form the basis for SRE-mediated protection.
The KSHV-positive B cell line bearing a doxycycline-inducible version of the major lytic transactivator RTA (TREX-BCBL-1) [83] was maintained in RPMI medium (Invitrogen) supplemented with 10% fetal bovine serum (FBS; Invitrogen), 200 μM L-glutamine (Invitrogen), 100 U/ml penicillin/streptomycin (Invitrogen), and 50 μg/ml hygromycin B (Omega Scientific). Lytic reactivation was induced by treatment with 20 ng/ml 2-O-tetradecanoylphorbol-13-acetate (TPA; Sigma), 1 μg/ml doxycycline (BD Biosciences), and 500 ng/ml ionomycin (Fisher Scientific) for 48h. 293T cells (ATCC) were grown in DMEM (Invitrogen) supplemented with 10% FBS. The KHSV-infected renal carcinoma cell line iSLK.219 bearing doxycycline-inducible RTA were grown in DMEM supplemented with 10% FBS [84]. KSHV lytic reactivation of the iSLK.219 cells was induced by the addition of 0.2 μg/ml doxycycline (BD Biosciences) and 110 μg/ml sodium butyrate for 48 h. For supernatant transfer experiments, the supernatant of iSLK.219 cells reactivated or not was collected after 48h, filtered through a. 45uM filter and spinfected onto fresh 293T for 1h at 1500rpm. Cells were then fixed and mounted onto slides to visualize with a confocal microscopy on a Zeiss LSM 710 AxioObserver microscope. 293TΔNCL were generated by lentiviral transduction. Briefly, psPAX2 and pMD2.G lentiviral plasmids were co-transfected with pTRIPZ plasmids encoding Doxyclyclin inducible shRNAs targeting NCL (V2THS_36645 and V2THS_36643 from Open Biosystems, kindly provided by Chih-Wen Peng at Tzu-Chi University). Supernatant containing viral particles was collected 48h later, filtered, complemented with 8 μg/mL polybrene and centrifuged onto target 293T cells.
For DNA transfections, cells were plated and transfected after 24h when 70% confluent using linear PEI (polyethylenimine). For small interfering RNA (siRNA) transfections, 293T cells were reverse transfected in 12-well plates by INTERFERin (Polyplus-Transfection) with 10 μM of siRNAs. siRNAs were obtained from IDT as DsiRNA and sequences are as described in S3 Table. 48h following siRNA transfection, the cells subjected to DNA transfection as indicated.
For time course experiments, half-live were measured by transfecting 293T or 293TΔNCL cells with the indicated plasmids in 6-well plates. The cultures were split after 6 h into 12-well plates and 12 h later treated with 5 μg/mL Actinomycin D (ActD) for the indicated times. The extracted RNAs were subjected to qPCR analysis and GFP mRNA levels were normalized to the level of 18S rRNA.
The full-length IL-6 cDNA in pCMV-SPORT6.1 was obtained from Invitrogen. Sequence numbering for IL-6 refers to Homo sapiens interleukin 6 (interferon, beta 2), mRNA (GenBank accession number BC015511.1). The GFP-IL-6 3’UTR and GFP-IL-6 SRE fusion constructs were described previously [18], and the GFP-IL-6 3’UTR ΔSRE construct was obtained by overlap PCR into the pcDNA3.1 IL-6 3’UTR plasmid cut with BlpI and XbaI with the following primers (primers sequences are described in S4 Table); IL-6 ΔSRE PCR1 and PCR2, forward and reverse. The Csy4 recognition motif was fused to the SRE or IL-6 nucleotide sequence 251–450 by PCR with the Csy4 primers (S4 Table) and cloned into the KpnI and XhoI sites of pcDNA3.1.
NCL was obtained from 293T total cDNA and cloned into the Gateway entry vector pDON207 (Invitrogen) using the following primers (S4 Table): NCL-Forward and Reverse. It was then transferred into the gateway-compatible destination vector pCiNeo-3xFlag to generate Flag-NCL fusions. For the NCL ΔRGG mutant, the same forward primer was used, but with ΔRGG Reverse primer was. Other mutations were introduced with the Quickchange site directed mutagenesis protocol (Agilent) using the following primers: NCL ΔNLS; NCLmutRBD mutant was generated in a two-step process to introduce mutations both in the RNA binding domains 1 and 2 as described in [21]: RBD1 mutating F347 and Y349 into D in NCL RBD1 domain; to generate the final NCLmutRBD, this RBD1 mutant was further mutated at residues I429 and Y431 into D in the RBD2 domain. The final NCLmutRBD mutant thus contains 4 mutations.
Csy4 H29A/S50C was expressed and purified using the same protocol as wild-type Csy4 (generously provided by R. Haurwitz, H.Y. Lee, and J. Doudna) [19,85]. Plasmids expressing the Csy4 RNA binding motif fused to segments of IL-6 were in vitro transcribed using the T7 Maxiscript kit (Ambion). Transcribed RNA (20 μg) was mixed with purified recombinant Csy4 protein (200 pmol) and magnetic beads for 2h in lysis buffer [10 mM HEPES (pH 8.0), 3 mM MgCl2, 5% glycerol, 1 mM dithiothreitol (DTT), 150 mM NaCl, 0.1% octyl β-d-glucopyranoside, 10 mM imidazole, 1× protease inhibitor]. Lysate from TREX-BCBL1 or 293T cells (1 mg) was then added to the beads for 2h, whereupon the beads were washed 7 times with lysis buffer containing 150 to 300 mM NaCl. RNA and its associated cellular proteins were released from the Csy4-bound beads by the addition of 500 mM imidazole for 2h to activate the cleavage activity of Csy4. Eluates were processed, trypsin digested, and concentrated for LC-MS/MS. Digested peptide mixtures were analyzed by LC-MS/MS on a Thermo Scientific Velos Pro ion trap mass spectrometry system equipped with a Proxeon Easy nLC II high pressure liquid chromatography and autosampler system.
Specific protein bands were excised from a gel and subjected to in-gel tryptic digestion. The gel bands were reduced with 10 mM dithiothreitol (Sigma-Aldrich) at 56°C for 1 hour, followed by alkylation with 55 mM iodoacetamide (Sigma) at room temperature in dark for 45 minutes. The samples were then incubated overnight with 100ng trypsin (Promega) at 37°C. The peptides formed from the digestion were extracted using 50% acetronitrile and 5% formic acid, and then re-suspended in 10 μl of 0.1% formic acid in water and analyzed by on-line LC-MS/MS technique. The LC separation was performed using a NanoAcquity UPLC system (Waters) while the MS/MS analysis was performed using a LTQ Orbitrap XL mass spectrometer (Thermo Scientific). During the LC separation step, 0.1% formic acid in water was used as the mobile phase A and 0.1% formic acid in acetonitrile was employed as the mobile phase B. Following the initial equilibration of the column in 98% A /2% B, 5 μL of the sample was injected. A linear gradient was started with 2% B and increased to 25% B in 33 mins followed by an increase to 60% B in the next 12 mins at a flow rate of 400 nL/min. The subsequent MS analysis was performed using a top six data-dependent acquisition. The sequence includes one survey scan in the FT mode in the Orbitrap with mass resolution of 30,000 followed by six CID scans in LTQ, focusing on the first six most intense peptide ion signals whose m/z values were not in the dynamically updated exclusion list and their intensities were over a threshold of 1000 counts. The analytical peak lists were generated from the raw data using an in-house software, PAVA [86]. The MS/MS data were searched against the UniProt database using an in-house search engine Protein Prospector (http://prospector.ucsf.edu/prospector/mshome.htm).
Total RNA was harvested using Zymo RNA extraction columns following the manufacture's manual. cDNAs were synthesized from 1 μg of total RNA using AMV reverse transcriptase (Promega), and used directly for quantitative PCR (qPCR) analysis with the DyNAmo ColorFlash SYBR green qPCR kit (Thermo Scientific). Signals obtained by qPCR were normalized to 18S.
Cells were lysed in low-salt lysis buffer [NaCl 150mM, NP-40 0.5%, Tris pH8 50mM, DTT 1mM, protease inhibitor cocktail] and protein concentrations were determined by Bradford assay. Equivalent quantities of each sample were incubated overnight with the indicated antibody, and then with G-coupled magnetic beads (Life technologies) for 1h. Where indicated, specific beads coupled to antibodies were used (M2 anti-flag beads; Sigma). Beads were washed extensively with lysis buffer. Samples were resuspended in Western blot loading buffer before resolution by SDS-PAGE. Where indicated, RNAse A and T1 were added to the lysates.
Cell lysates were prepared in lysis buffer and quantified by Bradford assay. Equivalent amounts of each sample were resolved by SDS-PAGE and Western blotted with the following antibodies: Rabbit anti-NCL (Abcam), Mouse anti-NCL (Santa Cruz), Rabbit anti-eIF4H (Cell signaling). Rabbit anti-Flag (Sigma), Mouse anti hnRNPC1/C2 (Abcam), Rabbit anti-H3 (Cell Signaling), Rabbit anti-Xrn1 (Sigma), Mouse anti-Strep (Qiagen). Primary antibodies were followed by HRP-conjugated goat anti-mouse or goat anti-rabbit secondary antibodies (Southern Biotechnology, 1:5000).
293T or TREX-BCBL1 cells were grown on coverslips, and fixed in 4% formaldehyde for 20 min at room temperature. Cells were then permeabilized in 1% Triton-X-100 and 0.1% sodium citrate in PBS for 10 min, saturated in BSA for 30 min and incubated with the indicated antibodies. After 1h, coverslips were washed in PBS and incubated with AlexaFluor594 or AlexaFluor488 secondary antibodies at 1:1500 (Invitrogen). Coverslips were washed again in PBS and mounted in DAPI-containing Vectashield mounting medium (VectorLabs) to stain cell nuclei before visualization by confocal microscopy on a Zeiss LSM 710 AxioObserver microscope.
All results are expressed as means ± S.E.M. of experiments independently repeated at least three times. Unpaired Student's t test was used to evaluate the statistical difference between samples. Significance was evaluated with pValues as follows: * p<0.1; ** p<0.05; *** p<0.01.
|
10.1371/journal.pgen.1004680 | The Chromosomal Association of the Smc5/6 Complex Depends on Cohesion and Predicts the Level of Sister Chromatid Entanglement | The cohesin complex, which is essential for sister chromatid cohesion and chromosome segregation, also inhibits resolution of sister chromatid intertwinings (SCIs) by the topoisomerase Top2. The cohesin-related Smc5/6 complex (Smc5/6) instead accumulates on chromosomes after Top2 inactivation, known to lead to a buildup of unresolved SCIs. This suggests that cohesin can influence the chromosomal association of Smc5/6 via its role in SCI protection. Using high-resolution ChIP-sequencing, we show that the localization of budding yeast Smc5/6 to duplicated chromosomes indeed depends on sister chromatid cohesion in wild-type and top2-4 cells. Smc5/6 is found to be enriched at cohesin binding sites in the centromere-proximal regions in both cell types, but also along chromosome arms when replication has occurred under Top2-inhibiting conditions. Reactivation of Top2 after replication causes Smc5/6 to dissociate from chromosome arms, supporting the assumption that Smc5/6 associates with a Top2 substrate. It is also demonstrated that the amount of Smc5/6 on chromosomes positively correlates with the level of missegregation in top2-4, and that Smc5/6 promotes segregation of short chromosomes in the mutant. Altogether, this shows that the chromosomal localization of Smc5/6 predicts the presence of the chromatid segregation-inhibiting entities which accumulate in top2-4 mutated cells. These are most likely SCIs, and our results thus indicate that, at least when Top2 is inhibited, Smc5/6 facilitates their resolution.
| When cells divide, sister chromatids have to be segregated away from each other for the daughter cells to obtain a correct set of chromosomes. Using yeast as model organism, we have analyzed the function of the cohesin and the Smc5/6 complexes, which are essential for chromosome segregation. Cohesin is known to hold sister chromatid together until segregation occurs, and our results show that cohesin also controls Smc5/6, which is found to associate to linked chromatids specifically. In line with this, our analysis points to that the chromosomal localization of Smc5/6 is an indicator of the level of entanglement between sister chromatids. When Smc5/6 is non-functional, the resolution of these entanglements is shown to be inhibited, thereby preventing segregation of chromatids. Our results also indicate that DNA entanglements are maintained on chromosomes at specific sites until segregation. In summary, we uncover new functions for cohesin, in regulating when and where Smc5/6 binds to chromosomes, and for the Smc5/6 complex in facilitating the resolution of sister chromatid entanglements.
| In order to maintain chromosome stability, cells need to overcome topological problems caused by the structure of the DNA molecule. One example of such topological problem is DNA supercoiling induced by replication or transcription. Another is sister chromatid intertwinings (SCIs), which is the wrapping of chromatids around each other (Figure 1A and B). If not resolved by topoisomerases, supercoiling inhibits transcription and replication, and SCIs block chromosome segregation. While both type I and type II topoisomerases can resolve supercoils by making transient DNA breaks, the type II variant, called Top2 in the budding yeast Saccharomyces cerevisiae (S. cerevisiae), is main responsible for the resolution of SCIs (Figure 1C). [1]–[3]. If Top2 is rendered non-functional before anaphase, chromosome segregation with unresolved SCIs leads to DNA breakage and cell death [4]–[6]. In addition to presenting an obstacle for segregation, sister chromatid tethering by SCIs has been proposed to contribute to proper segregation by counteracting the force of the mitotic spindle, thereby facilitating chromosome alignment during metaphase [7]. The idea of such a positive function for SCIs was, however, challenged when the cohesin protein Scc1 (also known as Mcd1) was shown to be essential for sister chromatid cohesion [8], [9].
The cohesin complex, with a core consisting of Smc1, Smc3, Scc1 and Scc3, is a so-called Structural Maintenance of Chromosomes (SMC) protein complex. In addition to the four core subunits, the Pds5 protein associates to the complex via Scc1 [10], [11]. When either of the subunits is non-functional, sister chromatids are not held together, and chromosome alignment and segregation fail. Cohesin is loaded onto chromosomes before replication, and localizes to intergenic regions between genes that are transcribed in a convergent manner in S. cerevisiae [9], [12]. Several observations indicate that transcription drives the translocation of cohesin to these regions after initial loading by the Scc2/4 complex at centromeres and other, so far mostly undefined, chromosome arm sites [12]–[14]. In addition to loading, cohesin has to be modified in order to establish cohesion. A key regulator of this process is the acetyltransferase Eco1 (also called Ctf7) [15], [16], which acetylates Smc3 and thereby prevents the Pds5-associated protein Wpl1 (also called Rad61), to destabilize the cohesin-chromosome interaction [17]–[19]. In the absence of Wpl1, Eco1 becomes largely dispensable for cohesion establishment [18], [20], [21]. At anaphase, chromosome segregation is made possible by removal of cohesin from chromosomes by separase, a protease which cleaves the Scc1 subunit [22]. In higher eukaryotes, the proteolytic cleavage is preceded by a specific dissociation of cohesin along chromosome arms in prophase, leaving cohesin only in the centromeric region until anaphase [23], [24]. The identification of cohesin as the main constituent of chromatid cohesion provided an explanation of how sister chromatid cohesion could be maintained without the risk of chromosome breakage, which is inevitably linked to cohesion created by SCIs. A more recent study by Farcas et al. shows, however, that cohesin protects SCIs from resolution by Top2 on circular mini-chromosomes [25], suggesting SCIs could contribute to cohesion.
Intriguingly, the cohesin-related Smc5/6 complex (Smc5/6) has also been connected to Top2 function [26]. Smc5/6 consists of Smc5, Smc6 and six non-SMC proteins (Nse1, Mms21, and Nse3-6), and is best known for its function in DNA repair and recombination (reviewed in [27]). The complex is recruited to DNA breaks in a process dependent on Mre11, and central repair factor which accumulates early at the site of damage [28]. When Smc5/6 is non-functional, unresolved recombination intermediates accumulate between sister chromatids in the repetitive ribosomal DNA in unchallenged cells, and during S-phase repair of induced DNA damage [29]–[32]. Since DNA repair in the absence of proper Smc5/6 function is taken to a step that inactivates the DNA damage checkpoint, the unresolved DNA links will inhibit the subsequent segregation event. Also in meiosis, repair of DNA breaks without Smc5/6 leads to similar formation of unresolved recombination intermediates with following segregation failure [33]–[35]. In addition to this, Smc5/6 appears to have non-repair functions. In S. cerevisiae, Smc5/6 has been proposed to function in replication termination [36], and removal of replication-induced supercoiling [26]. Moreover, Smc6 has been reported to allow full removal of cohesin at anaphase when Top2 function is partially compromised in the fission yeast Schizosaccharomyces pombe (S. pombe) [37]. Concerning the chromosomal association of Smc5/6 in the absence of DNA damage, it is independent of Mre11, but requires the replication process as such, and increases after inactivation of the temperature-sensitive top2-4 allele in S. cerevisiae [26], [28]. This opens for the possibility that the chromosomal association is triggered by the presence of SCIs, or another feature which accumulates in top2-4. Regardless if the complex is recruited to SCIs or recombination structures in the absence of DNA damage, its chromosomal association should require that sister chromatids are in close proximity to one another. This predicts that the levels of Smc5/6 present on the replicated genome should decrease in the absence of cohesion, which leads to a separation of chromatids before anaphase. However, our earlier ChIP-on-chip analysis (Chromatin immunoprecipitation (ChIP), combined with analysis on microarrays) of FLAG-tagged Smc6 indicated that the chromosome binding of Smc5/6 changed into more numerous, but narrower, binding sites in scc1-73 cells [28]. The finding that the chromosomal association of Smc5/6 was not reduced in the absence of cohesin argued against it being triggered by a structure which requires the proximity of sister chromatids. In contrast, the scc2-4 mutation, which inhibits cohesin loading, was shown to reduce the levels of chromosome-bound Smc6. This, together with the aberrant binding pattern of Smc6 in scc1-73 cells, made it difficult to draw a definite conclusion on how cohesin influences the chromosomal association of Smc5/6. Using ChIP-sequencing (ChIP-seq, ChIP combined with DNA sequencing), together with ChIP-qPCR (ChIP combined with quantitative PCR) and in situ immunofluorescence, we now show that Smc5/6 chromosome binding is cohesin-dependent. The majority of the chromosome-bound Smc5/6 also co-localizes with cohesin in the vicinity of centromeres, and specifically accumulates along chromosome arms after Top2 inactivation. Evidence is provided that this accumulation is independent of recombination, DNA breaks and fork stalling. Our results also show that the amount of chromosome-bound Smc5/6 predicts the level of missegregation in top2-4 cells, and that the complex promotes the segregation of short chromosomes in the mutant. Altogether, the presented data suggests that Smc5/6 indicates the presence of SCIs in the duplicated genome, and that the complex promotes their resolution, at least when Top2 is inhibited.
Triggered by the observations that Smc5/6 accumulates on chromosomes in top2-4 mutants [26], and that cohesin is a protector of SCIs [25], we revisited the chromosomal association of S. cerevisiae Smc5/6 using ChIP-seq. This method is more quantitative than ChIP-on-chip, and provides more clearly defined binding sites (Figure 2A). The difference is likely caused by the requirement of additional amplification of the immunoprecipitated DNA in ChIP-on-chip, which increases the risk of false positive signals due to the preferential augmentation of certain DNA molecules. Moreover, 50 base pairs (bp) reads are mapped to a reference genome in ChIP-seq, while the amplified material is hybridized to 25 bases long oligonucleotides, each representing a specific genomic sequence, in ChIP-on-chip. The short length of the oligonucleotides, and the requirement for hybridization as such (the efficiency of which varies from oligonucleotide to oligonucleotide), makes the ChIP-on-chip method less accurate.
In contrast to previous results, the ChIP-seq analysis showed that the levels of Smc6 found on chromosomes were markedly reduced in the scc1-73 mutant after an S-phase at restrictive temperature (Figure 2B). Western blot analysis confirmed that this reduction was not due to a general down-regulation of Smc6-FLAG protein levels in the mutant (Figure S1A). At core centromeres the signal remained high, but at all other specific Smc6 binding sites it was abolished (Figure 2B). The reduction of Smc6 was confirmed by ChIP-qPCR in the scc1-73 mutant (Figures 2C and S2). At arm loci, the amount of Smc6 was reduced to levels similar to those in G1-arrested wild-type cells, reflecting the background level before the complex has associated with chromosomes. The Smc6 signal around centromeres was also significantly reduced but remained at up to one third of the wild-type level (Figure 2C). Thus, the ChIP-qPCR results show that the ChIP-seq data is quantitatively accurate, apart from at core centromeres where the signal is overestimated in ChIP-seq, when few other binding sites are present.
Even though this indicates that Smc5/6 is largely absent from chromosomes in the cohesin mutant, it is possible that the reduction only reflects the spreading of the complex to an even distribution over the chromosomes. Such redistribution would make it undetectable by ChIP-seq and lead to a reduction in the ChIP-qPCR signal. To test if this was the case, immunofluorescence (IF) was utilized to detect the association of HA-tagged Smc6 on chromosome spreads. In scc1-73 cells, the fluorescence signal was reduced towards the levels detected in untagged cells (Figure 2D). As for Smc6-Flag, the signal reduction was not due to lower levels of the Smc6-HA protein (Figure S1B), showing that the chromosomal association, and not only positioning, of Smc5/6 requires a functional cohesin complex.
The reduction of Smc6 binding in scc1-73 mutants indicates that Smc5/6 requires sister chromatid cohesion to associate with chromosomes. To test this further, Smc6 localization was analyzed in other cohesion-disrupting mutants. First, ChIP-seq and ChIP-qPCR analysis confirmed the earlier result that Smc5/6 requires the cohesin-loading protein Scc2 for chromosomal association (Figure 3A and C) [28]. The reduction of Smc6 binding in scc1-73 and scc2-4 mutants as measured by ChIP-qPCR was similar (Figures 2C and 3C), and the reason for the difference previously seen by ChIP-on-chip remains unknown [28]. We also found that binding of Smc6 was prevented in the temperature sensitive pds5-101 mutant (Figure 3A and C). ChIP-seq was also performed on Smc6-FLAG in eco1-1 cells, in which formation of sister chromatid cohesion is inhibited even though cohesin remains bound to the chromatids [15], [16]. The reduction of Smc6 binding (Figure 3A and C) in this mutant therefore shows that the chromosomal association of Smc5/6 requires cohesion, and not merely the presence of cohesin on chromosomes. This was further supported by the observation that Smc6 chromosome binding in eco1-1 cells was increased by deletion of Wpl1 (Rad61) (Figure 3A and C), which restores cohesion [18], [20], [21]. On the other hand, the localization of cohesin remained unchanged in an smc6-56 mutant after an S-phase at restrictive temperature, showing that although cohesin controls Smc5/6, the reverse is not true (Figure 3B). Altogether, this shows that sister chromatids have to be held together for Smc5/6 to bind to the duplicated genome.
To take full advantage of the higher resolution obtained by ChIP-seq as compared to ChIP-on-chip, we reinvestigated the chromosomal association of Smc5/6 during the cell cycle. This confirmed that the complex is mostly absent from chromosomes in G1. Similarly to the Smc6 binding pattern in G2/M-arrested scc1-73 cells, ChIP-seq also revealed an association to the core centromeres in this cell cycle phase (Figure 4A), but ChIP-qPCR analysis showed that the levels are low compared to the binding in G2/M-arrested cells (Figure 2C). As shown before, Smc5/6 is detected at stalled forks in cells arrested in early S-phase by the addition of hydroxyurea (HU) [38], which is a binding pattern that differs from the distribution found after completion of replication (Figure 4B–D). This could indicate that the Smc5/6 is associated with the fork and follows fork progression, and to test this, Smc6 binding was analyzed in cells progressing through S-phase at 18°C. This condition is generally applied to slow down replication and to improve cell cycle synchronization. As in the HU-arrested cells, Smc6 displayed a different binding pattern as compared to in G2/M, but the signals were less well defined, likely due to a lower level of synchronization (Figure 4B–D). Even though this left the question whether Smc5/6 follows the replication fork unanswered, it shows that the binding pattern detected in G2/M is not present in early S-phase.
In addition to this, the following new features were revealed. First, Smc5/6 is absent from chromosomes in telophase cells, arrested through inactivation of the mitotic exit network kinase Cdc15 (Figure 4E), well in line with the dependency on cohesin, which is removed from chromosomes at anaphase onset. Second, robust Smc5/6 binding sites are concentrated around the centromeres in G2/M-arrested cells, and all of these sites are found between convergently transcribed genes and co-localizes with cohesin (Figures 4D and S3). A third new feature of Smc6 chromosomal association was detected when comparing the level of association with the length of each chromosome. Earlier analysis showed that Smc5/6 enrichment per chromosome increased with its length [26], [28]. Due to the new observation that strong Smc5/6 chromosome interaction sites clusters around centromeres, this analysis was repeated focusing on this region. Smc6 enrichment was calculated in a 100 kb region spanning the centromere (Figure S4), and when compared to chromosome length, a positive correlation was confirmed (Figure 4F). In our earlier analysis, we suggested that this binding pattern reflects that SCIs can swivel off chromosome ends [26]. If so, enrichment on each side of the centromere should also correlate to the length of the corresponding chromosome arm. This is because kinetochores are re-attached to microtubules directly after their replication, which should confine SCIs to each individual arm [39]. However, the correlation between Smc6 enrichment in a 50 kb region on either side of the centromere and the length of corresponding chromosome arm is low, arguing against such an interpretation (Figure 4G). On the other hand, the levels of Smc6 in the entire 100 kb region showed a stronger correlation with the distance to the closest telomere, i. e. the length of the shortest chromosome arm (Figure 4H). This shows that the further away from a chromosome end a centromere is positioned, the more Smc5/6 will accumulate in its vicinity.
Having determined that the chromosomal localization of Smc5/6 depends on cohesin and cohesion, the chromosomal binding pattern in top2-4 cells was determined using ChIP-seq and ChIP-qPCR. These analyses showed that Smc6 binding around centromeres in top2-4 was not significantly changed as compared to wild-type cells. However, along chromosome arms, Smc6 was strongly enriched at specific sites (Figures 5A and S3). Such an accumulation of Smc6 was also detected after depletion of Top2 by induced protein degradation, showing that it is was not specific effect of the top2-4 allele (Figure S5). In contrast to top2-4 cells, however, the Smc6 signal was increased at core centromere 9 after Top2 depletion, opening for a functional difference at these sites. The reason for this difference is unknown and here we focus on the increase along chromosome arms, which is common to both conditions. Similar to the binding sites in the pericentromeric region, the new binding sites were mainly found in intergenic regions between convergently oriented genes and co-localized with cohesin in top2-4 cells (Figure 5A, B and E). The binding pattern of Scc1, on the other hand, remained unaltered in top2-4 cells, showing that the change in Smc6 association does not reflect alterations in cohesin's chromosomal localization (Figure 5B). Moreover, the levels of chromosome-bound Smc6, as determined by ChIP-seq and IF, were reduced not only in scc1-73 cells, but also in top2-4 scc1-73 cells after an S-phase under restrictive conditions (Figures 5C and 2D). This reduction was confirmed by ChIP-qPCR (Figure 5D), and as in scc1-73 cells, it was not due to lower Smc6 protein levels in top2-4 scc1-73 cells (Figure S1A and B). This suggests that the chromosomal binding of Smc5/6 in wild-type and top2-4 is due to the same underlying cohesin-dependent mechanism. However, even though the IF analysis showed that the level of chromosome-bound Smc6 was lower in top2-4 scc1-73 than in top2-4 cells, the signal was significantly stronger than in the scc1-73 single mutant (Figure 2D). This, together with the ChIP results, indicates that some Smc5/6 remains on chromosomes in top2-4 scc1-73, but distributes differently from cells with functional cohesin.
Knowing that Top2 is needed for removal of transcription-induced supercoils [2], [40], the accumulation of Smc5/6 in top2-4 could be controlled by transcription alone. To investigate this, ChIP-seq and ChIP-on-chip analysis was performed after 1 hour of Top2 inactivation in G2/M-arrested cells, or after 30 minutes inactivation in a G1-arrest. The results revealed that without passage through S-phase at restrictive conditions, there was no alteration in Smc6 chromosomal interaction pattern as compared to the wild-type binding pattern (Figure 6A and B). This shows that like in wild-type cells, the chromosomal positioning of Smc5/6 is set under replication in the top2-4 mutant.
It is well established that Smc5/6 is recruited to DNA double-strand breaks (DSBs) and facilitates resolution of recombination intermediates [28], [32], [41]. To test whether Smc5/6 chromosome association in top2-4 cells was dependent on these structures, ChIP-seq and ChIP-qPCR analysis of Smc6 was performed on cells lacking RAD52 or MRE11. Deletions of these genes inhibit recombination and Smc5/6 recruitment to DSBs, respectively [28], [42]. The results showed that Smc6 still accumulates on chromosomes when Top2 is inhibited in these mutants, demonstrating that the complex binds chromosomes independently of DNA breaks and recombination in top2-4 cells (Figure 7A and B). This was further supported by western blot analysis of the Rad53 kinase, which is part of the damage cell cycle checkpoint and becomes phosphorylated upon DNA damage [43], [44]. This phosphorylation can be detected as a slower migrating form of Rad53, and this was readily observed after replication inhibition through the addition HU to both wild-type and top2-4 cells (Figure 7C). However, after passage through S-phase in the restrictive temperature without addition of HU, no phosphorylation was detected. This indicates that no DNA damage accumulates upon inhibition of Top2 during S-phase.
Smc5/6 is also known to associate to stalled replication forks [38], and Top2 has been shown to facilitate termination of replication [45]. It is therefore possible that Smc5/6 marks stalled forks that are still present in G2/M-arrested top2-4 cells. To test this, the chromosomal localization of the DNA polymerase epsilon subunit Dpb3 was analyzed. This showed that even though Dpb3 was detected on S-phase chromosomes, it was not found in G2/M-arrested top2-4 cells (Figure 8A and B). Moreover, Smc6 did not accumulate on chromosomes in a helicase rrm3Δ mutant, known to elicit replication fork stalling (Figure 8C) [46]. Finally, to assay directly if replication or recombination intermediates accumulate at Smc5/6 binding sites, two-dimensional gel electrophoresis was performed at two loci displaying abundant Smc5/6 binding in top2-4 cells (Figure 9A). At both loci, replication intermediates were detected in S-phase in wild-type and top2-4 cells, but not in G2/M-arrested cells, when Smc5/6 binding is most abundant (Figure 9B). This shows that the accumulation of Smc5/6 at these loci in top2-4 mutant is not due to the presence of a DNA structure that can be detected by a standard two-dimensional gel electrophoresis assay. In addition, the UPB10-MRPL19 locus was investigated using two-dimensional gel electrophoresis on DNA prepared using a CTAB-extraction method [47]. This method preserves specific X-shaped structures, which have been suggested to be hemicatenated sister chromatids, and in early S-phase, these could be detected at a positive control locus, ARS305, in wild-type and top2-4 cells (Figure 9C). In the end of S-phase, no difference between wild-type and top2-4 cells could be seen at the Smc5/6 binding sites. This shows that it is not an increase in hemicatenane-like structures that causes Smc5/6 to accumulate after Top2 inhibition. Altogether, this indicates that the Smc5/6 binding pattern detected in top2-4 cells is independent of DNA breaks, recombination and replication fork stalling.
So far, the data presented here show that Smc5/6 complex is recruited to a chromosome structure which requires sister chromatids that are held together by cohesin. It also accumulates at cohesin sites along chromosome arms after replication under Top2-inhibting conditions. The structure is not a recombination intermediate, a DNA break, nor a replication fork. Neither does it appear after inactivation of Top2 in G1- or G2/M-arrested cells. Altogether this points to that the chromosomal association of Smc5/6 indicates the presence of a recombination-independent structure, which forms during replication on cohesed sister chromatids, and normally is removed by Top2. To test if Smc5/6 accumulation on chromosomes in top2-4 was sensitive to Top2 activity after the completion of DNA replication, the chromosomal association of Smc6 was investigated after restoration of Top2 function in G2/M-phase. Mutant top2-4 cells were first taken through an S-phase at restrictive temperature, and when arrested in G2/M, the temperature was decreased to permissive during 1 hour before sample preparation. Cell survival experiments suggest that SCIs are removed under these conditions [48], and we confirmed this by showing that the temperature down-shift rescues the segregation of chromosome 5, and removes the accumulation of SCIs on a reporter plasmid (Figure 10A–C). Under these conditions, ChIP-seq and ChIP-qPCR showed that Smc6 dissociates from chromosomes in top2-4 cells to levels similar to those found in wild-type (Figure 10B, D and E). If the temperature instead is maintained during the prolonged G2/M-arrest, Smc6 levels remained high. This indicates that the chromosomal association of Smc5/6 correlates with a segregation-inhibiting structure that can be removed by Top2 after the completion of replication, but persist during a prolonged G2/M-arrest when Top2 is non-functional.
To investigate the correlation of the chromosomal association of Smc5/6 and missegregation in top2-4 further, we analyzed chromosome segregation after inactivation of Top2 in G2/M. Under these conditions Smc5/6 chromosomal association remains at wild-type levels (Figure 6), in contrast to the accumulation of Smc5/6 on chromosomes when Top2 is inactivated from G1 until G2/M (Figure 5). This allows segregation analysis under Top2-inhibiting conditions of chromosomes with either wild-type or increased levels of Smc5/6. In preparation for this analysis, we first investigated chromosome segregation in wild-type and top2-4 cells released from a G1-arrest into restrictive conditions for the mutant. Using a system based on the association of fluorescently labeled tetracycline repressors with multiple repeats of tetracycline operators [9], the centromere- and telomere-proximal regions of a short (chromosome 1), an intermediate (chromosome 5) and a long chromosome (chromosome 4) were observed (see material and methods for details) (Figure 11). All three chromosomes were marked 35 kb away from the centromere and within 100 kb from one of the telomeres. Note that on the short chromosome 1, the centromere and telomere marker is one and the same. To get as detailed a picture of the segregation event as possible, sister chromatid separation was scored in relation to elongation of the mitotic spindle, and segregation was scored in relation to the separation (Figure 11A). Chromatids were logged as separated as soon as two fully separated fluorescent dots were visible, and noted as segregated when these dots were partitioned into mother cell and bud. Both separation and segregation occurred simultaneously at the centromere of all three chromosomes in wild-type cells (Figure 11B–D). Separation and segregation of the telomere-proximal regions of chromosomes 4 and 5 took place later, with the longer being most delayed (Figure 11B–D). This result is expected since segregation starts at the centromeres due to their attachment to the mitotic spindle. In top2-4, chromatid separation and segregation proceeded slower in the pericentromeric regions of all three chromosomes, and the delay was most pronounced on the longest chromosome 4 (Figure 11B–D). Both separation and segregation of the telomere-proximal region of chromosomes 4 and 5, but not 1, were severely impaired. These length-dependent delays are in accordance with the observation that long linear chromosomes break more frequently than short ones in top2-4 cells, which was suggested to reflect the ability of SCIs to swivel off the ends of shorter chromosomes [6].
Having established this, segregation was scored after inactivation of Top2 in G2/M. Again, since chromosome-bound Smc5/6 is maintained at wild-type levels under these conditions, the segregation defects should be less severe than after Top2 inactivation in G1, if the complex indicates the presence of the chromosome segregation-inhibiting structures in top2-4 cells. Moreover, a shorter chromosome is expected to segregate more efficiently than a longer one. We therefore analyzed segregation of telomeric markers on chromosome 4 (long) and 5 (intermediate) after a shift to restrictive conditions for the top2-4 allele in G2/M-arrested cells. This showed that in sharp contrast to the severe segregation defect of both chromosomes when Top2 is inactivated in G1, the partitioning of the intermediate-size chromosome 5 now occurred at close to wild-type levels, while the telomeric marker on the long chromosome 4 still exhibited severely defective segregation (Figure 12A and B). When the experiment was repeated, analyzing a region on chromosome 4 which was located at the same distance from the centromere as the telomeric marker on chromosome 5 (approximately 350 kb from the centromere), an intermediate improvement of segregation was detected, as compared to when Top2 was inactivated in G1 (Figure 12C).
In yet another test of how well Smc5/6 chromosome association correlates with missegregation after Top2 inactivation, segregation was scored in an scc1-73 top2-4 double mutant after a G1-release into restrictive conditions. This analysis was also prompted by the observation that entanglements remains between circular mini-chromosomes in the double mutant [25]. Well in line with the ChIP and IF analyses, which indicate that some, but not all, Smc5/6 on chromosomes is retained but de-localized in scc1-73 top2-4 (Figures 5A, C, D and 2D), these cells displayed an intermediate missegregation phenotype (Figure 11E and F). While centromere-proximal regions of chromosomes 1 and 5 displayed premature separation similar to the scc1-73 single mutant, the telomere proximal site of chromosome 5 was inhibited as in top2-4 cells.
If Smc5/6 accumulates in response to the accumulation of segregation-inhibiting structures in top2-4 mutants, it is expected to execute a function at these sites. To test this, we analyzed segregation of chromosome 1. This chromosome segregates at near to wild-type levels in top2-4 cells despite the occurrence of new Smc6 binding sites along the arm (Figure 11C). In an smc6-56 top2-4 double mutant however, there was a threefold increase in missegregation (Figure 13A). This indicates that the additional Smc5/6 complexes recruited upon Top2 inhibition facilitate resolution of this chromosome. To test if this function is to promote removal of cohesin from mitotic chromosomes, levels of FLAG-tagged Scc1 was analyzed by western blot and ChIP-qPCR. This showed that both total levels and chromosome-associated Scc1 was equally reduced in telophase-arrested smc6-56 top2-4 and wild-type cells (Figure 13B–E).
This investigation was launched to understand why Smc5/6 accumulates on chromosomes under Top2-inhibiting conditions. Based on the current knowledge of both the complex and the topoisomerase this could either be due to the accumulation of SCIs or an increased level of sister chromatid recombination structures. Since Top2 impairment also delays replication termination, there is also a possibility that the accumulation of Smc5/6 is due to remaining forks in the mutant [45]. If recombination and/or SCI are the triggers, a central feature for Smc5/6 chromosome association should be a dependency on the proximity of sister chromatids. Using high-resolution ChIP-seq, ChIP-qPCR and IF in combination with a variety of mutations which disrupt sister chromatid cohesion, we show that this is the case in both wild-type and top2-4 cells (Figures 2 and 5). While it was already established that the cohesin loader Scc2 is needed for Smc5/6 chromosomal association [28], the role of cohesin was more uncertain, making it possible that Scc2 directly loaded Smc5/6 on to chromosomes. However, the here presented results indicate that the absence of chromosome-bound Smc5/6 in scc2-4 cells is due to the lack of cohesion, and not to a direct role of Scc2 in Smc5/6 loading (Figure 3). On a more general level, the results also argue that phenotypes of mutations which disrupt cohesin function are caused by the combined loss of chromosome-bound cohesin and Smc5/6. Mutations that change the localization of cohesin might also influence where Smc5/6 is found on chromosomes. Possibly, Smc5/6 contributes to some of the many functions assigned to cohesin (reviewed in [49]). Importantly, however, while cohesin impairment leads to cohesion loss, inhibition of Smc5/6 only creates minor cohesion defects [50], [51], with replication delays and/or perturbations of chromosome structure and segregation being more common phenotypes. This suggests that the complex regulates a process and/or structure which is specific for tethered sister chromatid pairs.
In addition to reveal that the chromosomal association of Smc5/6 in top2-4 cells is dependent on cohesion (Figure 5), this study shows that there are no signs of unfinished duplication in the mutant at the sites where Smc5/6 accumulates (Figures 8 and 9). This argues against the possibility that Smc5/6 binding is triggered by the presence of remaining replication forks. This is further supported by the persistence of chromosome-bound Smc6 during a prolonged G2/M-arrest (Figure 10D and E), since termination of replication has been shown to be delayed but not prevented in mutants of Top2 [45]. Also, Smc6 does not accumulate on chromosomes in rrm3Δ cells (Figure 8C), in which fork pausing is frequent. It is also unlikely that the trigger for Smc5/6 binding in top2-4 is a DNA break or a recombination structure since the damage checkpoint protein Rad53 remains un-phosphorylated, and the Top2-dependent increase in Smc6 binding is still present in top2-4 mre11Δ and top2-4 rad52 cells (Figure 7). Moreover, there are no signs of recombination intermediates detected by two-dimensional gel electrophoresis in the DNA regions bound by Smc5/6 in top2-4 mutants (Figure 9).
This leaves SCIs as the most likely candidates as triggers for Smc5/6 binding, and the following results argue in favor for this assumption. First, as stated above, the buildup of Smc5/6 in top2-4 cells requires the proximity of chromatids. Second, the accumulation requires that the mutant pass through S-phase under restrictive conditions. After inactivation of Top2 in G1- or G2/M-arrested cells, the levels of Smc5/6 binding are unchanged, i. e. under conditions when Top2 inhibition is expected to perturb transcription only (Figure 6). Third, Smc5/6 dissociates from chromosomes when Top2 function is restored after replication, under conditions when Top2 resolves SCIs (Figure 10). Fourth, the level of Smc5/6 chromosome enrichment correlates to the degree of missegregation in top2-4 cells (Figures 4 and 11). Moreover, inactivation of Top2 in G2/M, which leaves the amount of Smc5/6 binding at wild-type levels, also leads to a lower degree of missegregation than after an S-phase without Top2 function (Figure 12). In addition to this, the observation that Smc5/6 is needed for segregation of short chromosomes in top2-4 cells (Figure 13), reveals yet another functional connection between Smc5/6 and SCIs. Results from our earlier analysis suggest that Smc5/6 facilitates formation of SCIs during replication, at least in top2-4 cells. This function was attributed a role of the complex in facilitating fork rotation, thereby decreasing the level of replication-induced supercoiling [26]. The here presented data suggests that Smc5/6 also is needed for Top2-independent resolution of SCIs when replication has been completed (see below). Whether the replicative and post-replicative functions are functionally connected remains to be determined.
In addition to providing evidence for Smc5/6 being controlled by the presence of SCIs on chromatids, the level of its chromosomal association indicates that it senses replication-induced superhelical tension. It is difficult to envisage another mechanism that would lead to a correlation between levels of chromosome-bound Smc5/6 and the length of the shortest chromosome arm (Figure 4H). In a previous investigation, we proposed that the link between Smc5/6 binding and chromosome length reflected the ability of SCIs to swivel off chromosome ends [26]. But the relatively poor correlation between Smc6 enrichment and the length of each chromosome arm detected in this investigation (Figure 4G) argues against this, since SCI movements are expected to be confined between the microtubule-attached kinetochore and each telomere. We propose instead that the chromosomal association of Smc5/6 reflects the dissolution of replication-induced superhelical stress through rotation of the shortest arm. Such unidirectional dissolution should be possible since kinetochores become unattached from the mitotic spindle during their replication in early S-phase [39], [52], [53]. With increasing length of the shortest arm, the more difficult it will be to rotate, which will lead to higher levels of superhelical stress around the centromere. In addition to this, the chromosomal localization of Smc5/6 has to be promoted by a centromere specific-factor since superhelical tension is expected to reach high levels at centrally located, non-centromeric, regions of chromosomes as well. The specific maintenance of Smc5/6 close to the centromeres after Top2 reactivation in G2/M (Figure 10D) argues that this factor works by preventing Smc5/6 dissociation.
Taken together, the presented results are consistent with a scenario where chromosome-bound Smc5/6 indicates the presence of SCIs in the duplicated genome. Based on the observations that cohesin protects SCIs from Top2-resolution, and that Smc5/6 facilitates their resolution, it is conceivable that SCIs are positioned at Smc5/6-containing cohesin sites. Even though this cannot be formally proven until SCIs are directly observed at these sites, we use the following sections to speculate based on this model and discuss what the distribution of the complex, taken into the context of chromosome segregation in wild-type and top2-4 cells, suggests about SCI dynamics in budding yeast (summarized in Figure 14).
Smc5/6 distribution indicates that SCIs are preferentially found in the vicinity of centromeres in wild-type cells, and accumulate along chromosome arms when Top2 is inactivated during replication (Figures 14 and S3). During chromosome segregation in wild-type cells, the pericentromeric SCIs are removed by Top2, which gain access to its substrates after proteolytic cleavage of cohesin. When Top2 is inactivated from G1 and onwards, SCIs accumulate also along chromosome arms and persist after cohesin cleavage in anaphase. The specific inhibition of segregation of intermediate and long chromosome arms under these conditions suggests that the pulling forces of the mitotic spindle drive SCIs from the centromere towards the ends of the chromosome. This will allow separation of all pericentromeric regions, and passive, Top2-independent, separation of short chromosome arms. If Top2 instead is rendered non-functional in G2/M, only centromere-proximal SCIs remain after cohesin removal, and this lower level of SCIs allows segregation of intermediate-sized chromosomes, and partial separation of central regions of a longer ones (Figure 14). Importantly, based on our observation that missegregation of the short chromosome 1 is increased in the smc6-56 top2-4 mutant as compared to both singles (Figure 13A), Top2-independent SCI resolution appears to be facilitated by Smc5/6 function. Whether the complex achieves this by actively promoting SCI resolution via a separate mechanism and/or by preventing SCIs to be transformed into a structure which cannot be passively resolved over chromosome ends, remains to be established. However, in contrast to S. pombe, Smc5/6 does not appear to facilitate chromosome segregation in the absence of fully functional Top2 by promoting cohesin removal from mitotic chromosomes in S. cerevisiae (Figure 13D and E). This difference might reflect that Top2 inhibition specifically perturbs cohesin removal which occurs independently of Scc1 cleavage [37]. Such a pathway has been reported to exist in fission, but not budding, yeast [54]. Regardless, taking the role of Smc5/6 in the resolution of late recombination intermediates into account, it is possible that recombination structures and SCIs have something in common which allows Smc5/6 to promote their resolution.
In addition to the above, the premature chromatid separation of centromere-proximal regions in top2-4 scc1-73 (Figures 11E and F), and the reduction in Smc5/6 chromosome association (Figure 5), suggest that cohesin does more to SCI dynamics than protecting them from Top2 resolution. If this was not the case, the segregation phenotypes of the double mutant should be identical to that of top2-4 cells, i. e. there should be a delay in segregation at all sites tested. A possible scenario is that cohesin is also needed to prevent SCI mobility along chromosome arms, leading to an even dispersal of SCIs in the top2-4 scc1-73 mutant. Moreover, in the lack of cohesin-imposed constraint, the pulling on the chromosomes by the mitotic spindle would be able to displace SCIs from the centromere-proximal region more readily than in a wild-type background. As a result, regions in the vicinity of centromeres would separate prematurely, while chromosome arm regions on longer chromosomes would remain entangled. On the shorter chromosomes, SCIs would also be passively resolved over chromosome ends more easily. In summary, this leads to a scenario where SCIs are resolved by Top2 decatenation and passive resolution in the scc1-73 mutant, and only by passive resolution in top2-4 scc1-73 cells. This is supported by the IF analysis which shows that there is more Smc6 left on chromosomes in top2-4 scc1-73 than in scc1-73 cells (Figure 2D).
Finally, in the light of the possibility that cohesin acts as a direct protector of SCIs we see two explanations for their preferential accumulation around centromeres in wild-type cells. One possibility is that SCI protection not only depends on cohesin, but also on a centromere-specific factor, as discussed above. The observation that reactivation of Top2 in G2/M allows removal of Smc5/6 from cohesin sites along chromosome arms, but not at centromeres (Figure 10D), argues in favor for this. Another, not mutually exclusive, scenario is that SCIs only form when the topological tension reaches a certain threshold. In wild-type cells this would only occur in the vicinity of centromeres, while in top2-4 cells, in which replication-induced topological tension accumulates due to its function in supercoil relaxation, it would also happen at certain cohesin sites along chromosome arms. If so, chromatid entanglement after Top2 inhibition might not only be caused by lack of SCI resolution as the common view predicts, but also to an increase in SCI formation.
In conclusion, this investigation reveals that cohesin and cohesion are required for the chromosomal association and localization of Smc5/6. It also provides evidence that the chromosomal localization of Smc5/6 indicates the presence of SCIs, and that the complex is needed for their Top2-independent resolution. The localization of Smc5/6 to pericentromeric regions in G2/M-arrested cells thus opens for the possibility that SCI are maintained until anaphase, and therefore could contribute to chromatid cohesion, also on linear chromosomes. Taken together with the observation that the chromosomal localization of Smc5/6 is correlated to the length of the shortest chromosome arm, this leads to the unexpected prediction that replication-induced superhelical stress can influence chromosome segregation via the formation of SCIs.
All strains are of W303 origin (ade2-1 trp1-1 can1-100 leu2-3,112 his3-11,15 ura3-1) RAD5 with the modifications listed in Table S1. Primer sequences used for site directed gene-modifications are available upon request.
Strains used for live cell imaging: To integrate multiple copies of tetracycline operators at other sites than 35 kb away from centromere 5, which is the location of the endogenous ura3-1 gene, ura3-1 was first replaced with the NAT gene, which confers resistance to nourseothricin. The ura3-1 gene was also cloned into the PFA6a-KanMX4 plasmid, which contains the kanamycin resistance gene (KAN). Both ura3-1 and KAN were amplified by PCR using the primers listed in Table S2. The resulting constructs were used in transformations, and correct integration at the chosen genomic sites was controlled by Southern blot. Finally, the TetO plasmid (pWJ1378) containing multiple copies of tetracycline operons and URA3, was integrated at the ura3-1 sites. Correct integration was again controlled by Southern blotting. If not stated otherwise, cultures were grown in YEP medium (1% yeast extract, 2% peptone, 40 µg/ml adenine) supplemented with 2% glucose as carbon source, with the exception of the live cell imaging analysis, see below. For synchronization in G1 and a following release at restrictive temperature, 3 µg/ml α factor mating pheromone (Innovagen) was added every hour for 1.5 generation times. When a complete G1-arrest was achieved, cells were incubated at the restrictive temperature for 30 minutes, unless otherwise stated. For release into a synchronous S-phase, cells were filter-washed by three volumes of pre-heated YEP medium and subsequently resuspended in fresh medium. To achieve a subsequent arrest in the following G2/M, the release medium contained 15 µg/ml nocodazole (Sigma).
Chromatin immunoprecipitation was carried out as previously described [26], [55] with the modification that cells were lysed using a 6870 Freezer/Mill (SPEX, CertiPrep). Briefly, cells were crosslinked by 1% formaldehyde and then washed three times in ice-cold 1× TBS, before being lysed in the Freezer/Mill. Cell lysate was thawed on ice and suspended in lysis buffer. Chromatin was then sheared to a size 300–500 bp by sonication and IP reactions, with anti-FLAG antibody (F1804, Sigma) conjugated to Dynabeads Protein A (Invitrogen), were allowed to proceed over night. After washing and eluting the ChIP fraction from beads, crosslinks were reversed for input and ChIP fractions and DNA was purified. The DNA samples were then processed for sequencing (see below), qPCR or hybridization to microarrays. qPCR was performed using SYBR green (Applied Biosystems) and primers listed in Table S2 on Applied Biosystem 7000 Real-Time PCR System according to the manufacturer's instructions. For ChIP-on-chip, hybridization of ChIP and input fractions to GeneChip S. cerevisiae Tiling 1.0R Array (Affymetrix) was performed as described [26], [55]. BrdU-IP was performed as previously described [55] using monoclonal anti-BrdU antibody (clone Bu 20a, Dako) and Dynabeads Sheep Anti-Mouse IgG (Invitrogen).
DNA from ChIP and WCE fractions was further sheared to an average size of approximately 150 bp by Covaris (Woburn, MA). Samples were then prepared for sequencing according to the manufacture's standard protocol (Applied Biosystems SOLiD Library Preparation protocol) and were sequenced on Applied Biosystems SOLiD platforms (SOLiD3, 4 and 5500) to generate single-end 50 bp reads. Sequenced reads of DNA-seq were aligned to the S. cerevisiae genome obtained from Saccharomyces Genome Database (http://www.yeastgenome.org/) using Bowtie [56], allowing three mismatches in the first 28 bases per read and filtering reads having more than 10 reportable alignments (-n3 -m10 option). Each aligned read was extended to a predicted fragment length of 150 bp. Reads were summed in 10 bp bins along the chromosomes for ChIP and WCE, and further normalized and smoothed as previously described [57], Nakato R., et al, 2013). For the number of total and mapped reads in each sample, see Table S3. Sequence data are available at the Sequence Read Archive (http://www.ncbi.nlm.nih.gov/sra) with the accession number SRP018757. To call peaks for Smc6 and Scc1, we calculated the fold enrichment (ChIP/WCE) for each bin and identified bins which fulfilled following criteria: (1) fold enrichment was more than 2.0; (2) the maximum read intensity in ChIP bins was more than 1; and (3) fold enrichment of no tag sample was less than 1.8. Chromosome arms (Figure 5E) were defined as the whole chromosomes excluding: 25 kb pericentromeric region spanning the centromere; subtelomeric regions (20 kb proximal to each telomere); and long terminal repeats (LTR). LTRs, defined by Saccharomyces Genome Database (http://www.yeastgenome.org/), were excluded from the upstream to the downstream open reading frame neighboring each LTR. The significance of Smc6 peaks clustering around pericentromeric regions (Figure S3) was assessed with the binomial test by assuming that the Smc6 peaks distributed to the whole genome uniformly. The enrichment values of Smc6-FLAG for each chromosome (Figure 4F–H) were calculated by summing up the difference of fold enrichment between Smc6-FLAG and a no tag control experiment in 100 kb regions spanning the centromeres of each chromosome (Figure S4). Detailed information on the sequencing results is found in Table S3.
Mitotic spreads were prepared as described [58] with the exception that 5% Lipsol (Dynalab) was used as a detergent. Wild-type and mutated Smc6-3×HA-expressing cells were arrested in G2/M after a synchronous S-phase at 35° before preparation of spreads. Monoclonal rat-anti-HA (Roche) was used as the primary antibody followed by Cy3-conjugated goat-anti-rat (Invitrogen) to detect Smc6-3×HA on spreads. Each image was acquired under identical exposure conditions using a Leica microscope and 100× objective. Image analysis was carried out in Volocity (Perkin Elmer). Signals from >50 chromosome spreads were quantified using the analysis tools provided by the Volocity software (Perkin Elmer), and background staining in adjacent regions of the same size were subtracted. Box plots were made using standard statistical tools and represent all values measured between the maximum and the minimum. Statistical analysis to measure significance of differences between strains was done using a two-tailed T-test, with Welch's correction, which was used because the two populations compared had unequal variance. P-values greater than or equal to 0.05 were considered insignificant.
If not stated otherwise, cells were grown at 23°C in synthetic medium lacking histidine and uracil supplemented with 2% glucose. For G1-release experiments, cells were first arrested by of alpha factor at a final concentration of 3 µg/ml, and moved to 35°C thirty minutes prior to release. 500 µl of cell suspension was then applied to Concanavalin A (Sigma) coated glass coverslips (∅ 12 mm), and were allowed to settle for 2 minutes. Medium was subsequently removed and 1 ml fresh medium without alpha factor was added. Cells were allowed to settle to the glass surface for another 40 minutes and were finally imaged through the following mitosis at 35°C. For G2-release experiments in Figure 10A, cells were first arrested in G1 as above, and after 30 minutes at 35°C, released into pre-warmed medium containing nocodazole at a final concentration of 15 µg/ml. Cells were then grown for one hour at 35°C and then either moved to 23°C or kept at 35°C for an additional hour prior to release from the G2/M-arrest. For experiments in Figure 6A, top2-4 cells were arrested in G1, released and allowed to grow at 23°C for 90 minutes in medium containing nocodazole to reach a complete G2/M-arrest. The arrest was then maintained at 35°C for one hour prior to release. 500 µl of cell suspension was then put on Concanavalin A (Sigma) coated glass coverslips (∅ 12 mm) and were allowed to settle for 2 minutes. Medium was then removed and 1 ml fresh 23°C or 35°C medium was added as appropriate. Cells were allowed to settle on the glass surface for another 5 minutes and then imaged through the following mitosis at either 23°C or 35°C. For both type of experiments, images consisted of a 7-layer Z-stack, with layers 0.8 µm apart. These were collected every 30 seconds in green (GFP) and red (tdTomato) channels, for a total of 70 minutes. Control experiments using wild-type and recombination-deficient rad52Δ cells showed that this setup left cell cycle progression unperturbed, and is therefore unlikely to introduce any significant DNA damage. The microscope used was Deltavision Spectris (Applied Precision), and acquired images were analyzed using ImageJ (version 1.44i). Automated tracking of spindle length was performed using CellProfiler version r10997 [59]. Briefly, images were segmented for nuclei based on tetR tdTomato fluorescence and each nucleus was tracked over time. Within each nucleus, the EGFP-tubulin structure was segmented and tracked over time. Spindle elongation was considered when the EGFP-tubulin structure exceeded 10 pixels in length, which is equal to 3.18 µm.
Cells containing the plasmid pRS316-URA3 were collected and immediately fixed in ice-cold 70% ethanol. These cells were subsequently pelleted and incubated at 37°C for 30 minutes in 400 µl buffer containing 0.5 mg/ml zymolyase (Seikagaku Biobusiness), 0.9 M sorbitol, 0.1 M EDTA (pH 8.0) and 14 mM β-mercaptoethanol (Sigma). After a second centrifugation, spheroblasts were resuspended in 400 µl of TE buffer and incubated at 65°C for 30 minutes with 90 µl of 270 mM EDTA (pH 8.0), 460 mM Tris-base and 2.3% SDS. Thereafter, 80 µl of 5 M potassium acetate was added, and samples were kept on ice during 60 minutes, subsequently centrifuged for 15 minutes at 13 000 rpm, and finally, the supernatant was collected into new tube. DNA was then precipitated using 1 ml of 100% ethanol, and resuspended in 500 µl of TE buffer. After treatment with 0.1 mg/ml RNaseI at 37°C for 30 minutes, the DNA was precipitated with 2-propanol, washed by 70% ethanol and resuspended in 50 µl of TE buffer. For nicking enzyme treatment, DNA was incubated with Nb.BtsI (New England Biolabs) for 2 hours at 37°C according to manufacturer's protocol. DNA samples were separated by electrophoresis in 0.8% agarose (Lonza) 0.5× TBE gel with 2.7 V/cm for 24 hours. Plasmids were detected by Southern blotting under standard conditions using radioactive probe that was generated by PCR using primer FW (GTTCCAGTTTGGAACAAGAGTC), primer BW (CATTAAGCGCGGCGGG) and pRS316 as template.
Genomic DNA isolation to study replication intermediates was performed according to [60]. Isolation of genomic DNA with CTAB extraction to preserve X-shape structures was performed according to [47]. Digestion was performed using PstI-HF (New England Biolabs) for the loci UBP10-MRPL19 and MPP10-YJR003C, and EcoRI and HindIII (Roche) for ARS305 locus. The DNA was then precipitated by the addition of 2 volumes ethanol containing 0.5 M potassium acetate and incubated at −80°C for 30 minutes. The precipitated DNA was spun down for 15 minutes at 13 000 rpm and washed with 70% ethanol, before being resuspended in loading buffer. The first dimension gel running was run in 0.35% agarose (Melford, Molecular Biology Grade, MB1200) in 1× TBE at 1 V/cm, in room temperature for 24 hours. The second dimension gel running was run in 0.875% agarose (same as above) in 1× TBE with 0.3 µg/ml ethidium bromide at 5 V/cm, at 4°C for 8 hours, with buffer circulation from anode to cathode at 50 ml/min. Specific loci were detected by Southern blotting under standard conditions using radioactive probe that was generated by PCR using primer pairs GTTCGCCAGTCTCCGTTATT and CTGGGATACCCGAATGTGTATG for ARS305; ATGGTGAAGACATCGGCGAAGACA and AGTGGTAGAAGTGGTGGCTGAAGT for UBP10-MRPL19; GCTTCAGCGTATTGTAGCATTT and GCTCGTGGAACCTATCCTTATT for MPP10-YJR003C, with genomic DNA as template.
To detect Rad53, wild-type and top2-4 cells were G1-arrested at permissive temperature (23°C), incubated at restrictive temperature (35°C) for 30 min, before being released into 0,2M HU or 15 µg/ml nocodazole at 35°C for 75 min. Cells were then collected and protein extracted using trichloroacetic acid (TCA)-precipitation. To detect Scc1-FLAG in telophase and G2/M-arrests, cells were G1-arrested as above before being released into media with or without 15 µg/ml nocodazole at 35°C for 2 hours. To detect Smc6-FLAG and -HA in various strains, cells were prepared as in Figure 2. Cells were then collected and protein extracted a glass-bead disruption method [61]with the modifications that 1× PhosSTOP (Roche) was added to the lysis buffer and that after cell lysis, 2 µl of Benzonase nuclease (Novagen 70664) and NaCl to 200 mM final concentration was added and incubated 30 min at 4°C to promote the release of chromatin-bound proteins. Bradford assay was then used to estimate protein concentration and 20 µg of protein was loaded for each sample. For Rad53, Smc6-FLAG and Smc6-HA detection, membranes were cut after the blocking step and the lower part was incubated with anti-beta Actin antibody and the upper part of the membranes were incubated with anti-Rad53, anti-FLAG and anti-HA, respectively. To detect Scc1-FLAG, the membranes were not cut. Instead, the membranes were incubated with anti-FLAG and anti-beta Actin antibody simultaneously. The following antibodies were used for detection: anti-Rad53 (Abcam, ab104232), anti-FLAG (SIGMA, F1894), anti-HA (Roche, clone 3F10) and anti-beta Actin to detect Act1 (Abcam, ab8224).
|
10.1371/journal.pgen.1000875 | Identification of the Regulatory Logic Controlling Salmonella Pathoadaptation by the SsrA-SsrB Two-Component System | Sequence data from the past decade has laid bare the significance of horizontal gene transfer in creating genetic diversity in the bacterial world. Regulatory evolution, in which non-coding DNA is mutated to create new regulatory nodes, also contributes to this diversity to allow niche adaptation and the evolution of pathogenesis. To survive in the host environment, Salmonella enterica uses a type III secretion system and effector proteins, which are activated by the SsrA-SsrB two-component system in response to the host environment. To better understand the phenomenon of regulatory evolution in S. enterica, we defined the SsrB regulon and asked how this transcription factor interacts with the cis-regulatory region of target genes. Using ChIP-on-chip, cDNA hybridization, and comparative genomics analyses, we describe the SsrB-dependent regulon of ancestral and horizontally acquired genes. Further, we used a genetic screen and computational analyses integrating experimental data from S. enterica and sequence data from an orthologous regulatory system in the insect endosymbiont, Sodalis glossinidius, to identify the conserved yet flexible palindrome sequence that defines DNA recognition by SsrB. Mutational analysis of a representative promoter validated this palindrome as the minimal architecture needed for regulatory input by SsrB. These data provide a high-resolution map of a regulatory network and the underlying logic enabling pathogen adaptation to a host.
| All organisms have a means to control gene expression ensuring correct spatiotemporal deployment of gene products. In bacteria, gene control presents a challenge because one species can reside in multiple niches, requiring them to coordinate gene expression with environmental sensing. Also, widespread acquisition of DNA by horizontal gene transfer demands a mechanism to integrate new genes into existing regulatory circuitry. The environmental awareness issue can be controlled using two-component regulatory systems that connect environmental cues to transcription factor activation, whereas the integration problem can be resolved using DNA regulatory evolution to create new regulatory connections between genes. The evolutionary significance of regulatory evolution for host adaptation is not fully known. We studied the convergence of environmental sensing and genetic networks by examining how the Salmonella enterica SsrA-SsrB two-component system, activated in response to host cues, has integrated ancestral and acquired genes into a common regulon. We identified a palindrome as the major element apportioning SsrB on the chromosome. SsrB binding sites have been selected to co-regulate a gene program involved in pathogenic adaptation of Salmonella to its host. In addition, our results indicate that promoter architecture emerging from SsrB-dependent regulatory evolution may support both mutualistic and parasitic bacteria-host relationships.
| Precise gene regulation is crucial to the successful activation and execution of virulence programs for all pathogenic organisms. The acquisition of genes through horizontal gene transfer, a widespread means of bacterial evolution [1], requires a process to integrate new coding sequence into pre-existing regulatory circuitry. Silencing of horizontally-acquired genes by DNA binding proteins like H-NS is one way some incoming genes are initially controlled [2],[3], which can then be subject to regulatory evolution by mutating cis-regulatory operator regions to select for optional gene expression. The eventual promoter architecture selected to deploy virulence genes is often modular and should reflect a design that maximizes organismal fitness while limiting fitness trade-offs and antagonistic pleiotropy [4]. Both simulated [5] and functional experiments [6],[7] show that mutation of cis-regulatory sequences can be rapid, and that plasticity - the degree to which regulatory mutation can perturb the larger gene network - can be well tolerated in bacterial systems.
Promoter architectures that control quantitative traits such as bacterial virulence are, in fact, modular and evolvable [8]. For instance, Salmonella enterica has a multi-faceted pathogenic strategy fine-tuned by several transcriptional regulators. Intracellular survival and persistence of Salmonella requires a type III secretion system (T3SS) encoded in a horizontally-acquired genomic island called Salmonella Pathogenicity Island-2 (SPI-2) [9],[10]. T3SS are complex secretion machines that deliver bacterial effector proteins directly into host cells through an injectisome during infection [11],[12]. Several ancestral regulators control the genes in the SPI-2 genomic island including the two-component systems EnvZ-OmpR and PhoQ-PhoP, and the regulatory protein SlyA [13]. SsrA-SsrB is another two-component regulatory system co-inherited by genetic linkage with the SPI-2 locus that is essential for gene expression in SPI-2 [14]–[16]. SsrA is a sensor kinase activated in the host environment that phosphorylates the SsrB response regulator to create an active transcription factor needed for spatiotemporal control of virulence genes [13],[17]. In the Salmonella genus, the SPI-2 genomic island is found only in pathogenic serotypes of Salmonella enterica that infect warm-blooded animals and is absent from Salmonella bongori, which colonizes cold-blooded animals [18]. It is generally accepted that SPI-2 was acquired by Salmonella enterica after divergence from S. bongori, providing a useful pedigree to study regulatory evolution influenced by SsrB.
We recently demonstrated the evolutionary significance of cis-regulatory mutations for pathoadaptation of Salmonella enterica serovar Typhimurium (S. Typhimurium) to an animal host [19]. Our focus was on SsrB because of its broad conservation among the pathogenic Salmonellae and its essentiality for animal infection, suggesting it coordinates fundamental aspects of Salmonella pathogenesis beyond the SPI-2 genomic island. In this study we investigated how regulatory evolution assimilates horizontally acquired and ancestral genes into the SsrB regulon on a genome-wide scale using an integrated set of experimental methods. Combining our data with previous biochemical work, along with comparative genomic analyses with an orthologous T3SS-encoding genomic island in the tsetse fly endosymbiont, Sodalis glossinidius, we reveal the flexible DNA palindrome that distributes SsrB in the genome to influence transcriptional activation of the SPI-2 T3SS and almost all of its accessory effector proteins. Our data uncovers the principal SsrB circuitry that appears to have been conserved to support multiple bacterial lifestyles, including parasitic and mutualist symbioses.
To begin to understand regulatory evolution and network expansion of the SsrB response regulator, we profiled the transcriptome of an ssrB mutant and compared it to S. Typhimurium wild-type cells grown in an acidic minimal medium that activates the SsrA-SsrB two-component regulatory system [20]. We identified 133 genes that were significantly down-regulated in the ssrB mutant [z <−1.96] [21] (Table 1). This included almost all genes in the SPI-2 genomic island as well as effector genes encoded throughout the genome (Dataset S1). Next, we performed a Clusters of Orthologous Groups (COG) analysis [22] on the 118 genes that had an ortholog in the annotated genome of S. Typhimurium strain LT2 [23]. Among these, 45 genes lacked a functional COG assignment and the 73 remaining genes were distributed among 86 COGs (Figure 1). The majority of functions in the latter groups are in transport, secretion, and trafficking of cellular components in addition to protein and membrane modification.
The SsrA-SsrB system was acquired by horizontal gene transfer into the S. enterica species after divergence from what is now extant S. bongori. As such, S. bongori has evolved in the absence of SsrA-SsrB and its regulatory architecture has not been influenced by it. Orthologous genes ancestral to both species but regulated by SsrB in S. enterica provide evidence for network expansion and regulatory evolution that we previously showed can be mapped to a single cis-input location by using functional and comparative genomics [19]. To expand on this, we used a reciprocal BLAST-based analysis and identified 47 orthologs in S. bongori among the 133 genes whose transcription was down-regulated in an ssrB mutant (Table 1). In ΔssrB cells, the mean fold-change of the orthologous genes (−6.1-fold) was subtler than for the S. enterica-specific gene set (mean −21.2-fold), which included the T3SS and associated effector genes (Dataset S1). We also determined the distribution of down-regulated genes among genomic islands [24], including prophages, pathogenicity islands (SPI-islands) and additional regions of difference (ROD) between S. enterica and S. bongori (Dataset S2). For this we used a BLAST-based comparison of genome-wide synteny between S. Typhimurium and S. bongori and identified 50 ROD that included 17 previously reported SPI-islands and prophages. Of the 133 down-regulated genes identified, 56 were present within genomic islands (Table 1), with a mean change in gene expression of −25.3-fold in ΔssrB cells.
To examine SsrB allocation on the chromosome in vivo, we isolated functional SsrB-DNA interactions using chromatin immunoprecipitation and examined the bound DNA by chip analysis (ChIP-on-chip) using an S. Typhimurium SL1344 array containing 44,021 probes. With this method, we identified 256 significant interaction peaks distributed throughout the genome that were enriched under SsrB-activating conditions and with interaction scores three standard deviations greater than the mean probe score (Figure 2A and Table 1). Of these 256 peaks, 126 (49%) occurred within coding regions of genes (CDS) and 130 (51%) were in intergenic regions (IGR). Given the strong influence of SsrB on horizontally acquired genes (Table 1), we plotted the ChIP-on-chip data against all genomic islands in S. Typhimurium SL1344. From this analysis, 62 of 256 SsrB binding peaks (24.3%) occurred within genomic islands (Figure 2B).
SsrB ChIP peaks were observed upstream of previously identified SsrB regulated genes indicating that our ChIP-on-chip data captured functional interactions (Dataset S3). To generate a consensus set of SsrB regulated genes, we performed an analysis to identify operons in the S. Typhimurium SL1344 genome that encoded at least one gene down-regulated in ΔssrB cells and that possessed an SsrB binding peak in the upstream regulatory region as defined by our ChIP-on-chip analysis. From this, the 133 down-regulated genes mapped to 86 operons, 49 of which had an SsrB interaction upstream or within the first gene of the operon (Table 2). This analysis captured all five reported operons in the SPI-2 genomic island in addition to ten operons outside of this island that encode SPI-2 translocated effectors.
In order to rigorously evaluate our genome-wide functional genomics data, we compared it against traditional biochemical experiments describing SsrB-DNA interactions at the SPI-2 locus. Previous data reported SsrB footprints upstream of 6 genes in SPI-2: ssrA, ssrB, ssaB, sseA, ssaG, and ssaM [25],[26]. Our ChIP-on-chip data showed discrete SsrB binding at all of these promoters except for the promoter reported to be between ssrA and ssrB [25] (Figure 3). We attempted to verify functional activity at this site, but could not using transcriptional fusions (data not shown). Our data also identified three additional SsrB binding peaks upstream of sseA, within the CDS of ssaJ, and in the IGR upstream of ssaR (Figure 3). Functional interactions were confirmed for sseA and ssaR in subsequent reporter experiments described below.
Previous attempts by others to identify a conserved SsrB DNA recognition motif have been unsuccessful. To overcome this, we employed a bacterial one-hybrid screen originally developed to define binding site preferences for eukaryotic transcription factors [27]. We fused the DNA binding domain of SsrB (SsrBc) to the α-subunit of RNA polymerase and screened a prey library of ∼108 DNA molecules previously counter-selected against self-activation (Figure 4A). We used the PhoP response regulator from E. coli as a control because a DNA recognition sequence for it was known [28]. Bait-prey combinations surviving selection on medium lacking histidine were purified, and preys were sequenced and analyzed using the motif-finding program MEME [29]. From 189 unique sequences isolated for SsrBc, over 80% contained a degenerate consensus motif, mCCyTA (Figure 4B). In control screens with the PhoP-αNTD fusion, the PhoP box sequence (G/T)GTTTA was identified in 11% of sequenced preys (12/109, data not shown) but this sequence was never captured by SsrB-αNTD and vice versa, demonstrating specificity of the bacterial one-hybrid system for prokaryotic regulatory proteins.
Next, we examined our ChIP-on-chip data for the presence of a conserved regulatory motif. We extracted sequence data from the local maxima of the 256 binding peaks and analyzed the sequences with the computational program MDscan [30]. Using the highest-ranking probes to generate an initial prediction followed by lower-ranking probes for refinement, this analysis identified motifs that represented either the forward or the reverse complement of the consensus sequence ACmTTA, which shares consensus with the motif identified in the bacterial one-hybrid screen (Figure 4B). We identified variations of this motif within footprinted regions of SsrB-regulated promoters [25],, however sequence degeneracy made it difficult to precisely map the input functions.
The analysis of regulatory evolution is particularly challenging because it is difficult to distinguish neutral stochastic change from functional divergence. To solve this problem in the context of mapping the SsrB binding element, we used comparative genomics to search for conserved promoter architecture in another organism with a similar genomic island to Salmonella SPI-2. The tsetse fly endosymbiont Sodalis glossinidius contains the Sodalis Symbiosis Region-3 (SSR-3) that is similar in content and synteny with the S. enterica SPI-2 locus [31]. Gene conservation includes the entire T3SS structural module extending to the regulatory genes ssrA-ssrB and all other genes except the effectors sseF and sseG, which are not present in Sodalis SSR-3. We aligned the sequences of the five mapped promoters in SPI-2 with the orthologous SSR-3 regions to identify local conservation. Highly conserved sites within the promoters were restricted to regions previously footprinted by SsrB [25],[26], whereas adjacent sequence showed substantial drift (Figure 4C). Within the conserved sites we identified a heptameric sequence in 7-4-7 tail-to-tail architecture that created an 18-bp degenerate palindrome. This palindrome was found in all SPI-2 and SSR-3 T3SS promoters with the exception of the sseA promoter that had only one reasonably well-conserved heptamer in the footprinted region (Figure 4C and Figure S1). Interestingly, two copies of the palindrome occur upstream of the ssrA-ssrB operon in S. Typhimurium within the same footprint, and the conservation of either site in Sodalis was weak. Evaluation of the heptamer motif in the palindrome showed high similarity to the motifs identified by the bacterial one-hybrid screen and the ChIP-on-chip experiments (Figure 4B and 4D), giving us confidence that we had identified the major recognition module for transcriptional input by SsrB. In accord with a previous observation [26], there was not a strict requirement in the spacing between the SsrB binding site and the downstream transcriptional start site.
The presence of a conserved palindrome sequence in SPI-2 promoters and in related sequences from the endosymbiont S. glossinidius suggested that regulatory input by SsrB was through a palindrome sequence architecture. However, other lines of evidence suggested that the recognition site architecture was flexible in nature: (i) our bacterial one-hybrid screen isolated functional single hexamer sequences, (ii) the SsrB footprint at the naturally evolved sseA promoter within SPI-2 [26] contained only one reasonably well-conserved heptamer, and (iii) degenerate or non-ideal palindromes exist in the genome. In order to deconstruct this architecture, we designed a set of experiments to test the palindrome's tolerance to mutation. We chose the ssaG promoter for these experiments because 16 of 18 bases were identical between SPI-2 and SSR-3 from S. glossinidius, differing only in the 4-bp spacer between heptamers (Figure 4C). We mutated the palindrome in a series of transcriptional reporters that were otherwise identical to the evolved ssaG promoter (Figure 5A) and promoter activity was compared to that of a wild-type palindrome sequence. Variants in which the first half-site (7′) or second half-site (7″) was deleted produced similar transcriptional activity to the wild-type palindrome, verifying that a single well-conserved heptamer is sufficient for transcriptional input under these experimental conditions (Figure 5B). Deletion of the 4-bp spacer sequence between the heptamers - the most degenerate element of the palindrome - also generated wild-type promoter activity. However, the orientation of individual heptamers was essential for transcriptional input since rewiring the palindrome in any head-to-head orientation produced negligible promoter activity. However, if the two half sites were swapped front-to-back so that they maintained tail-to-tail orientation (construct labelled “Reverse” in Figure 5), wild-type promoter activity was restored. Precise deletion of the entire 18-bp palindrome lead to ∼10% residual activity in wild-type cells, which was reduced to less than 1% in an ssrB mutant (Figure 5C). To determine whether the remaining 10% transcriptional activity was a result of an SsrB-dependent feed-forward mechanism or transcriptional read-through of our chromosomally integrated reporter, we constructed an ectopic deletion reporter. Assessment of reporter activity for this construct in addition to wild-type constructs in wild-type and ssrB mutant backgrounds showed a similar level of activity to the ssrB mutant (Figure 5D).
The results for the half-site deletion constructs, which retained activity similar to wild type, were unexpected. Therefore, we compared the sequences generated upon mutation against a consensus palindrome matrix generated from all SPI-2 and other identified putative elements. The 7′-4-X, X-4-7″ and 7′-X-7″ mutations introduced a number of base transitions and transversions never occurring in the matrix, however the modified 7 base pair heptamer retained 4–5 naturally-occurring bases along with the unchanged wild-type sequence in the other heptamer (Figure S2). The possibility existed that this modified heptamer, although now weaker in consensus, could still be sufficient for recruitment of a functional form of SsrB when paired with the other wild-type heptamer. To test this, we created an additional ectopic transcriptional fusion construct in which the left half (7′) of the palindrome was mutated to bases never occurring in the consensus matrix. When tested in promoter activity experiments, this reporter was unable to activate transcription and was identical to the X-X-X mutant construct (Figure 5D).
Salmonella SsrB and the Sodalis ortholog (SG1279) are 69% identical and 81% similar at the amino acid level. All of the critical residues in the dimerization helix and HTH motif required for specific transcriptional activity by SsrB [32] are conserved in the Sodalis ortholog (Figure S3). To demonstrate a functional role for the palindrome identified in Sodalis, we engineered luciferase transcriptional reporters that either contained (7′-4-7″) or lacked (X-X-X) the identified palindrome from the Sodalis SG1292 promoter (ssaG ortholog) and transformed them into wild-type S. Typhimurium and an ssrB mutant. The transcriptional activity from a wild-type Sodalis palindrome sequence was high, but was completely abolished in an ssrB mutant and in experiments where only the palindrome sequence was precisely deleted (Figure 5D). These experiments demonstrated a functional role for the conserved palindrome in Sodalis and the requirement for SsrB for transcriptional activation.
The above results identified the conserved, yet flexible, palindrome sequence defining DNA recognition by SsrB. To examine the extent to which regulatory evolution has been selective for this genetic architecture, we created a position weight matrix (PWM) for the five strongest palindrome sites in SPI-2 and the orthologous sites in Sodalis SSR-3. We then searched for representative candidates of this motif in the S. Typhimurium genome using the simple scoring algorithm MotifLocator [45],[46]. This analysis recovered the motifs upstream of ssaB, ssaG, ssaM, and ssaR that were used to generate the PWM. The palindrome in the ssrA promoter was not used to create the PWM due to its weaker consensus in the left heptamer, however, it was recovered in the computational analysis in a second group of lower-scoring motifs (Figure 6A). We identified 24 palindromes co-occurring with ChIP-on-chip peaks upstream of 24 different SsrB-regulated genes or operons. Applying a stringent threshold to the output allowed us to identify two groups - genes with high-scoring upstream palindromes (ssaB, ssaG, ssaM, ssaR, sopD2, sifA, sifB, sseK2, sseK3, sseL, sseA′, steC, and srcA) and those with medium-scoring palindromes (0.7–0.8 threshold; ssrA, STM1633, sseI, slrP sspH2, pipB, sseJ, pipB2, srfN, sseA and steB) (Figure 6A and Dataset S4) (sseA′ denotes the SsrB palindrome sequence upstream of sseA that falls within the ssaE CDS, while sseA refers to the SsrB-footprinted IGR site with only one conserved heptamer defined in Figure 4C). Remarkably, this accounted for 17 of 22 SL1344 genome-encoded effectors translocated by the SPI-2-encoded T3SS (exceptions are chromosomal steA, gogB, and sseK, and plasmid-encoded spvB and spvC). These genes either lacked an upstream ChIP peak above our 3-standard deviation cut-off (sseK) or had such a peak but did not reach statistical significance in our transcriptional profiling experiments (steA, gogB).
Our ChIP-on-chip data revealed three additional strong SsrB binding peaks within SPI-2: one in the IGR directly upstream of ssaR, a second within the CDS for ssaJ, and a third within the CDS for ssaE that would be predicted to influence transcription of the downstream effector/chaperone operon beginning with sseA. The analysis described above recovered SsrB palindrome sequences at the sseA' and ssaR locations prompting further validation of these sites. No palindrome was identified for the ssaJ interaction peak and so further characterization was not pursued. For the IGR palindrome upstream of ssaR, we tested both a chromosome-integrated transcriptional fusion and an autonomous episomal reporter. In wild-type cells these reporters were as active as other SPI-2 promoters, whereas promoter activity was abrogated in ΔssrB cells, implicating this IGR as a functional promoter for ssaR (Figure 6B and Figure S4). We next tested the function of the intragenic palindrome within ssaE (sseA'). For this, we constructed a single-copy transcriptional reporter that either contained (WT PsseA) or lacked (PsseA del) the single heptamer site in the sseA IGR and integrated this reporter into wild-type cells and mutants lacking either ssrB or the ssaE coding sequence that removed the high-scoring intragenic palindrome sseA'. These experiments showed that the sseA' sequence contributes approximately 75% of transcriptional output at the sseA promoter (Figure 6C), since deleting the single heptamer in the sseA IGR had little effect on transcriptional output in any of the strain backgrounds. These reporter data are in line with the respective binding scores for the ChIP-on-chip interaction peaks (Figure 3) and the sequence similarity for these elements with respect to the consensus palindrome (Figure 6A and Figure 4D). Together, these data provide compelling evidence for the identity of the DNA recognition element that has been selected through evolution to co-regulate an SsrB-dependent gene program involved in adaptation to a host.
Horizontal gene transfer is a well-recognized mechanism of bacterial evolution that gives rise to new phenotypes due to the coordinated expression of novel genetic components [1]. A good example of this is acquisition of type III secretion by mutualists and pathogenic bacteria enabling new colonization strategies within a host [33],[34]. Evolved changes to regulatory circuitry can also give rise to phenotypic diversity at the species level [19]. In both cases, regulatory evolution is required to correctly deploy gene products during infection, yet the extent to which regulatory evolution contributes to pathogenic adaptation is only beginning to be realized [8]. The SsrA-SsrB two-component regulatory system in S. enterica has been the focus of our efforts to understand the significance of regulatory evolution for pathogenic adaptation. This regulatory system was co-acquired with a T3SS encoded in the SPI-2 pathogenicity island and likely contributed to immediate and gradual phenotypic diversity as new regulatory nodes were explored and acted upon by natural selection.
Extensive work has been reported on the characterization of SsrB dependent genes, including functional evaluation of genes encoded within SPI-2 in addition to genome-wide transcriptional studies [14],[15]. In this study we identified genes co-expressed under SsrB-inducing conditions and found those with strong levels of expression localized predominantly to mobile genetic elements, recently acquired genomic islands or other annotated islands. We also identified many weakly co-expressed genes, some of which may represent ancestral Salmonella genes recruited into the SsrB regulon like the previously reported srfN [18]. Some of these genes may not be directly regulated by SsrB and will require further experimental investigation.
Direct profiling of SsrB-DNA interactions using ChIP-on-chip was used to identify SsrB binding sites in the genome. This analysis identified many interactions which have not been previously described and interaction sites within coding regions of genes which may represent non-canonical functions for SsrB. Other groups have reported the existence of similar numbers of ChIP-on-chip interactions within intragenic regions for other transcription factors [35],[36] suggesting that this phenomenon is not restricted to SsrB. In light of the disparate number of microarray genes in comparison to ChIP-on-chip peaks we attempted to generate a more comprehensive picture of the SsrB regulon by combining these data sets at the operon level. In doing so we believe that the nineteen operons containing differentially expressed genes determined by microarray and containing a ChIP-on-chip peak three standard deviations above the mean captured by this analysis represent the genes directly activated by SsrB (Table 2). Those operons having a ChIP-on-chip peak directly upstream in the IGR region encompass the majority of known SsrB regulated genes while those possessing a ChIP peak within the CDS of the first gene may represent non-functional interactions that deserve follow-up experimental investigation.
The ChIP-on-chip data not only provided information on the identities of SsrB-regulated genes but also gave insight as to the identity of the SsrB recognition element specified by the interaction site sequences. The regulatory architecture governing SsrB input has been elusive despite several SsrB footprints being defined biochemically [19],[25],[26]. Our ChIP-on-chip data further suggested that SsrB binding within SPI-2 was specific, with binding peaks overlapping precisely with regions of the DNA footprinted by SsrB [25],[26]. By using a genetic screening strategy together with functional and comparative genomics, we were able to define the essential SsrB regulatory element as being an 18-bp palindrome with a conserved 7-4-7 internal organization.
In support of the palindrome as the functional entity we showed the loss of SsrB dependence as a result of deletion of this element for the ssaG promoter. Evaluation of the 7-4-7 palindrome in the ssaG promoter revealed the minimal architecture and sequence orientation required for transcriptional input. Deletion of the entire palindrome resulted in less than 1% activity in wild-type cells, an equivalent level of activity to those lacking ssrB entirely. A search of the S. enterica genome for this palindromic motif revealed candidates upstream of the previously noted SsrB dependent genes, including two additional SPI-2 sites; one IGR site upstream of ssaR that until now had been cryptic, and one intragenic palindrome upstream of sseA in the ssaE CDS. In both cases these input sites were found to be functional.
Although palindrome architecture was conserved upstream of SsrB-regulated genes, degenerate palindromes in which one half-site was more conserved were also functional. As a result of our mutational analyses we conclude that so long as the orientation of a single heptamer of the palindrome is conserved with respect to the downstream gene, SsrB is tolerant of degeneracy in the adjacent spacer and heptamer sequences. While we were able to identify a number of limited palindrome-like sequences from our bacterial one-hybrid screen, this tolerance in addition to the library size required to pull out an 18-bp palindrome in large numbers may explain why we isolated functional single heptamer sequences and why degenerate palindromes naturally exist in the genome. A recent report by Carroll et al, postulated that SsrB first interacts with DNA as a monomer, followed by dimerization [32]. Our findings also suggest that dimerization is likely required for transcriptional activation however strong recognition by one monomer may stabilize interaction of a second monomer with a less than ideal sequence. The finding that a flexible palindromic sequence can be selective for SsrB input raises many interesting questions around the nature of regulatory evolution. The ability to use a short functional half-site adjacent to an uncharacterized threshold level of tolerated bases would reduce the period of neutral evolution required to generate an inverted repeat sequence twice the length [37], and would limit the loss of intermediate variations to drift while a more desirable palindrome is created by regulatory evolution. For bacteria that make use of horizontal gene transfer, this could increase the tempo with which new DNA is integrated into the regulatory circuitry of the cell.
We showed that the SsrB regulatory palindrome is also present in the orthologous SSR-3 island of the endosymbiont Sodalis glossinidius and that the palindrome evolved in Sodalis can act as a cis regulatory input function in Salmonella. Thus, in addition to supporting a pathogenic lifestyle within a host in Salmonella, it seems probable that this common promoter architecture may direct the activation of the SSR-3 T3SS of S. glossinidius in its endosymbiotic relationship with the tsetse fly host, although we acknowledge this requires experimental validation. The SSR-3 region in S. glossinidius is fully conserved in gene synteny and content with that of SPI-2 [31], with the exception of two effector genes missing in SSR-3 (sseF and sseG) that are required to localize vacuolar Salmonella to the perinuclear Golgi in host cells [38],[39]. The SsrB ortholog in S. glossinidius is ∼30% divergent with SsrB at the protein level, initially leading us to think that they might have different binding site preferences. To the contrary, high local conservation in the promoters evolved in Salmonella and Sodalis was the crux in defining the functional SsrB input among stochastic noise. This analysis revealed strong palindrome sequence conservation in five promoters identified in SPI-2 and in the orthologous sequences in Sodalis SSR-3.
Among palindrome-containing promoters, the ssrA promoter is exceptional for two reasons: a lack of conservation between Salmonella and Sodalis, and the evolution of tandem palindromes in Salmonella. One possible interpretation of this divergent regulatory architecture in front of ssrA might relate to bacterial lifestyle. Salmonella may have retained or evolved SsrB input here to create a positive feedback loop on the regulatory system to rapidly adapt to the host environment during infection, similar to transcriptional surge described for the PhoP response regulator [40]. The endosymbiotic relationship of Sodalis with the tsetse fly - where long-term vertical transmission has ostensibly been formative in shaping regulatory circuitry at certain promoters - may obviate the need for rapid transcriptional surge, leading to regulatory drift or selection against positive feedback. With the structure of SsrB available [32] and its recognized sequence now identified, future studies will be able to build a picture of how SsrB interacts with both its target DNA, RNA polymerase and potentially other transcription factors including nucleoid associated proteins in order to direct transcription of its regulon.
In summary, this work highlights the evolutionary significance of cis-regulatory mutation for the adaptation of Salmonella to a host animal. The DNA module that choreographs SsrB-mediated pathogenic behaviour in Salmonella appears to have been conserved for mutualism as well, thereby shedding new light on the significance of cis-regulatory mutations for bacteria evolving in different ecological settings.
All experiments with animals were conducted according to guidelines set by the Canadian Council on Animal Care. The Animal Review Ethics Board at McMaster University approved all protocols developed for this work.
The Salmonella strain used for microarray and ChIP-on-chip analysis was Salmonella enterica serovar Typhimurium strain SL1344. Bacterial strains and plasmids used in this work are described in Table S1. Primer sequences used to generate constructs are available upon request. Bacteria were grown in LB medium unless otherwise indicated. Low-phosphate, low magnesium (LPM) medium was used as bacterial growth medium for microarray, ChIP-on-chip, and transcriptional reporter experiments [20]. Liquid cultures were routinely grown at 37°C with shaking. Antibiotics were added to media as follows when necessary: ampicillin (Amp, 100 µg/mL), chloramphenicol (CM, 34 µg/mL) kanamycin (Kan, 30 or 50 µg/mL), and streptomycin (SM, 50 µg/mL). NM medium was used in the bacterial one-hybrid experiments as described previously [27].
Microarray experiments were conducted and analyzed as described previously [16]. cDNA was synthesized from RNA harvested from wild-type cells and an ΔssrB mutant. cDNA from 2 replicate experiments was hybridized to InGen arrays and analyzed using ArrayPipe version 1.309 [21]. Probe signals underwent a foreground-background correction followed by a printTipLoess normalization by sub-grid. Duplicate spots were merged followed by averaging of the two replicates. Local intensity z scores were calculated for determination of significance.
For operon analysis, S. Typhimurium SL1344 operons were defined as groups of genes encoded on the same strand with a maximum intergenic distance of 30-bp. Operons selected for further investigation were those possessing at least one significantly down-regulated gene from the cDNA microarray analysis of an ssrB mutant. A cDNA microarray score was assigned based on the average fold-change value of all genes within the operon. For ChIP-on-chip analysis, a top ChIP interaction score was defined as that of the highest scoring probe within the first gene or the intergenic region upstream of the first gene of the operon. For the analysis of regions of difference (ROD) between S. enterica serovar Typhimurium and Salmonella bongori, a reciprocal-best BLAST analysis was performed to identify orthologous genes between S. Typhimurium and S. bongori. Orthologs were defined as reciprocal best BLAST pairs with E-values less than 0.005. Comparison of gene synteny between regions encoding orthologous genes was performed to identify regions of low conservation including gene deletions and insertions. The location and names of genes flanking the comprehensive list of genomic islands is provided in Dataset S2 and were compared to those predicted using IslandViewer [41].
The bacterial one-hybrid (B1H) experiments were conducted as outlined previously using a single-step selection procedure [27]. Full-length phoP from E. coli and the C-terminal domain of ssrB (ssrBc) from S. Typhimurium were cloned into pB1H1 to create a fusion to the αNTD of RNA polymerase. Each bait vector was transformed into E.coli ΔhisB ΔpyrF, purified, and then cells were transformed again with purified prey library that was previously counter-selected for self-activating preys using 5-fluoro-orotic acid. Transformants were recovered for 1 h in SOC medium, washed with NM medium supplemented with 0.1% histidine (NM + his) and allowed to grow for 2 h in this medium. Cells were washed four times with water, once with NM medium lacking histidine (NM –his), then resuspended in NM –his and plated on 150×15 mm dishes containing NM –his media supplemented with either 1 mM (for PhoP screen) or 5 mM (SsrBc screen) 3-aminotriazole. Selection was for ∼48 hours at 37°C. Individual clones were selected from plates containing <600 colonies, the prey plasmids were isolated and sent for sequencing (Macrogen USA). Sequences were parsed to extract the 18-bp prey sequence, then inputted into MEME (version 4.1.1) for motif generation [29]. MEME was run with default parameters and included searching for motifs of length 5–17 bp in either forward or reverse direction and with no limit on the number of occurrences within an input string. Motif logos were generated using Weblogo, version 2.8.2 [42].
Chromatin immunoprecipitation-on-chip (ChIP-on-chip) was conducted as described previously using an SL1344 strain containing an ssrB-3xFLAG allele on the chromosome [19]. The primer sequences used to generate the DNA for recombination were: 5′GAG TTA CTT AAC TGT GCC CGA AGA ATG AGG TTA ATA GAG TAT GAC TAC AAA GAC CAT GAC GG3′ and 5′ATC AAA ATA TGA CCA ATG CTT AAT ACC ATC GGA CGC CCC TGG CAT ATG AAT ATC CTC CTT AG3′. This strain was generated by an allelic replacement method described previously [43] and causes lethal infection of C57BL/6 mice similar to wild-type SL1344 (Figure S5). Immunoprecipitated DNA from three experiments under SsrB-inducing conditions (LPM growth medium) and one experiment under non-inducing conditions (exponential growth in LB medium) was hybridized to a single chip printed with four whole genome arrays designed on S. enterica serovar Typhimurium strain SL1344 (Oxford Gene Technology, Oxford UK). Signals for each probe within an experiment were normalized to the median channel signal for the respective array. Signal ratios were obtained for both inducing and non-inducing conditions by calculating the ratio of the control probe value and experimental probe value. A final interaction score was obtained by taking the log2 value of the ratio between the non-inducing and inducing conditions for each probe to remove SsrB interactions that occur under non-inducing conditions. Parsing and data analyses were performed using the Python scripting language. Genome-wide ChIP-on-chip data was plotted using Circos v.0.51 [44].
ChIP probes were ordered according to their position on the S. Typhimurium SL1344 genome and local maxima for ChIP interaction scores were defined as interaction peaks. Peaks with scores greater than three standard deviations from the mean probe signal were considered significant ChIP interaction peaks and were ranked in order of descending interaction score. The sequence of the top-scoring probe for each peak was exported to a text file and used for analysis by MDscan [30]. The background parameter was run with output generated by the included genomebg program from the S. Typhimurium SL1344 genome sequence. The initial motif was generated from sequences from the top ten SsrB interaction peaks and refined using the top 25 peak sequences.
To identify instances of the palindrome motif in S. Typhimurium, ten 18-bp 7-4-7 palindromic motifs in the promoters of the SPI-2 genes ssrA, ssaB, ssaG, ssaM, ssaR and their orthologous SSR-3 genes were used as input for MDScan to identify a consensus motif and to determine the position specific scoring matrix (PWM). This PWM was used with MotifLocator to identify instances of this motif in the S. Typhimurium SL1344 genome. A background file specific to SL1344 was generated using the associated script called CreateBackgroundModel. A stringent threshold value of 0.8 was used [45],[46].
Transcriptional fusions to lacZ for the ssaG and sseA promoter palindrome analysis were generated using chemically synthesized double-stranded DNA (Genscript Corp). Promoter DNA was ligated into pIVET5n, then the plasmid was subsequently conjugated into SL1344 to generate single-copy transcriptional fusions integrated on the chromosome as described previously [20]. Luciferase reporter constructs for the Sodalis glossinidius SG1292 and ssaR promoters were generated by PCR amplification of promoter regions from genomic DNA templates. The luciferase reporter construct for the PssaG scrambled substitution was created by PCR product splicing via overlap extension using the existing ssaG promoter cloning primers and two additional internal primers containing the desired mutation sequence (DTM0061R, 5′CGC GAA AGC AAC GAT TAC TCC GGC GCA CG3′ and DTM0061.1F, 5′GAG TAA TCG TTG CTT TCG CGA TAC CGG ATG TTC ATT GCT TTC TA3′). This DNA was ligated into pCS26 and transformed into SL1344 to generate plasmid-based reporters. Overnight cultures of Salmonella were sub-cultured into LPM medium and grown with shaking for 7 h. Samples were removed hourly to measure β-galactosidase activity via a chemiluminescence-based assay as described previously [20] or luminescence directly from cultures (EnVision, Perkin-Elmer). Output was relative light units (RLU) normalized to OD600. Each experiment was performed in triplicate then averaged. Reporter activity from mutant and rewired promoters was normalized to that from wild-type promoters.
All ChIP-on-chip data can be retrieved from the NCBI Gene Expression Omnibus at http://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE20192. Data files for viewing in Artemis (http://www.sanger.ac.uk/Software/Artemis/) are available upon request.
|
10.1371/journal.pcbi.1003102 | Models of Self-Peptide Sampling by Developing T Cells Identify Candidate Mechanisms of Thymic Selection | Conventional and regulatory T cells develop in the thymus where they are exposed to samples of self-peptide MHC (pMHC) ligands. This probabilistic process selects for cells within a range of responsiveness that allows the detection of foreign antigen without excessive responses to self. Regulatory T cells are thought to lie at the higher end of the spectrum of acceptable self-reactivity and play a crucial role in the control of autoimmunity and tolerance to innocuous antigens. While many studies have elucidated key elements influencing lineage commitment, we still lack a full understanding of how thymocytes integrate signals obtained by sampling self-peptides to make fate decisions. To address this problem, we apply stochastic models of signal integration by T cells to data from a study quantifying the development of the two lineages using controllable levels of agonist peptide in the thymus. We find two models are able to explain the observations; one in which T cells continually re-assess fate decisions on the basis of multiple summed proximal signals from TCR-pMHC interactions; and another in which TCR sensitivity is modulated over time, such that contact with the same pMHC ligand may lead to divergent outcomes at different stages of development. Neither model requires that T and T are differentially susceptible to deletion or that the two lineages need qualitatively different signals for development, as have been proposed. We find additional support for the variable-sensitivity model, which is able to explain apparently paradoxical observations regarding the effect of partial and strong agonists on T and T development.
| T cells develop in the thymus, where they are vetted – they must respond weakly to self-antigens, but not so strongly as to risk causing autoimmunity. This selection process involves developing T cells being exposed to a large sample of self-peptides presented on specialised cells in the thymus, and deciding to die or to differentiate into mature T cells of either conventional or regulatory lineages. The rules by which T cells assimilate information from these interactions to make these decisions are not known. In this study we use previously published data to assess and discriminate between different models of thymic selection and find the most support for a model in which T cells vary their sensitivity to self-peptides during their development. This allows fate decisions to be made on the basis of as few as one peptide at a time, which allows for fine specificity in the selection process.
| Conventional T cells (T) and T regulatory cells (T) are essential components of the adaptive immune system. Conventional T cells develop effector function in response to foreign antigens, while natural T regulatory cells produced in the thymus play a key role in the maintenance of tolerance to self-antigens and prevent autoimmune diseases (reviewed in, for example, [1]). Both populations are derived from precursors in the thymus that develop, undergo selection and differentiate into different T cell lineages. The differentiation of a thymocyte into the mature T cell repertoire is dependent on the engagement of its T cell receptor (TCR) with endogenous peptides presented by major histocompatibility complex (MHC) molecules on thymic antigen presenting cells. Continued very weak or null interactions between the TCR and peptide-MHC ligands (pMHC) lead to failure to positively select (‘death by neglect’) while excessively strong TCR-pMHC interactions lead to negative selection, removing highly autoreactive cells from the T cell repertoire. However, the precise rules underlying T cell precursor fate are not well understood; based on its exposure to a sample of pMHC, how and when does a thymocyte decide to become a T, a T, or be deleted?
Studies using fetal thymic organ cultures have shown that there exists a sharp avidity threshold between positive and negatively selecting ligands [2], [3]. There is substantial evidence indicating that T are induced by TCR signals that lie below this negative selection threshold, but above that required for selection into the conventional T cell pool [4]–[8]. However, many uncertainties remain. It has been shown that expression of cognate antigen (which we loosely refer to ‘agonist peptide’) in the thymic epithelium is required for the generation of T [9]–[13], but a recent study showed that T commitment occurs over a wide range of TCR affinities for a ubiquitously expressed self antigen [14]. Further, the partitioning of fates with increasing strength of recognition for self (deletionTTdeletion) appears to be questioned by a study in which both expression of an agonist and a weaker partial-agonist could enhance deletion, but only the agonist was able to induce the formation of regulatory T cells [15], suggesting that either the mapping of avidity to fate is more complex or that qualitatively different signals are required for T and T selection.
Many experimental models using TCR transgenic cells (clonal populations of T cells with identical TCR) have shown that these cells can develop into both the regulatory and conventional lineages together in the same environment. This observation implies that there is stochasticity in fate determination. This stochasticity can be partitioned conceptually into two sources that are not mutually exclusive. First, there may be heterogeneity at the early double positive stage of development, even within a clonal population, that pre-disposes cells to different fates. This heterogeneity might derive, for example, from differences in expression of factors determining the baseline levels or dynamic range of TCR signalling, or other signalling proteins related to lineage commitment. Second, stochasticity may be present later in the selection process, arising at least in part because each thymocyte encounters an independent sample of self-peptide ligands. Evidence for the latter comes from observations that probabilities of deletion and T generation have been shown to vary with levels of agonist-peptide expression; in-vivo studies in TCR transgenic mice [16]–[18] and in-vitro fetal thymic organ culture [19]–[21] have shown that the efficiency of T selection increases with modest increases in agonist-peptide expression, but drops when expression is high. (We use the term efficiency here interchangeably with the probability of experiencing a given fate.) The efficiency of selection into the T lineage also decreases in the presence of increasing numbers of cells of the same specificity [14], [15], [22]–[24]. Thus the availability of relevant ligands, either in absolute terms or through competition, can influence fate decisions.
The timing of an interaction with a ligand may also influence fate. There is evidence that the sensitivity of thymocytes to TCR stimulation is increased during maturation through the subcellular localisation of signalling molecules such as tyrosine kinase Lck [25]; the inhibition of extracellular signal-regulated kinase (ERK) activation and increased expression of inhibitory tyrosine phosphatase SHP-1 [26]; the upregulation of the negative regulator CD5 [27]; and the increased expression of ZAP-70, a downstream target of TCR signalling [28]. However, the expression of miR-181a, a microRNA that enhances sensitivity to TCR stimulation, is reduced during thymic development [29], [30], and TCR signalling in response to low-affinity pMHC ligands is strongest in immature thymocytes [31]. The net effect of changes in TCR signal activating and inhibiting factors is not clear, but it is possible that stimulation with the same ligand will lead to different levels of activation in the same thymocyte at different stages of development.
The challenge of synthesising these observations and describing quantitatively how the affinity, number and timing of pMHC contacts shape the developing T cell repertoire invites a mathematical modelling approach. Models of thymic selection have been successful in providing insight into the relationship between diversity of self peptides sampled in the thymus and the cross-reactivity [32], [33], alloreactivity [34], size [35], [36] and CD4SP/CD8SP ratio [37] of the selected repertoire. Models have also helped us understand the relation of HLA phenotype to viral epitope recognition [38] and the trade-off between MHC and T cell receptor diversity [39]. In this study we use stochastic (probabilistic) models to describe previously published in vivo data describing T and T commitment of a transgenic cell population in the presence of varying densities of agonist peptide in the thymus [16]. These data allow us to test and discriminate between models of how developing thymocytes might integrate signals received from pMHC ligands to make lineage decisions.
Models of thymic selection must relate the physical interaction between a TCR and a pMHC ligand to the signal interpreted or integrated by the thymocyte. There is evidence to support competing models of TCR-pMHC interactions in which the level of T cell activation is determined by TCR-pMHC dwell times, through the kinetic proofreading model [40], [41], TCR occupancy [42]–[44] and overall pMHC ligand affinity [45]. Previous approaches to quantitative modelling of TCR-pMHC interactions can be divided into three broad categories: (i) detailed modelling of signal transduction immediately downstream of TCR-pMHC engagement [46], [47]; (ii) kinetic models of binding events using measured rates of TCR-pMHC association and disassociation [48], [49]; and (iii) the use of a ‘string model’ framework in which the strength of an interaction is determined by pairwise interaction energies between peptides and the aligned residues of amino acids on the variable CDR3 loop of randomly generated TCRs [32], [33]. However, binding kinetic parameters are not available for the full range of endogenous peptides that are encountered during thymic development, and uncertainty remains in the relation between avidity and the signalling thresholds determining fate decisions. Here, we abstract from the mechanistic model of signal strength derived from molecular interactions. Instead we assume a distribution of signal strengths that a given TCR derives from pMHC ligands, in which low-strength signalling events occur with the greatest likelihood and, in line with our knowledge of the specificity of T cell recognition, stronger signalling events occur with decreasing probability. We show that our conclusions are insensitive to the precise form of this distribution.
We explore candidate mechanisms of T and T selection using the canonical hypothesis that signals associated with T commitment are stronger than those required for T commitment but are below the threshold for negative selection. We use the data of van Santen et al. [16] to reject a simple model of the selection process in which thymocyte fate is based on testing sequential single TCR-pMHC interactions. Instead, we find the data can be explained with two generalisations of this model in which perceived TCR signal strength correlates to a strict hierarchy of cell fates (neglectTT negative selection). In both models, thymocytes are continuously initiating fate decisions based on measuring the strength of binding to self peptide-MHC ligands in a series of encounters with antigen-presenting cells. In one class of model, the model, cells measure the avidity of each encounter, each of which comprises binding to a sample of multiple self-peptide-MHC ligands simultaneously. The model is motivated by studies implicating the integration of signals from multiple pMHC interactions in the priming of mature T cells by antigen [50]–[53]. In the second class of model, the two-phase model, we examine the consequences of TCR sensitivity of thymocytes varying during development. In this model a cell's interpretation of the signal derived from a given ligand depends on whether it occurs early or late in selection. We show that both models are able to describe the data, and also make predictions that are consistent with studies quantifying the efficiency of T selection with avidity [14]. However, we argue that variable TCR sensitivity is required to explain the effect of partial and full agonist peptide expression in the thymus on T generation and negative selection reported by Cozzo Picca and colleagues [15].
We use data from van Santen et al. (2004) [16] in which the frequency of high affinity intra-thymic ligands was manipulated in vivo. Briefly, a mouse line was used that employs the tetracycline inducible system to conditionally express an invariant chain mutant, bearing the T cell epitope from moth cytochrome c (MCC) in place of the class-II associated invariant chain peptide (CLIP)-encoding region (TIM). TIM was expressed in both cortical and medullary thymic epithelial cells (cTEC and mTEC), and at controllable and graded levels. The mice also contained a transgene encoding a TCR specific for this peptide, such that in the absence of induced TIM, these cells differentiated efficiently into mature CD4 single positive thymocytes. Expression of TIM, measured by TIM RNA transcripts via real-time PCR, influenced both T and T formation in a non-linear fashion (Figure 1).
In Figure 1 we see that (i) low frequencies of a strong agonist (TIM) do not affect the selection of TCR-specific (AND) thymocytes into the conventional T cell pool; (ii) moderate increases in agonist expression lead to increasing efficiency of selection of AND cells into T ((T) against (Relative TIM RNA) between ; Pearson correlation ) and a concurrent drop in the efficiency of T selection; and (iii) high frequencies of a strong agonist lead to the deletion of AND T cells. A very similar trend was observed by Cozzo Picca et al. [17] using TCR transgenic cells specific for an epitope of influenza virus in the presence of different levels of expression of this agonist. Atibalentja et al. [18] also observed this trend following intravenous injection of varying concentrations of hen egg-white lysozyme (HEL), which was rapidly processed and presented in the thymus, resulting in the negative selection of specific TCR transgenic T and an increase in TCR transgenic T at low, but loss at higher, HEL concentrations.
Developing thymocytes survey pMHC ligands presented on the surface of thymic epithelial cells. In all models we assume that fate decisions are continually reassessed based on ‘encounters’, each of which is the sum of interactions with pMHC (Figure 2), where . Each thymocyte participates in encounters at most, where a thymocyte might undergo negative selection, or initiate development into the T or T lineages, before reaching its -th encounter. We assume that each encounter with one or more pMHC can be divided into four categories determined by its affinity or avidity and the resulting signal through the T cell receptor(s). These are (i) a weak or null signal below that required for positive selection; (ii) a signal sufficient for selection into the T lineage; (iii) a signal that initiates selection into the T lineage; and (iv) a strong signal that leads to deletion.
We considered two classes of models. In one, the distribution of signal strengths resulting from encounters is constant throughout the selection period - the model. In the other, the two-phase model, we allow for the possibility that this distribution shifts during selection as a result of temporal changes in TCR sensitivity.
The key parameters of interest were the encounter size for the model, and the number of encounters in each phase in the two-phase model. Other parameters were estimated simultaneously, but several quantities were taken as inputs to the models because the data from [16] did not allow us to parameterise them directly. These were (i) the parameters specifying the relation between relative RNA expression and absolute peptide abundance; (ii) the distribution of signal strengths obtained by the AND TCR from randomly sampled self pMHC ligands; and (iii) the relation between selection probabilities and absolute cell numbers. First, we explored ranges of parameters defining the mapping function (equation 6); . We chose to use this generic sigmoid dose-response curve given our ignorance of the mechanistic relation between RNA expression and peptide-MHC abundance on thymic epithelial cells. However, we were able to partially validate this choice of function, and the region of parameter space that we explored, using data from the study by Obst et al. [58]. They characterised the relation between the degree of activation of adoptively transferred AND T cells and the relative TIM RNA expression on MHC class II-expressing cells, using a similar tetracycline-inducible expression system to that used in [16]. Their readout of immune activation was the fraction of AND cells that had divided 60 h following induction of TIM expression. Assuming this fraction is linearly related to peptide availability we used the data from Obst et al. to estimate the parameters of the mapping function (equation 6). We found that both the recruited fraction and an alternative measure of immune activation, the estimated per capita rate of recruitment into division, yielded mappings within the envelope of functions generated with our parameter ranges. These mappings also lay well within the 95% uncertainty envelope generated by the best-fitting parameters from our analysis of the data from [16]. For details, see Text S3, Figure S1 and Table S1. Second, we assumed the logarithm of the signal strength derived from a single AND-TCR endogenous-pMHC interaction is normally distributed with zero mean and unit variance. The scale of the distribution of signal strengths is arbitrary and its coefficient of variation does not influence our conclusions (see Results). Third, the models provide the probabilities of selection into the T and T lineages and the data are absolute numbers of these populations in the thymus. We relate the numbers to probabilities through a scaling constant derived from the proportion of AND TCR cells that fail negative selection in control mice (Text S1).
The key features of the data are (i) T numbers decline monotonically with agonist expression and (ii) modest increases in agonist expression lead to an increase in the absolute number of AND T, with numbers then decreasing at higher levels of TIM expression (Figure 1). Assuming there is a positive relationship between TIM peptide presentation () and relative TIM RNA expression, equation 4 shows that a model in which fate decisions are re-evaluated after single TCR-pMHC contacts () can describe the T data, which falls progressively with .
However, we can see using a graphical argument (Figure 6, upper panel) that the model with constant TCR sensitivity will only be able to capture the trend in T numbers if encounters comprise TCR signals integrated over multiple pMHC bindings (). If the strength of an interaction between a single AND-TCR and agonist TIM () lies within the T-selecting range , we would expect to see a monotonic increase in T numbers with increasing agonist peptide expression; as agonist becomes more abundant, progressively more probability mass is contained within this area, boosting the probability of T selection (Figure 5, upper panel; Figure 7A, dotted-blue curve). Here, the one-hit model predicts that the absolute increase in T numbers is greater than or equal to the absolute decline in T numbers. Conversely, if is above the threshold for negative selection, , then we would predict a continuous decrease in T as agonist peptide becomes more abundant and increases the probability of deletion (Figure 5, upper panel; Figure 7A, dashed red curve). Neither of these trends are what is observed and so we rule out these scenarios. Finally, we can exclude the possibility that lies within the T -selecting range; if , increasing TIM expression would then increase the probability of selection into T, which we do not observe. Thus we can reject the simple one-hit model for selection of AND thymocytes.
Extending the argument above, to explain the rise and fall of T numbers with agonist peptide expression () requires the probability mass within the T-selecting region to increase then decrease with . This becomes possible when thymocytes read multiple TCR-pMHC bindings simultaneously (). Qualitatively, this is because when , replacing an increasing fraction of endogenous peptides with TIM () right-shifts the distribution of encounter strengths and, in contrast to the case, increases the probability of an encounter within both the T and negative-selection regions (Figure 3B). The probability contained below the T selection threshold falls with , consistent with T numbers falling; the probability of an encounter occurring within the T zone first increases then decreases with , as required to explain the data; and the probability of negative selection continually increases (Figure 6, middle panel).
We explored this quantitatively and sought to identify the parameters of the model from the data. They cannot all be identified uniquely. As described in Methods we took the approach of exploring a range of plausible parameters governing the function mapping RNA expression to endogenous peptide replacement by TIM, and a range of values of the maximum number of encounters, .
Remarkably, all values of yielded equivalent descriptions of the data, and the encounter size was highly insensitive to other parameters; it lay between 2 and 5 for all models, with best fitting value , independent of . We also found that a range of mapping functions were able to describe the data equally well (Table S1). In particular, we predict that at maximum RNA expression, TIM replaces beween 0.1% and 12% of endogenous peptides. Representative fits to the data are shown in Figure 7. Panel A illustrates the failure of the one-hit model, with the best fit obtained by forcing . Panel B shows the fit achieved with the model with a free parameter.
The estimate of is also independent of the variance of the TCR-pMHC signal strength distribution . This also derives from the fact that the key quantities are just the probabilities of interactions lying between the different thresholds . However these thresholds become increasingly spaced with (that is, as the log-normal distribution becomes increasingly fat-tailed). The less heavy-tailed the distribution of signal strengths, the smaller is the window of affinity/avidity for triggering T development with respect to the mean signal strength. Small increases in affinity can shift TCR signals from positively to negatively selecting [2], [3], and so if signal strength relates linearly to affinity or avidity [14], our model predicts that the distribution of encounter strengths with self may not be strongly heavy-tailed.
Anything between ten and a few hundred pMHC have been shown to be required for T cell activation (see for example, [60]) and as few as 3–5 for pMHC recognition by cytotoxic T cell effector function [61], although with the extent of TCR binding influencing the degree of activation [42]. However, data interpreted using the kinetic proofreading model suggest that multiple interactions with very weak ligands may not lead to activation at the whole cell level (see, for example, [40], [41], [62]). Therefore we wanted to test whether the low estimates of are an artefact of the assumption that every TCR-pMHC interaction generates a signal and so an encounter comprising weak TCR-pMHC bindings might still lead to strong signalling.
To do this, we extended the model such that only a fraction () of self-peptides are capable of inducing a signal through the AND TCR, and the remaining fraction are classifed as null. This introduces a stochastic element to the number of TCR contributing to the signal from each encounter. We found that increasing the abundance of null ligands increases the estimated TCR engagements per encounter (Table S1). For example, we estimate the number of proximal TCR-pMHC engagements per encounter () to be between 20–190 if 99% of peptides fail to trigger the TCR, and between 350–1000 when 99.9% of peptides are null. Intuitively, the increase in derives from the dilution of the information content of each encounter by the presence of null peptides. For each encounter to be a unit of sufficient information with which fate decisions can be triggered, the sample size must increase in the presence of null interactions. As for the simpler model () the estimate of is also independent of the number of encounters, .
Therefore, this extended model predicts that in the AND TCR system the expected number of productive TCR-peptide MHC interaction per encounter remains remarkably small (of the order 1). This is perhaps unsurprising, as low values of will allow thymocytes to discriminate between ligands with small differences in affinity.
Next we explored the implications of a time-varying sensitivity of thymocytes to TCR stimulation during maturation. The two phase model, as described in Methods, extends the one-hit model to include time-varying TCR sensitivity. Its predictions are independent of the direction of variation, but to illustrate we assume an interaction with agonist leads to T commitment during phase A early in development, but causes deletion in phase B when the same peptide is capable of inducing a stronger downstream TCR signal (Figure 4). Selection into the T lineage is still possible in both phases; what changes between phase A and phase B is a right-shift in the distribution of signal strengths with respect to the selection thresholds. This shift in probabilities within the different fate-determining affinity ranges yields the required trends in T and T production with TIM expression (Figure 6, lower panel).
The details of parameter estimation for this model are in Methods and in Text S2. The unknowns are , the number of encounters in phase A (Since , this is the maximum number of pMHC sampled in phase A), , the number of encounters in phase B, and , the probability that a randomly sampled pMHC in the low-sensitivity phase A will lead to negative selection.
As for the model, a range mapping functions described the data equally well. A representative fit using the two-phase model is shown in Figure 7C. We found a clear inverse relationship between the value of and the total number of encounters, + (Figure 8A). The model predicts that between 1–2% of encounters occur in the lower sensitivity phase (Figure 8B).
We have used data in which there is a profound loss of conventional T cells in the presence of relatively low frequency of agonist peptide, while T numbers are maintained and even initially increase with moderate increases in agonist frequency (Figure 1). These observations suggested the hypothesis that regulatory T cells are intrinsically more resilient to deletion by agonist peptides than conventional T cells [16]. Further, the study by Cozzo Picca et al. [15] showed that a partial agonist can induce deletion of conventional T cells but only an agonist could boost regulatory T cell generation. This led to a hypothesis that agonist peptide may deliver a qualitatively different signal that induces regulatory T cells.
We argue that neither of these hypotheses need be invoked. We have shown that both models can explain the first set of observations within a single affinity/avidity framework with different thresholds, without the need to assume differential susceptibilities of T and T to deletion. Further, we can see immediately that the model will not explain the partial/full agonist observations in ref. [15]. Their observation that partial agonist increases the probability of deletion with no increase in T suggests that the presence of the partial agonist shifts the distribution of the sum of interactions far to the right of the wild-type distribution, such that the bulk of the distribution is contained above the negative selection threshold. It follows that strong agonist must push this distribution even further rightwards, and so the probability of signals lying within the T-inducing zone must fall. This is inconsistent with the observed increase in T with agonist strength.
In contrast, the simple two-phase model can explain the effect (Figure 9). Assume that the partial agonist is not strong enough to induce T commitment in phase A when the TCR is relatively insensitive, but in the more sensitive phase B delivers a signal that lies above the negative selection threshold. Then the net effect of introducing a weak agonist is to increase deletion and have little effect on T numbers, as is observed. In contrast, suppose the strong agonist triggers T commitment when the TCR in phase A, but is negatively selecting in phase B (Figure 9, right hand columns). Then (i) expression of the strong agonist will always lead to a fall in conventional T cell numbers, and (ii) moderate levels of strong agonist, while increasing the overall probability of negative selection, can drive a net increase in T production by boosting the probability of receiving a T-inducing signal in phase A.
Figure 9 illustrates this effect for the case , but the same argument holds for a generalisation of the two-phase model with (Figure S2).
How does a thymocyte decide to become a conventional or regulatory T cell, or die, and when are these decisions most likely to take place? To address these questions, we have used experimental data to test models of how thymocytes interpret TCR interactions with self-peptide MHC to make fate decisions. We showed that the data cannot be explained with a model in which (from a given TCR's perspective) individual self-peptide MHC ligands are classed as positively selecting, negatively selecting or promoting T development, and in which a single engagement with a strong agonist pMHC is sufficient to influence cell fate. Instead we found stronger and roughly equivalent support for two alternative models, in which (i) a thymocyte bases decisions on TCR signals summed from multiple engagements with pMHC ligands, and/or (ii) a two-phase model in which TCR sensitivity alters during development, and so an engagement with the same pMHC ligand may lead to divergent outcomes at different stages of development. A robust prediction of the model is that the number of non-null TCR-self-peptide interactions per encounter that contribute to fate decisions is remarkably small. The two-phase model predicts that the initiation of differentiation into the T lineage is most efficient during a relatively short temporal window during which the TCR is less sensitive to stimulation. Importantly, the models express the probabilistic aspect of thymic selection that likely underlies the heterogeneity in lineage decisions within a clonotypic population.
We focused here on a system in which agonist availability could be manipulated. We can also use the models to make predictions regarding selection efficiency as a function of TCR affinity. Lee et al. [14] quantified the efficiency of T selection in the rat insulin promoter (RIP)-mOVA system for a range of TCR clones with varying affinity for OVA. They found T were generated across a broad range of reactivities and found an increase in T selection efficiency with increasing affinity for a self peptide. Both the and variable-sensitivity models are able to explain these observations (Text S4 and Figure S3), but they make distinct predictions. The model predicts that the probability of the summed-proximal signals at each encounter exceeding the T threshold is lower for weaker clones; but this probability remains constant throughout development. A variable-sensitivity model predicts that T selection efficiency is determined by the duration of the window in which agonist contact triggers T commitment. In an increasing-sensitivity scenario, clones that recognise OVA more weakly will take longer to reach a level of TCR sensitivity that can induce T, and so a prediction of the model is that lower-affinity TCR clones will commit to the T lineage later in development.
These models might be distinguished by manipulating thymocytes' TCR sensitivity, for example through altered expression of signalling proteins, at different stages of thymocyte development. The model predicts that the efficiency of T selection would be altered equally for all clones, whereas a model of increasing TCR sensitivity predicts that damping of TCR signalling later in development would most strongly reduce T selection efficiency for clones with low affinity for self peptide.
We deliberately did not use model selection criteria to discriminate between models, in part because it is not possible to identify all parameters uniquely. Instead we identified regions of parameter space for each model that provided reasonable descriptions of the data. Importantly, the predicted values of the encounter size for different abundances of null peptides () are insensitive to the remaining parameters (Table S1). We consider the and two-phase models capable of describing the data equally well because they are able to capture the decline in T and increase then decrease in T with TIM abundance. Both models capture this behaviour provided TIM abundance increases monotonically with RNA expression, which we expect to be the case. Further, model selection criteria are not required to reject the simplest one-hit model, nor to compare the abilities of the and two-phase models to describe the partial agonist observations.
One prediction of an avidity-based model is that thymocytes may be deleted if they interact simultaneously with several pMHC at moderate affinities. We believe this prediction is not necessarily unreasonable; such events may be an inevitable byproduct of a selection process that is inherently probabilistic and which can result in cells with identical TCR experiencing divergent fates.
The two mechanisms we explore here are not mutually exclusive. We illustrated the two-phase model assuming that decisions are made based on single, rather than summed interactions with pMHC. However it can be generalised to an extended version of the model in which encounters are interpreted differently as TCR sensitivity increases. This model will be able to explain the observations a least as well as the or two phase models, at the cost of extra parameters. Also, we illustrated the impact of increasing TCR sensitivity with a simple model that divided development into two discrete phases, while increases in TCR sensitivity are likely to be continuous. Modelling smooth changes in TCR sensitivity will introduce additional parameters but we expect such a model to yield qualitatively similar results. Importantly, our analysis does not exclude the possibility that T are more resistant to deletion and/or that qualitatively different signals are involved in T and T commitment; we simply show that these mechanisms need not be invoked to explain the observations.
There is evidence that positive selection requires multiple low-affinity contacts with self-peptides in the thymus [63]–[65]. In contrast, in our model, a single encounter above the threshold is sufficient for positive selection. We assumed this threshold is very low for the AND thymocytes, which are strongly selecting under normal conditions. These cells will presumably receive essentially continuous positively selecting signals. In the general case, and particularly for weakly signalling TCRs that are near the threshold for death by neglect, the threshold would need to be included in the parameter search and the model would be extended to include a memory of recent interactions; one possibility is a model in which levels of survival or fate-determining proteins are increased by TCR signalling but decay in its absence.
There is substantial evidence that increasing the frequency of a given clonal (single TCR specificity) population reduces its efficiency of selection and in particular the probability of being directed into the T lineage [14], [22], [24]. Our models treat cells as independent entities and do not explicitly incorporate the possibility that competition between thymocytes of similar TCR specificities might influence the availability of selecting ligands. However one mechanism of competition can be represented quite straightforwardly in the models. If the strength of an encounter correlates with its duration, or perhaps increases the probability of internalisation of the pMHC ligand by the thymocyte, the TCR-specific cells will compete for and possibly sequester agonist and other high-avidity pMHC ligands. This will shift the apparent distribution of signal strengths leftwards towards lower avidity (and more available) interactions, reducing the probability of acquiring T-selecting signals. This model of competition for higher-avidity pMHC ligands may also explain the observation that the efficiency of T selection can increase with precursor frequency [22]. However it remains an open question whether competition for pMHC plays an important role in selection at physiological precursor frequencies.
Signalling through the IL-2 receptor is a requirement for T development [66]–[68]. It is thought that strong TCR signalling below the negative selection threshold may sensitise cells to IL-2, licensing progression towards the T lineage. Whatever the precise role for IL-2, it must operate downstream of fate-determining signals if selection is governed by a hierarchy of TCR avidity thresholds. Nevertheless, if IL-2 is limiting it may provide an upper bound on the total rate of production of T, either by redirection of cells to the T lineage or through loss. We argue however that competition for non-specific factors is unlikely to play a significant role in the system we are working with. First, the source of the IL-2 is unclear but we can reasonably suppose that IL-2 production in this system is independent of TIM expression. T numbers increase with TIM at low expression levels, indicating that IL-2 cannot be limiting in this region (Figure 1). Similarly it cannot be limiting at higher TIM levels as T decrease. It remains possible that a capping of T production through competition for IL-2 may be occuring in a small flat region near the peak in T numbers, but competition for non-specific factors alone cannot explain the key aspects of T development we are attempting to describe.
Early neonatal thymectomy experiments suggested that the development of T is delayed compared to conventional T cells [69]–[71]. A key T marker, the transcription factor Foxp3, is predominantly observed in the mature CD4 single positive stage of thymocyte development [72]. However, there may be a lag between initiation of T development and the stable expression of Foxp3; and it is possible that factors required for T development such as IL-2 [66]–[68] or medullary thymic epithelial cells [73] may only be required later in the maturation process. Thus the timing of T commitment remains unclear. The two models explored here make different predictions regarding this timing. The model suggests that diversion into the regulatory T cell lineage occurs with constant probability per encounter throughout selection. In contrast, the key prediction of the two-phase model is that the development of AND TCR T is triggered most efficiently within a relative short window during which thymocytes are relatively insensitive to TCR signalling. This window is estimated to span roughly 2% of the total pMHC ligands sampled, with the caveat that the two-phase model is an abstraction of what is more likely to be a temporally graded shift in sensitivity.
The two-phase model's predictions are identical whether the shorter, less sensitive phase occurs early or late in development. However, expression of the downstream TCR-signalling protein Zap70 increases progressively between the double positive (DP) and single positive stages of thymocyte development, and is associated with increasing sensitivity to TCR stimulation [28]; immature DP thymocytes display lower surface expression of TCRs, as compared to mature single positive ( or ) thymocytes, and TCR signalling may be actively inhibited in immature DPs [74]. Thymocytes' signalling environment may also change as they develop. Selection begins in the thymic cortex, where pMHC are encountered on cortical thymic epithelial cells, before cells migrate to the medulla where they encounter pMHC on medullary thymic epithelial cells and dendritic cells. It is thought that levels of co-stimulation and antigen presentation are generally lower in the cortex than in the medulla [75]–[77], suggesting that there may be an effective increase in TCR signalling during development. Using these observations, the two-phase model predicts that T development begins predominantly, but not exclusively, during a short window at the earliest double positive stage of selection. Clearly, definitive experimental identification of when T development is initiated will substantially increase our understanding of how thymocytes process information.
Reliable recognition and discrimination of self and nonself ligands requires both TCR sensitivity and specificity. Specificity will decrease as the number of pMHC integrated per encounter () increases — when is large, many pMHC engagements are integrated at each encounter, and so the thymocyte is then just sampling the mean of the distribution of pMHC affinities, and information is lost. This may explain why the optimum values of in the model are at the lower end of the reported numbers of pMHC engagements required for T cell activation; T cell activation may invole a relatively small number of informative TCR recognition events, together with many brief engagements with null or very low affinity peptide ligands. Our analysis of the model places small lower bounds on the number of non-null TCR engagements per encounter; but the two-phase model explains the data by allowing even a single non-null pMHC recognition event to influence fate. We speculate that varying TCR sensitivity with time in the thymus may allow for increased specificity in self-nonself discrimination.
|
10.1371/journal.pntd.0006929 | Organization of oversight for integrated control of neglected tropical diseases within Ministries of Health | Neglected tropical diseases (NTDs) are communicable diseases that impact approximately 1 billion people, but receive relatively little research, funding, and attention. Many NTDs have similar treatments, epidemiology, and geographic distribution, and as a result, the integration of control efforts can improve accountability, efficiency, and cost-effectiveness of programs. Here, we examine the landscape of efforts towards NTD integration across countries with the highest burden of disease, and review the administrative management of integration in order to identify approaches and pathways for integration.
We utilized a standardized system to score countries for NTD endemnicity to create a list of 25 countries with the highest overall burden of NTDs. We then conducted a literature review to characterize the NTD control programs in the focus countries. Six countries were selected for key informant interviews to validate literature review results and gather additional data on opportunities and obstacles to NTD integration, from an administrative perspective. The majority of countries included in the study were located in Africa, with the remainder from Asia, North America, and South America. Multiple models and pathways were observed for the integration of NTD programs, in combination with other NTD programs, other diseases, or other health programs. Substantial heterogeneity existed with respect to the NTD control programs, and no country had integrated all of their NTD control efforts into a single program. NTDs that can be treated with preventative chemotherapy were frequently integrated into a single program. Leprosy control was also frequently integrated with those of other communicable diseases, and notably tuberculosis. Barriers to NTD integration may result from internal administrative obstacles or external obstacles.
Although many countries have begun to integrate NTD control efforts, additional work will be required to realize the full benefits of integration in most of the countries examined here. Moving forward, NTD integration efforts must ensure that administrative structures are designed to maximize the potential success of integrated programs and account for existing administrative processes.
| Neglected tropical diseases (NTDs) are communicable diseases that impact billions of people but receive disproportionately little attention from researchers and funders. Many of these diseases have similarities in their epidemiology and control measures, rendering the integration of control programs a practical option to improve accountability, efficiency, and cost-effectiveness. Efforts to integrate NTD control programs have begun across many of the countries with the highest overall burden of NTDs, although no standardized approach for integration exists. Our research sought to examine the landscape of approaches for NTD integration, across the 25 countries with the highest burden of NTDs, to identify models that could be used for countries seeking to integrate their NTD programs. Integration often first targets diseases that can be treated with preventative chemotherapy, though multiple administrative pathways and models exist, including integrating NTD control programs with other NTDs, other communicable diseases, or other health initiatives. Still, no country has yet fully integrated all of their NTD control efforts into a single program. This may be due to internal and external factors that impede the integration of NTD control. Future NTD integration efforts must account for these factors to maximize the potential success of integrated programs.
| Neglected tropical diseases (NTDs) are a collection of infectious diseases caused by parasites, viruses, and bacteria. These diseases affect approximately one billion of the world’s poorest people, and most often impact populations living in sub-tropical climates with inadequate access to health care, clean water, sanitation, housing, education and information [1]. All low-income countries are affected by at least five NTDs simultaneously, and 149 countries are affected by at least one NTD [2]. Other estimates suggest that NTDs are some of the world’s most common conditions, accounting for greater than 2 billion infections globally [3]. Countries have worked to combat the impacts of NTDs by implementing “vertical” programs, aimed at preventing specific diseases using top-down approaches. Vertical programs, which are often supported by external funders and organizations, help countries measure success by implementing treatments and lowering prevalence levels. Because of the high burden of NTDs, having successful control programs in place is necessary for decreasing their prevalence and associated morbidity.
Many NTDs have similarities in treatment measures, epidemiology, and geographic distribution [4]. Accordingly, many NTDs have similar strategies for control and eradication. Among the 15 most common NTDs, seven are controlled using preventative chemotherapy in NTD endemic countries [5]. Traditional approaches to NTD control often relied on the aforementioned vertical programs within these countries working in parallel to one another, using the same treatments in the same areas and populations [6]. As a result, although vertical control programs are effective tools in combating specific diseases, integrated disease control programs could enhance control efforts by combining efforts to control multiple diseases into a single intervention.
WHO now recognizes the integration of NTD efforts as a crucial activity for tracking progress, ensuring accountability, and informing the development of policies and strategies [7]. It is in this context that NTD control programs may be incorporated into broader public health systems providing opportunities for countries to advance their NTD control by increasing efficiency, improving the overall quality of health services, covering a larger percentage of the population, and reducing the disparities associated with control programs [8]. Recent disease integration efforts have also yielded considerable savings both financially and in personnel time [9], and modeling efforts have identified opportunities for epidemiological benefits at a population level under some conditions [10]. Thus, the positive impacts of large-scale integrated disease control programs–both for the burden of NTDs, as well as the cost-effectiveness of interventions–may render them the best option for many countries [6]. However, there is no standardized approach to integration, allowing for substantial heterogeneity at the country-level in the implementation, administration, and oversight of integration efforts.
Generally, integrated disease control efforts are administratively placed within Ministries of Health (MOH), and thus the leadership, management, and organizational structures of the ministry can impact the ability to integrate programs. The goal of this work was to understand and present the various was by which NTD endemic countries have approached the integration of NTD control from an administrative standpoint. By observing the different approaches taken by NTD-endemic countries, we hoped to be able to extract common elements which might serve as recommendations or lessons learned that could be provided as a model to other countries that have yet to integrate their NTD control programs.
We sought to identify countries most impacted by NTDs, and then selected a sub-set of nations to examine the structure of their respective MOHs, specifically looking at the units and programs overseeing NTD control, and the extent to which they have been integrated. To accomplish this, we used a mixed methods approach that combined a literature review that assessed published evidence on the administrative integration of NTD control efforts, followed with purposely selected key informant interviews to validate the review results and provide additional information not captured in the literature.
To narrow the scope of the study from the almost 150 countries affected by at least one NTD, we chose to focus on 25 countries with multiple endemic NTDs. To do so, we assembled two lists of countries affected by NTDs. The first list examined all countries of the world for the presence of all priority NTDs as defined by the World Health Organization (WHO) (Table 1) [11], focusing on data from 2010 onwards. Using primarily the information provided on WHO’s NTD-specific websites (accessed in 2017 and 2018), which cover priority NTDs and where they are found globally, each country was “scored” for each NTD: 0 indicated a disease was not present within a country; 0.5 indicated that a disease was present within a country but not endemic; and 1 indicated that a country was endemic for a disease. For this work, “endemic” was defined as regular or established occurrence within the boundaries of the country, while “presence” was defined as any reported occurrence. The overall NTD burden was then calculated by totaling the numbers for each country, and the 25 countries with the highest burden of, as measured by our weighted scoring of number of disease presence and endemicity, were identified.
The second list repeated this scoring system, again across all countries in the world, but only using NTDs included within the London Declaration: Chagas disease, dracunculiasis, human African trypanosomiasis, leishmaniasis, leprosy, lymphatic filariasis (LF), onchocerciasis, schistosomiasis, soil-transmitted helminths (STH), and trachoma [12] (Table 1). This was done in efforts to align our analysis towards those countries with the greatest burden of London. Declaration NTDs, which are those prioritized for control and elimination, and are more likely to have existing control initiatives in endemic countries.
The two lists were then reviewed side by side to create a consensus list of the 25 countries with the highest overall burden of NTDs, which we then used to focus our literature review on the types of NTD control programs and approaches. By combining these lists, we sought to expand the geographic scope of the countries reviewed while maintaining a focus on priority NTDs.
We conducted a systematic literature review to characterize the nature of NTD control programs (vertical or integrated) in each of the 25 countries of interest. The review included a broad range of materials, including academic journals, published reports, “grey” literature and other publicly available guidance documents. MOH websites of the 25 countries were reviewed for relevant information on NTD integration efforts. Databases–including Google Scholar, JSTOR, and PubMed–were also searched for materials identifying NTD control programs in the countries of interest. Searches were performed by combining the name of a country and the term “NTD control program.” See S1 Appendix for the complete search strategy. Snowball sampling techniques [13] were used when reviewing these materials to identify other stakeholders involved in NTD integration. The websites of identified stakeholders were also reviewed for information about NTD integration efforts.
Eligibility for inclusion required items to focus on the integration of an NTD control program in a country of interest and to be published in the year 2000 or later. Language restrictions required documents to be written in English or French. No limitations were placed on publication type. This approach was justified to provide a thorough review of information relating the integration of NTD control efforts.
One author performed the initial search, screening of materials, review of full texts, and extraction of data; the search and subsequent screening, review, and extraction was re-performed by a second author to validate and corroborate results. All extracted data was reviewed by a third author for final validation; in cases of discrepancies with data characterization, the data sources were reviewed again and discussed among the remaining authors to reach a consensus. Data extracted characterized NTD control programs as either vertical or integrated. For the purposes of this paper, a program was considered vertical if it focused on a single, specific NTD; programs were considered to be integrated if they combined disease prevention efforts for two or more diseases or conditions (but not necessarily NTDs). In the event that is was unclear if integration had occurred, control programs were assumed to be vertical. If comprehensive control programs were not in place, we characterized programs based on if mapping or surveillance activities had occurred, as a likely precursor to establishment of a control program. Integrated control programs were further characterized based on what other diseases or health programs were integrated with the NTD(s).
Six countries (Brazil, Guinea, India, Kenya, Mali, and Mexico) were selected for further research. These countries were purposively selected to present a range in terms of country size and geography, diversity of endemic NTDs, and in some cases because of existing connections between the researchers and their respective MOH. Individuals from the MOH of these countries were contacted to identify one or more appropriate key informants, validate results from the literature review and gather additional data, particularly related to the opportunities and obstacles for integration of NTD control from an administrative perspective. For all six countries, semi-structured, open-ended interviews were conducted with one or more officials affiliated with NTD control programs within the country and served to highlight key challenges and opportunities with respect to integration efforts. In Mali and Guinea, in addition to MOH personnel, we interviewed officials from non-governmental organizations who collaborate with the MOH to implement national NTD control programs, and who were referred us by their MOH counterparts. A copy of the interview questions sheet used to conduct the semi-structured interviews is provided as Supporting Information.
The countries with the highest total burden of NTDs were Benin, Brazil, Burkina Faso, Cameroon, Central African Republic, Chad, Côte d’Ivoire, Democratic Republic of the Congo, Ethiopia, Ghana, Guinea, Guinea-Bissau, India, Indonesia, Kenya, Malawi, Mali, Mexico, Mozambique, Niger, Nigeria, South Sudan, Sudan, Tanzania, and Uganda. Geographically, 21 of these countries are located on the African continent, two are located in Asia, one in North America, and one in South America (Fig 1).
Results from the literature review suggest that the integration of NTD programs can be categorized into three groups–full integration of NTD programs, integration of select NTD programs with other NTD programs, and integration of select NTD programs with other public health programs or initiatives.
Substantial heterogeneity existed with respect to the NTD control programs in place. No country had integrated all of their NTD control efforts into a single program. Several countries (Central African Republic, Chad, Guinea, Guinea-Bissau, Indonesia, Mexico and South Sudan) had NTD control profiles that were characterized, per the information available to us, almost exclusively by vertical programs. Of the countries included in this study, Guinea-Bissau’s NTD control efforts were the least robust, as the country lacked control programs for three of the seven NTDs affecting the country.
Chagas disease was the NTD with the fewest programs in place, but it is also the NTD with the smallest geographic distribution. Both countries on our list that are endemic for Chagas–Brazil and Mexico–have control programs in place for the disease [14, 15]. Human African Trypanosomiasis (HAT) was the most geographically widespread disease that had the fewest control programs in place. The disease is present in 19 of the 25 countries considered in this study, but we were unable to find evidence of control programs in three countries–Guinea-Bissau, Niger, and Tanzania–and Ghana and Nigeria only have surveillance programs in place. Leishmaniasis is another disease that often lacks control programs. Leishmaniasis is considered present in Burkina Faso, Ghana, and Nigeria, but based on our literature review, do not currently have leishmaniasis control programs.
The administrative approaches used to integrate NTD programs ranged from multiple NTD combinations, to integrating NTD control with other communicable diseases, to integrating NTD control with Water, Sanitation and Hygiene (WASH) programs (Table 2). MOHs have most frequently integrated control efforts based on similarities in treatment, and the integration of schistosomiasis and STH programs was the most common integration effort. 11 countries–Benin, Brazil, Burkina Faso, Cameroon, Ghana, Kenya, Malawi, Mali, Nigeria, Tanzania, and Uganda–have further integrated schistosomiasis and STH efforts with those for lymphatic filariasis, onchocerciasis, and trachoma [16–22]. Control of these five diseases can be achieved using preventative chemotherapy. Leprosy control efforts were also frequently integrated with those of other communicable diseases caused by Mycobacterium species. Ethiopia [23], Tanzania [24], and Uganda [25] have integrated leprosy and tuberculosis (TB) control efforts, Benin has integrated leprosy and Buruli ulcer (BU) [26], Nigeria has integrated leprosy, BU, and TB control efforts [26], and Cameroon has integrated leprosy control efforts with those for BU, yaws, and leishmaniasis [27]. Ethiopia [23], India [28] and Sudan [29] have worked to integrate NTD control efforts, and particularly those with an arthropod vector, with other vector-borne diseases. The integration of NTD and WASH programs has occurred in Cote d’Ivoire (focused on dracunculiasis elimination) [30] and Ethiopia (trachoma) [23]. See S1 Data for a full summary of NTD control programs.
For some countries, we identified the existence of a plan, but were unable to verify the details or implementation of NTD control programs. For other countries, sources presented conflicting data. In Tanzania, the Health Sector Strategic Plan for 2015–2020 states that the country will work to improve the detection and management of HAT but does not specifically mention a control program [31]. However, according to WHO, HAT is a priority disease for control and elimination in Tanzania and a focal point exists who is responsible for the coordination and control of activities, which are integrated into an NTD master plan and a national One Health Strategy [32]. A 2010 report [33] suggested that Chad had a schistosomiasis program in place, but more recent documents suggest that no control programs currently exist [34, 35]. Similarly, a 2016 WHO document [36] states that mass drug administration (MDA) efforts for LF have not yet started in South Sudan, but the 2016 South Sudan NTD Plan [37] indicates that MDA is occurring in all parts of the country affected by LF.
Results from the literature review and key informant interviews also revealed important considerations for administration and management of integrated NTD control efforts. Multiple models (Fig 2) and pathways (Fig 3) exist for how integrated programs can be managed by MOHs. Models for integrated NTD programs may span from no administrative integration, with formal or informal coordination between vertically organized units, to partial administrative re-structuring for integration (usually treatment-oriented) (Fig 2A) to more comprehensively integrated units that strive to address all NTDs (Fig 2B), and create linkages with other communicable diseases and/or health services. Several unique pathways exist for transitioning to the integration NTD control programs (Fig 3). Thus far, integration efforts have included the creation of new, integrated MOH-endorsed programs, adjusting administrative structures to expand integrated control efforts, and intersectoral cross-over and collaborations within countries. The creation of new international initiatives and partnerships also represents a pathway towards the integration of NTD control programs.
Key informant interviews revealed that advantages of more fully integrated models included easier oversight of resources and timing of interventions, and stronger advocacy with Ministry leadership for continued integration. This advocacy is important as high-level decision makers within Ministries may not fully recognize the benefits of integration and thus not provide support for the type of administrative restructuring necessary for successful integration. Finally, there are notable administrative obstacles outside of Ministries’ control such as silo-ed funding streams and implementation partners that may, intentionally or not, make the integration of NTD efforts difficult. Interviews also revealed that even if a MOH is willing and able to integrate NTD control programs, earmarked aid from development partners or the organizational structure of implementing partners may prevent them from doing so.
This work highlights the distinct and varied approaches taken by different countries when integrating NTD control programs. Countries with at least one NTD control program frequently have multiple programs in place, often based on funding from non-governmental organizations and public-private partnerships for NTDs focused on vertical elimination and integrated control [6]. All countries included in this study had at least one vertical control program and many of these programs are supported by international development partners–such as the United States Agency for International Development (USAID), the United Kingdom’s Department for International Development (DFID), the World Bank, and the Bill & Melinda Gates Foundation–who can greatly influence the formation and implementation of program activities. Although implementation plans are largely collaborative and country-driven, disease or outcome-targeted funding may also affect opportunities for integration even if the country desires it, discussed in more detail below.
WHO has recently emphasized the need for the integration of vector management, treatment management, information systems, and sectoral collaboration [7] and the data gathered in this study demonstrate that countries are embracing the toward this recommendation of more integrated approaches. This represents a significant step forward for improving health outcomes and the cost-effectiveness of control strategies [8, 9]. Although integration allows for greater control over the allocation of resources, monitoring and evaluation of programs, and other critical activities, these programs also present additional challenges such as greater dependence on the political environment, such as requiring more will, leadership, and long-term resource investment.
Based on our findings, diseases requiring preventive chemotherapy are often first to be integrated administratively, while it is less common for NTDs that require individual case management. This is intuitive, as case management treatment options vary widely for NTDs and may not easily be combined; with lower prevalence levels overall, the incidence of co-infection tends to be lower for these diseases and thus there is reduced opportunity for synergizing patient care for multiple infections. However, given the intensive patient care required for treatment in these cases, there may be opportunities for integration of surveillance and disease management with other aspects of healthcare, such as maternal health visits or health promotion activities. In some instances, NTD integration efforts have focused on commonalities in transmission (e.g. integrated vector management) though fewer formal integrated programs exist in this regard, despite the beneficial opportunities such an approach might provide. One example of this is the Global Vector Control Response 2017–2030, endorsed at the 2017 World Health Assembly [38], which has focused on strengthening the prevention of NTDs through inter- and intra-sectoral action and collaboration, and expanding and integrating of vector-control tools and approaches [7]. This strategy also promotes the integration of NTD treatment through MDA campaigns or case management [7]. These actions ultimately should act to increase cost efficiency and help to expand the coverage and sustainability of NTD control efforts.
In accordance with one of their key recommendations of improved integrated surveillance and information systems, WHO has led the development of an integrated NTD database to improve planning and the management of NTD programs allowing for a central, single source of data concerning NTD programs, incorporating input and support from a large number of partners [39]. This platform provides key data on NTDs with the intent of leading to earlier detection of outbreaks [7]. Our literature review and key informant interviews did not uncover evidence that this database is being used by in-country stakeholders for control program planning, so concerted efforts to raise awareness about the availability of this resource may be beneficial. It is also worth noting that the utility of database could be improved by incorporating information regarding control efforts into a single, publicly available platform. This would help to clarify uncertainties surrounding NTD control efforts (e.g., Indonesia, Chad, South Sudan), add a higher level of validity to national NTD control programs through a WHO endorsement, and provide countries opportunities to share lessons learned and best practices with regard to NTD control.
While countries are making substantial progress with regard to conceptualizing NTD control programs, more work is needed. Partnerships between international organizations and national administrative structures may have a role to play in expanding NTD control, as several examples exist whereby vertical programs have been implemented across various countries [40, 41]. Cross-border and regional approaches may have further advantages, particularly where there are high levels of human movement across political boundaries, and allowing countries to benefit from economies of scale for implementation. The Central African Republic, Chad, and Sudan may be especially good candidates for these partnerships as all three of these countries are in close geographic proximity, have similar NTD burdens, and current NTD control efforts are dominated by vertical programs. This regional approach for integration represents one iteration of creating new international initiatives as a pathway for achieving integrated NTD control programs (Fig 3).
Other non-administrative routes may also exist for the integration of NTD control. India, for example, has several NTD control programs, but only leishmaniasis and LF control are integrated under the National Vector Borne Disease Control Programme. Still, in reality, the implementation of these control efforts is conducted by the same personnel who conduct control efforts for other NTDs such as leprosy and STH. Thus, although only two of these NTDs are formally integrated at a higher administrative level, the realities of implementation at a local or community level may nonetheless result in close coordination between control efforts, and thus mirror an integrated approach.
It is also important to consider potential obstacles for NTD control programs, as various factors can influence the delivery of health services. For example, geographical demands, poverty numbers and distribution, resource limitations, and political dynamics can all affect service delivery [42]. For the Central African Republic, Chad, and Sudan, underlying contextual factors may further determine the ability or inability to integrate disease control programs. For example, these countries have all suffered significant instability and civil unrest in recent years. Not only does this civil strife fuel the spread of NTDs–as it is difficult to implement NTD control programs in conflict zones and other non-permissive environments–but it may also deter foreign assistance [43]. Foreign assistance programs may be less willing to implement activities in zones with perceived security risks, or where there is a high chance of interruption of the program due to a renewal of conflict. There may be other restrictions placed on program implementation; for example, the US embargo on Sudan restricted the provision of certain medical equipment and supplies [44]. The sanctions may also have discouraged partners who were concerned with falling afoul of US law. The lifting of sanctions in 2017 [45] points to an opportunity to re-enlist US-based organizations to support NTD control in Sudan and integrate efforts with other priority public health activities.
Effective, integrated responses will also require improved intersectoral collaboration. Brazil launched their national integrated neglected tropical disease plan in 2012 and linked it to the national plan for poverty reduction. In doing so, the country formalized the links between poverty and NTD, which have facilitated implementing effective cross-sector approaches [46]. In another example, India has integrated control programs for soil-transmitted helminths with school health and nutrition programs. This intersectoral collaboration between health and education has acted to expand the reach of NTD programs–improving the health of children across the country [46]. NTD programs have also garnered strong global support–spurring partnerships between governments in NTD endemic countries, international agencies, pharmaceutical companies, international nongovernmental organizations, academia, civil society and United Nations agencies [7]. These collaborations must continue for the full benefits of NTD integration to be realized.
Silo-ed funding streams and the organizational structure of implementation partners may also pose challenges for the integration of NTD control efforts. Earmarked aid from development partners or the organizational structure of implementers may prevent countries from integrating efforts even if they wish to do so. This practice represents a clear divergence from the 2005 Paris Declaration on Aid Effectiveness, which lists alignment as one of the five fundamental principles for making aid more effective [47]. At a more theoretical level, reliance on international aid may also threaten the long-term sustainability of NTD control efforts, in cases where the implemented programs are donor-driven and do not promote country ownership. This could result in a situation in which progress toward integration achieved by MOHs is nullified in the event that development aid is withdrawn.
This study is subject to limitations. Although the countries considered spanned a large geographical range, most were located on the African continent–specifically in sub-Saharan Africa–which may limit the broader applicability of the findings. As our literature review was limited to online sources in English and French, it is possible that we missed information about programs that have been published in other languages, only available in hard copy, or not publicly available, which may have resulted in publication bias. The purposive sampling of personnel involved in NTD control, used for the key informant interviews, may have also biased results, although the primary objective of these interviews was to validate results from the literature review as opposed to primary data collection. Despite these limitations, the results and subsequent discussion presented in this study undoubtedly contribute to a better understanding of the administrative frameworks utilized for the integration of NTD programs.
Moving forward, it is of the utmost importance for advocates of NTD integration to clearly articulate the potential monetary and resource benefits of integration to high-level decision makers to garner political support. These advocates must also take into account the existing administrative structures and creatively engage them to manage the coordination of NTD programs. Finally, it is imperative for development partners to recognize the importance of NTD integration and align their own priorities with national or regional NTD integration efforts where appropriate. In parallel, research efforts should continue to analyze the successes and challenges of integration of disease control programs, in order to produce a robust evidence-base that can support additional refinement of standards and recommendations for future integration.
|
10.1371/journal.pbio.1002072 | Effective Connectivity of Depth-Structure–Selective Patches in the Lateral Bank of the Macaque Intraparietal Sulcus | Extrastriate cortical areas are frequently composed of subpopulations of neurons encoding specific features or stimuli, such as color, disparity, or faces, and patches of neurons encoding similar stimulus properties are typically embedded in interconnected networks, such as the attention or face-processing network. The goal of the current study was to examine the effective connectivity of subsectors of neurons in the same cortical area with highly similar neuronal response properties. We first recorded single- and multi-unit activity to identify two neuronal patches in the anterior part of the macaque intraparietal sulcus (IPS) showing the same depth structure selectivity and then employed electrical microstimulation during functional magnetic resonance imaging in these patches to determine the effective connectivity of these patches. The two IPS subsectors we identified—with the same neuronal response properties and in some cases separated by only 3 mm—were effectively connected to remarkably distinct cortical networks in both dorsal and ventral stream in three macaques. Conversely, the differences in effective connectivity could account for the known visual-to-motor gradient within the anterior IPS. These results clarify the role of the anterior IPS as a pivotal brain region where dorsal and ventral visual stream interact during object analysis. Thus, in addition to the anatomical connectivity of cortical areas and the properties of individual neurons in these areas, the effective connectivity provides novel key insights into the widespread functional networks that support behavior.
| The cortex of primates consists of many areas that are highly interconnected, forming widespread functional networks engaged in specific tasks. Cortical areas frequently consist of submodules, columns, or patches of neurons that share functional properties. The neuronal characteristics of such clusters of neurons are determined by their inputs (i.e., from which neurons they receive information) and outputs (i.e., to which neurons in other brain areas they project), but detailed information about the connectivity of small clusters of neurons is frequently lacking. We applied electrical microstimulation during functional magnetic resonance imaging to chart the connectivity of small patches of neurons in the Intraparietal Sulcus, a brain region that has been implicated in many cognitive operations, such as motor planning, spatial attention, 3-D vision, and grasping. We observed that the three patches of neurons we studied were embedded in very distinct functional networks, covering almost the entire cortex. The network of brain areas connected to each patch could, in turn, explain the properties of the neurons in that patch. Thus, the connectivity of clusters of neurons provides crucial information to understand how functional brain networks support behavior.
| Extracellular recording studies have provided detailed information on the properties of individual neurons and neuronal populations during task performance, which can be correlated with [1,2], and even causally related to, behavior [3,4]. However in order to fully understand the function of neurons in any given brain area and how these neurons subserve behavior, one also needs information about their anatomical connectivity, i.e., from which areas these neurons receive information (input) and to which areas they project (output). Anatomical tracer studies provide a general roadmap of connectivity but cannot identify how specific types of visual information are transmitted between different levels in the cortical hierarchy, since most far extrastriate cortical areas are highly heterogeneous and frequently contain specialized modules for different types of visual information or cognitive processes [5–7]. Functional magnetic resonance imaging (fMRI) provides a static bird’s-eye view of cortical activations elicited by specific stimuli or tasks [8], yet this indirect measure of brain activity cannot in itself determine how the different nodes of the network are connected and how information flows between these different nodes.
Electrical microstimulation in monkeys during fMRI (EM-fMRI) allows the study in vivo of how neural systems are connected (i.e., effective connectivity [9–14]) at a scale of patches or clusters of neurons. However, no study has used this approach to investigate the areas in the macaque intraparietal sulcus (IPS), which have been implicated in a large number of cognitive processes such as motor planning, spatial attention, decision, reward, timing, 3-D vision, and even categorization [15–21]. In this study, we wanted to relate function to connectivity by implementing single-cell recordings and EM-fMRI in the anterior lateral bank of the IPS. We first identified patches of neurons encoding the depth structure of objects in the anterior intraparietal area (AIP) and subsequently performed EM-fMRI experiments in these functionally defined patches of neurons. Although neurons in anterior and posterior AIP showed highly similar neuronal selectivity, we observed markedly distinct networks of cortical areas in occipital, parietal, frontal, and temporal cortex when stimulating each of these subsectors; anterior AIP was embedded in a somatomotor network, while posterior AIP was connected to areas involved in object processing. Our results demonstrate that the posterior subsector of area AIP may be a critical site of convergence of dorsal and ventral stream object information.
We recorded fMRI-guided single-unit activity (SUA) along the lateral bank of the IPS prior to the EM-fMRI sessions (see overview of EM/recording-positions in S1 Table, S1 Fig. for electrode locations). In two monkeys (M and K), we identified two grid positions in AIP with a high proportion of neurons selective for disparity-defined depth structure (e.g., convex versus concave) (424 recording sites in total, in 22 grid positions), one in anterior AIP (aAIP) and one in posterior AIP (pAIP, [22], Fig. 1A and B). In both subsectors of AIP, a high proportion of the neurons (45% to 65%) preserved their preferences for the depth structure of surfaces (3-D shape) across positions in depth [21], as illustrated by the example neurons in Fig. 1A and B. The aAIP and the pAIP patches contained neurons with highly similar selectivities for disparity-defined curved surfaces. Furthermore, consistent with previous research, we confirmed the presence of object-selective responses (single and multi-unit activity (MUA), monkeys M, K, and C, [23]) and grasping activity in pAIP (monkey C, [24]). As a control, we also recorded spatially selective saccadic activity (SUA and MUA; one-way ANOVA with factor target position; p < 0.01) in neighboring lateral intraparietal area (LIP) using a visually guided saccade task, in which the saccade target was positioned at seven to ten different locations in the contralateral visual hemifield (monkeys M, K, and T) (example neuron in Fig. 1E; average spike rate during visually guided saccades towards seven different target locations on the screen). Thus, electrophysiological recordings identified three functionally distinct stimulation sites covering two-thirds of the anterior-posterior extent of the lateral bank of the IPS.
We stimulated aAIP in three different animals (M, K, and C), in ten scan sessions (87 runs × 245 functional volumes, S1 Table for overview, S1 Fig. for electrode locations, data in [25]). Fig. 2A (left column) shows the t-score maps overlaid on coronal sections for monkey M during sedation (contrast EM versus NoEM; p < 0.001, uncorrected; n = 17 runs). Focal increases in fMRI-activity were observed in area AIP, consisting of both aAIP and pAIP (Fig. 2A, left column, first and second row), in the anterior lateral bank of the IPS. Furthermore, aAIP-EM elicited significant fMRI activations in the medial bank of the IPS (area MIP), in area PFG in the rostral portion of the inferior parietal lobule, in the most anterior sector of somatosensory area S2 and in ventral premotor cortex (PMv, or area F5; Fig. 2A, left column, rows 2–5). The fMRI activation evoked by aAIP-EM in PMv was located in the posterior bank of the inferior ramus of the arcuate sulcus, comprising both F5p and F5a [26].
Remarkably similar results were obtained when aAIP was electrically stimulated while the animal was awake and performing a task: the same somatomotor network consisting of pAIP, MIP, PFG, S2, and F5 was activated (Fig. 2A, second column, n = 41 runs). To quantify the similarity between awake and sedated fMRI-EM in monkey M, we considered 32 pre-defined regions of interest (ROIs) throughout the cortex (see Materials and Methods) and calculated the correlation between the percentage of significant voxels (p < 0.001, uncorrected) per ROI in both states (awake/sedated; data averaged over runs). Awake fMRI-EM correlated strongly (Pearson correlation: 0.78; permutation test, p = 0.0002) with fMRI-EM during sedation in the same animal, which allowed us to combine the data from the awake and sedated states in the group analysis. Similarly, the average t-value per ROI in the awake state correlated strongly with the average t-values in the sedated state (Pearson r = 0.85, permutation test: p < 0.0001).
The effects of aAIP-EM were not only similar in the awake and sedated states, but also in different individual animals (compare results obtained from monkey M in first and second column with results from monkeys K-awake and C-sedated in the two rightmost columns of Fig. 2A): although the centers of the activations varied slightly among animals, aAIP-EM consistently elicited fMRI activations in areas AIP (both aAIP and pAIP), MIP, PFG, S2, and F5 in all three animals. The percentage of voxels significantly activated (in the 32 pre-defined ROIs, see Materials and Methods) by aAIP-EM in monkey M was highly correlated with those in monkeys K (Pearson r = 0.60, p = 0.01) and C (Pearson r = 0.72, p = 0.0042). Moreover, the average t-value per ROI in monkey M correlated closely with those obtained in monkeys K (r = 0.36, p = 0.04) and C (r = 0.78, p = 0.01).
The activation pattern evoked by aAIP-EM was also evident in the group average (fixed effects analysis on all 87 runs; Fig. 2B, average of three monkeys and awake/sedated, p < 0.001 uncorrected; see also S2A Fig. for coronal sections). Note that qualitatively similar results were obtained when including the same number of runs per animal (S3 Fig.). An ROI-based analysis showed significant increases in percent signal change (PSC) in areas AIP, MIP, PFG, S2, and F5p during aAIP-EM compared to no-EM (Fig. 2C, t-test; p < 0.05 corrected for multiple comparisons [32 ROIs]; No-EM is set as the baseline with a zero-value), but not in any of the temporal, occipital or prefrontal ROIs. Note that we did not obtain a significant effect in the ROI of F5a (t-test; p = 0.37) in the group data, most likely because aAIP-EM activated only a fraction of F5a (Fig. 2B, bottom row). The group data in Fig. 2B also illustrate that the strongest activation in PMv during aAIP-EM was located in area F5p. aAIP-EM did not activate subcortical structures except the putamen (Fig. 2B inset, white arrow), even when the statistical threshold was lowered. Furthermore, the group analysis did not show an effect of aAIP-EM on the contralateral hemisphere (see examples in Fig. 2A, group results in S2A Fig.; PSC in S4A Fig.).
To assess the specificity of our aAIP-EM results, we performed a similar analysis of PSC on ROIs which are not connected to AIP (motor areas F1, F2, F3, F4, F6, and F7; note that early visual areas V1, V2, and V3 in Fig. 2C are also not connected to AIP [27]). We observed no significant increase in PSC in the latter areas during aAIP-EM (S5A Fig.), and the percentage of activated voxels in these areas was very low (S5B Fig.).
To verify the consistency of our results across animals we performed a conjunction analysis on the aAIP-EM data of all individual animals and states (at p < 0.05 uncorrected for each animal). The network connected to the anterior subsector of AIP consisted of MIP, PFG, and S2 (S6A Fig.). Although we observed significant AIP-EM–induced activations in area F5 in each animal, the conjunction analysis did not contain F5 due to interindividual differences in the location of the F5 activations (see also Fig. 2A, bottom row) and the very localized character of activations within F5p and F5a. To test the reciprocity of the anatomical connectivity of aAIP, we also stimulated in two target areas of aAIP, namely, PFG (monkey M, 2 sessions, 17 runs) and MIP (monkey K, 14 runs). As expected, PFG-EM activated AIP, S2, and F5p, and MIP-EM activated AIP, PMd, and S2 (S7 Fig.), consistent with the existence of reciprocal connections at this level in the hierarchy of cortical areas [28]. Thus, the most anterior subsector of area AIP—where neurons encoding the depth structure of objects were recorded—is connected to a network of somatosensory and motor areas implicated in reaching and grasping [29–31].
Although neuronal characteristics in pAIP were highly similar to those in aAIP, EM of pAIP activated a network of cortical areas that was markedly distinct from the network activated by aAIP-EM (eight scan sessions, 61 runs × 245 functional volumes; see S1 Table for overview per animal; see S1 Fig. for electrode locations; data in [25]). Application of EM to pAIP induced fMRI activations throughout the lateral bank of the IPS; not only in pAIP itself, but also in areas aAIP, LIP, the more posterior subsector of MIP, and in the caudal intraparietal area (CIP) (Fig. 3A, first three rows). In addition, we consistently obtained stimulation-induced activations in the temporal lobe, which was not observed during aAIP-EM (Fig. 3A, third and fourth row): in the lower bank of the superior temporal sulcus (STS), corresponding to the dorsal and ventral part of the posterior inferotemporal cortex (PITd and PITv); the occipitotemporal area (OT); the anterior part of the inferotemporal cortex (TE); the fundus of the STS (FST); and the contralateral STS (Fig. 3A, third row). The pattern of EM-induced fMRI activations was also distinct in the frontal lobe: in contrast to aAIP-EM, stimulating pAIP elicited significant activations in area 45b (in the anterior bank of the inferior ramus of the arcuate sulcus) and in area 46v (Fig. 3A, last two rows). Finally, pAIP-EM also caused scattered activations in and around the lunate and inferior occipital sulcus, corresponding to areas V3 and V4, and even in parts of primary visual cortex V1. As was observed for aAIP-EM, microstimulation of pAIP elicited similar results during awake and sedated experiments in the same animal (Monkey M: Fig. 3A, first and second columns, correlation between the percentage of significantly-activated voxels induced by pAIP-EM during awake versus sedated fMRI: 0.62, p = 0.0014). Similar results were obtained in monkeys C (sedated) and K (sedated; Fig. 3A, third and fourth columns; Pearson r = 0.66; permutation test: p < 0.0001; and r > 0.30; p < 0.05 between C and M, and K, and M). Likewise, the average t-values per ROI in monkey M showed a high degree of correlation between the awake and sedated state (r = 0.57, p = 0.0018), and between monkeys C and M (r = 0.73; p < 0.0001) and K and M (r = 0.52; p = 0.003) respectively. The consistency of the pattern of activations elicited by pAIP-EM across animals was confirmed by a conjunction analysis across all three animals, which showed activations in CIP, LIP, pAIP, aAIP, FST, middle temporal area (MT), the ventral part of medial superior temporal area (MSTv), PIT, OT, TE, and 45B (S6B Fig.).
Group data (fixed-effect analysis including all 61 runs; qualitatively similar results were obtained when including the same number of runs per animal, S3B Fig.) of the effect of pAIP-EM are shown in Fig. 3B (see S2B Fig. for coronal sections). In general, pAIP-EM did not evoke significant activations in subcortical structures such as the basal ganglia and the cerebellum. However, in contrast to aAIP, a restricted part of the pulvinar—possibly corresponding to the dorsal pulvinar—was significantly activated by pAIP-EM (Fig. 3B, inset, white arrow). Furthermore, we observed significant activations in posterior parietal cortex (AIP, LIP, CIP in the lateral bank, and MIP and posterior intraparietal area [PIP] in the medial bank of the IPS); in prefrontal areas 45B and part of 46v; and extensive activations in temporal areas FST, PITd/v, OTd, and TE (see S2B Fig. for coronal sections). The PSC was significantly greater than zero in all aforementioned ROIs during pAIP-stimulation compared to no-stimulation (t-test; p < 0.05, corrected for multiple comparisons [32 ROIs]; Fig. 3C), except for 46v (p = 0.37, most likely due to the relatively small part of the area activated by pAIP-EM). In contrast to aAIP, pAIP-EM also elicited contralateral activations in temporal cortex (Fig. 3A, top three rows; S2B Fig.; S4B Fig. for PSC), including areas FST, PITd/v, and OTd.
As for aAIP-EM, a similar analysis of PSC was performed on ROIs which were previously found not to be connected to AIP (motor areas F1, F2, F3, F4, F6, and F7). No significant increases in PSC in the latter areas were measured during pAIP-EM (S5A Fig.), and the percentage of activated voxels was very low in the latter ROIs (S5B Fig.). Conversely, the only cortical area found to be connected to AIP as described by tracer studies [27] that was not activated in the current study was the dorsal part of parieto-occipital area V6A.
To quantify the difference in effective connectivity between aAIP and pAIP, we performed an ANOVA on the PSC evoked by EM versus No-EM in all voxels of the 32 predefined ROIs (Materials and Methods). The interaction between the factors Stimulation [EM/NoEM] and area [aAIP/pAIP] was significant (p < 0.05) for early visual areas (V1, V2, V3, V3a, V4, V4t, V4A, and V6), temporal areas (OTd, PITd, PITv, TE, FST, MSTv, MT, and STP), prefrontal areas 45b and 45a and parietal areas (S2, MIP, LIP, PIP, and CIP), but not for AIP (see S2 Table). Hence the ROI of AIP was equally activated by aAIP- and pAIP-EM. These results remained essentially unchanged after correcting for the different number of runs (61 for pAIP-EM and 87 for aAIP-EM). Similarly, a conjunction analysis (p < 0.01 uncorrected, S2C Fig.) showed that areas AIP, part of area MIP, and a small region in the arcuate sulcus corresponding to F5a were the only areas activated by both aAIP-EM and pAIP-EM.
Taken together, electrical microstimulation in two functionally defined subsectors of area AIP—with very similar neuronal properties and in some cases separated by no more than 3 mm—activated markedly different networks of cortical areas in parietal, temporal, and frontal cortex.
As a control, we measured the effective connectivity of neighboring area LIP. Data were collected in three monkeys (K, M, and T, see S1 Fig. for electrode locations), in six sessions (60 runs × 245 functional volumes, see S1 Table for overview, data in [25]).
LIP-EM enhanced fMRI-activations in parietal areas LIP, MIP, CIP, and also in area FST in the fundus of the STS (Fig. 4A, group data in Fig. 4B, S2D Fig.). Furthermore, LIP-EM evoked focal increases in Frontal Eye Fields (FEF)-activity in individual stimulation sessions (three out of six sessions; illustrated in Fig. 4B, upper row). As with the aAIP and pAIP experiments, we obtained comparable results in the sedated and awake states, and similar results in all animals individually (Fig. 4A, correlations between percentage of significant voxels per ROI per monkey/state in 32 predefined ROIs: Pearson r > 0.58, permutation test: p < 0.0066; correlations between average t-values per ROI per monkey/state in 32 predefined ROIs: r > 0.62, p < 0.002). The consistency of the activation patterns elicited by LIP-EM across animals was confirmed by a conjunction analysis (S6C Fig.) across all three animals, which showed activations in CIP, MIP, FST, and V3A.
The group analysis of the LIP-EM experiments revealed a small but significant stimulation-induced increase in activity in FEF (Fig. 4C), in parietal areas LIP, MIP, CIP, and PIP and in temporal area FST. Note that qualitatively similar results were obtained when including the same number of runs per animal (S3C Fig.). Similarly, LIP-EM significantly increased the PSC in parietal areas LIP, MIP, CIP, and PIP, and in temporal area FST (Fig. 4D, t-test, p < 0.05, corrected for multiple comparisons). In contrast to pAIP-EM, LIP-microstimulation did not modulate the PSC in visual areas V4 and V4A, in temporal areas TE and PITd/v, nor in parietal area AIP or prefrontal area 45B (two-way ANOVA with factors microstimulation [EM versus no-EM] and area [pAIP versus LIP]; interaction: p < 0.05; S2 Table). Thus the IPS sector characterized by the presence of spatially selective saccadic activity (LIP) is effectively connected to a network of cortical areas that only partially overlaps with that of neighboring pAIP.
To our knowledge, our study provides the first causal evidence relating the properties of individual neurons to their effective connectivity in posterior parietal cortex. fMRI-EM in three functionally defined patches of neurons in the lateral bank of the IPS revealed distinct networks of cortical areas in parietal, temporal, and frontal cortex.
Our EM-fMRI results obtained in AIP are highly comparable to earlier work using traditional tracer injections in AIP [27], although the latter authors did not distinguish between anterior and posterior AIP. Our current results are in line with a previous EM-fMRI study, in which FEF-EM showed increased fMRI activations in areas previously found to be anatomically connected to area FEF [11,32,33] (see [34] for review). It is important to emphasize that almost all areas activated during fMRI-EM in aAIP and pAIP are monosynaptically and reciprocally connected with AIP [26,27,35–39]. The only possible exception may be the contralateral PITd/v and OTd activations elicited during pAIP-EM. Conversely, areas that are not directly connected to AIP (e.g., F1–F4, F6–F7, and V1–V3) were also not activated during EM-fMRI in AIP. Only very few brain regions for which AIP connections have been reported were never activated during EM-fMRI: examples include area V6Ad, which is weakly connected to AIP [27], and the cerebellum, which is connected to the possible homologue of AIP in the cebus monkey [40]. Note that with the current resolution of monkey fMRI, it is not possible to make claims about the laminar distribution (supra- versus infragranular) of the connections; hence, it is not possible to make conclusive inferences about feedforward versus feedback connections.
Unlike most tracer studies, we combined extensive single-cell recordings and effective connectivity measurements. However, despite the striking similarity between tracer studies and our results, the crucial advantage of in vivo EM-fMRI (or EM-optical imaging [41]) over tracer studies is that it allows the identification of the connections of specific clusters of neurons (in our case 3-D–shape selective neurons within AIP) with subsectors of other areas (e.g., in frontal cortex), which can then become the target of detailed investigations using (a combination of) single-cell recordings, EM-fMRI, and reversible inactivation during fMRI (see [34]). Thus, in vivo effective connectivity studies furnish the possibility to investigate neural populations for which the inputs and outputs have been accurately identified in the animal under study (possibly even without existing anatomical data), so that the signals can be traced throughout the hierarchy of extrastriate areas from occipital to frontal cortex (see also [11]). Moreover, the same in vivo procedure can be repeated for virtually unlimited numbers of target areas within the same subject. Furthermore, EM-fMRI may also provide important information for the interpretation of the behavioral effects of microstimulation, since a general overview of the effective connectivity of a cortical stimulation site can reveal which downstream areas influence behavior.
Previous tracer studies have shown connections between area LIP and many other (sub)cortical areas such as FEF, 46, parieto-occipital cortex (PO), dorsal prelunate area (DP), 7a, V3, V4, MT, MST, posterior area in inferotemporal cortex (TEO), and superior colliculus [36,37,42,43], in which the ventral stream areas are primarily connected with dorsal LIP while MT/V5 and FEF are primarily connected with ventral LIP [36]. Our current study, however, shows mainly increased activation in temporal area FST, and to a lesser extent in FEF. The difference between our current EM study and previous anatomical tracer studies was not surprising, given the known heterogeneity of area LIP (e.g. [7,44,45]) and the fact that we only stimulated a single site per animal in LIP as a control for pAIP. Moreover, it is conceivable that previous studies (e.g., [46]) have identified our pAIP site, located 9 to 15 mm from the anterior tip of the IPS and with foveal RFs, as anterior LIP (most likely also explaining the connections between area 45B and LIP [38]), whereas we observed 45B activations during pAIP but not LIP stimulation. In addition, our pAIP stimulation site was also located more dorsally in the lateral bank of the IPS, and previous studies have reported that the anterior part of dorsal LIP is strongly connected to ventral stream areas and to area 45B [36].
Other studies [47,48] have investigated effective connectivity in the motor system using a combination of EM and single-cell recordings and have demonstrated that EM can elicit both excitation and inhibition in the target neurons, particularly at higher current strengths. Although marked behavioral effects can be observed with relatively low intensity EM (25–35 μA) in extrastriate cortex [3,49], our pilot experiments showed no reliable fMRI activations when we stimulated with currents below 200 μA (awake) or 1,000 μA (sedated). Since the fMRI signal represents a sum of excitatory and inhibitory activity [50], we probably evoked both effects in our EM-fMRI experiments. It needs to be noted that previous FEF-EM experiments revealed extensive effective connectivity networks with currents below 50 μA in awake animals [11], which could also be obtained by optogenetic stimulation of the same areas [51].
The patterns of activations we observed were highly similar in awake and sedated sessions, similar to lateral geniculate nucleus (LGN)-EM [50]. This observation has important practical implications because it shows that it is possible to chart the connectivity of functional patches of extrastriate neurons in monkeys engaged in single-cell experiments but not accustomed to the scanner environment. However, the strong correspondence between awake and sedated sessions does not necessarily imply that EM-induced activations cannot be stimulus- or task-dependent (e.g., in the Frontal Eye Fields, [11,52]). We did not obtain sufficient data in the current experiment in the awake state to draw definitive conclusions, but future studies should investigate the task-dependency of the fMRI activations evoked by AIP-EM. Importantly, the possibility remains that similarities between the awake and sedated EM-fMRI results are region-specific.
Our effective connectivity results help to understand several anatomical and physiological observations. First, neurons in the posterior part of AIP tend to be more visual, whereas neurons in anterior AIP tend to be more motor-dominant [53], and this visual-to-motor gradient in AIP can now be linked to the connectivity of pAIP (object processing network including the ventral stream) and aAIP (somatomotor network). Our results clearly demonstrate the status of pAIP as a pivotal brain area where dorsal and ventral visual stream interact during object analysis. Before contact with the object, the anterior IPS regions may access information about object identity—which may assist in selecting the appropriate grasp—through these connections with the ventral stream [27,54]. This distinction between pAIP and aAIP, made possible using both effective connectivity and physiological assessments, could not be readily detected in anatomical studies [27]. Future studies should determine to what extent and under which conditions (e.g., immediate versus delayed actions [55,56]) these dorsal–ventral stream interactions become behaviorally relevant.
Secondly, both the aAIP and the pAIP patch contained a high proportion of 3-D–shape selective neurons. The patterns of effective connectivity we observed strongly suggest that 3-D–shape information is transmitted from pAIP to aAIP and subsequently to the motor system. This stream of 3-D–shape information runs along the lateral bank of the IPS and interacts with the ventral stream at the level of pAIP. Note that 3-D–shape selective clusters in AIP are also active during object grasping [57], and that reversible inactivation of these AIP clusters induces a grasping deficit (Verhoef and Janssen, unpublished observations). Therefore our results, in concert with previous findings in AIP, also contribute to our understanding of the organization of the 3-D–shape network.
Finally, the anterior-posterior extents of areas AIP and LIP have been a long-standing controversy, in which anatomical studies have conflicted with physiology studies [42,45,46]. Specifically, the region in the lateral bank of the IPS where spatially selective saccadic activity—interspersed with memory delay-period activity during saccades [7]—can be recorded (i.e., area LIP) is mostly confined to the posterior third of the lateral IPS [42]. In contrast, anatomical studies have claimed, based on the pattern of myelination of LIP [43,45], that LIP occupies a large part of the lateral bank of the IPS. Our data finally suggest a resolution for this issue. The functional properties of individual neurons at the pAIP stimulation site, located 9 to 14 mm posterior to the tip of the IPS, resembled those of aAIP neurons given the presence of strong grasping activity, 3-D–shape selectivity, and the absence of saccadic activity [24,58], and the most posterior tracer injections in AIP in [27] were also located 15 mm posterior to the anterior tip of the IPS. However, the connectivity of pAIP with frontal and temporal areas was strikingly distinct from that of both aAIP and LIP. Therefore, a functional parcellation of the lateral IPS would identify the anterior two-thirds of the lateral IPS, where grasping activity can be recorded, as AIP [53,59,60], and the posterior one-third of the lateral IPS, where spatially selective saccadic activity can be recorded, as LIP. Monkey fMRI studies have also demonstrated a representation of the fovea in the anterior lateral bank of the IPS [61,62], and in humans, two regions in the anterior IPS (DIPSA and DIPSM) are activated more strongly by curved surfaces compared to flat surfaces at different positions in depth [63], which may be homologous to monkey aAIP and pAIP.
Our results are also consistent with human fMRI studies on grasping and object processing. Although considerable differences exist between the human IPS and the macaque IPS (related to, e.g., 3-D structure-from-motion and tool use, [64–67]), the more anterior IPS sectors appear more action-related, whereas the more posterior IPS sectors are more visual [68]. Similarly, resting-state connectivity analysis in humans has indicated that a region in the Lateral Occipital Complex (LOC) responsive to images of hands and tools is selectively connected to the IPS regions involved in action-related processing of hands and tools. It is also noteworthy that distinct patterns of resting-state connectivity can be observed for adjacent seed regions in occipitotemporal cortex [69], similar to the distinct networks we observed when stimulating in different cortical sites that were merely 3 mm apart. Future studies will have to determine the correspondence between functional and effective connectivity as determined with EM-fMRI in the IPS areas, as already achieved in macaque somatosensory cortex [13].
It is remarkable that we observed such distinct networks of cortical areas when stimulating sites that were in some cases separated by no more than 3 mm. Electrical microstimulation at the currents we used undoubtedly activated large numbers of neurons, and most likely not exclusively 3-D–shape, object, or saccade selective neurons. Since the 3-D–shape patches we identified in AIP measured merely 1–2 mm (i.e., one or two grid positions) but were very homogeneous (containing up to 80% 3-D–shape selective neurons [21]), the effect of EM in these patches must have been dominated by the connectivity of 3-D–shape selective neurons. Not surprisingly then, most of the cortical areas connected to the aAIP and pAIP stimulation sites are sensitive to the depth structure of objects [3,5,21,70–75]. Area FST was effectively connected to both pAIP and LIP, consistent with anatomical studies [27,37]. Since FST neurons encode 3-D–shape defined by structure-from-motion (SFM) [76], and in view of the fact that both FST and AIP are selectively activated by SFM-defined 3-D surfaces compared to control stimuli (Mysore, Vogels, Vanduffel, and Orban, unpublished observations), the pAIP-FST connection may be important for the integration of two of the most powerful depth cues, binocular disparity and SFM.
The patterns of connectivity we observed appeared to result mostly from feedforward (e.g., aAIP to F5) and lateral (e.g., aAIP to PFG [27,30]) projections. However, in each stimulation site, EM activated its most likely input areas (feedback): aAIP-EM activated its input area pAIP, pAIP-EM activated LIP and CIP, and LIP-EM activated CIP and V3A, a pattern that is entirely consistent with an earlier anatomical tracer study demonstrating that the main connectivity pattern in the lateral IPS runs from posterior to anterior, from V3A to CIP–LIP–AIP [35]. Visual object information is then send to the motor system (F5p and F5a) and to the somatosensory system (area S2), an area connected to both AIP and F5 where many neurons respond during active hand manipulation of objects but not during passive hand stimulation [77]. Thus, charting the effective connectivity of functionally defined subsectors of areas or patches of neurons in the IPS provides crucial insight into the organization of cortical networks that support behavior.
All experimental procedures were performed in accordance with the National Institute of Health’s Guide for the Care and Use of Laboratory Animals and EU Directive 2010/63/EU, and approved by the Ethical Committee at the KU Leuven. The animals in this study were pair-housed with cage enrichment (toys, foraging devices) at the primate facility of the KU Leuven Medical School. They were fed daily with standard primate chow supplemented with nuts, raisins, prunes, and fruits. The animals received their daily water supply either during the awake experiments, or ad libitum in the cages before and after sedated experiments.
All experiments were performed in four male rhesus monkeys (C: 8 kg; K: 6 kg; M: 5 kg; T: 6 kg). All animals had a custom-made, magnetic resonance imaging (MRI)-compatible headpost and cylinder implanted on the skull using ceramic screws and dental acrylic. All surgeries were performed under isoflurane anaesthesia and sterile conditions. The cylinders were implanted in an oblique orientation (orthogonal to the IPS in monkey C, parallel to the IPS in monkeys M, K, and T) over the IPS at Horsley-Clark coordinates ranging from 10 to 0 P and from 10 to 20 L. In monkey M, the recording cylinder was repositioned before the fMRI-EM experiment in aAIP from an orientation orthogonal to the IPS (S1 Fig., upper row, red arrow) to an oblique orientation parallel to the IPS to allow electrode penetrations parallel to the IPS, targeting the aAIP patch as defined by its neuronal characteristics. Three monkeys (K, M, T) were trained in passive fixation and saccade tasks in a mock fMRI-setup. They were seated in a sphinx position [78] in a plastic monkey chair directly facing an LCD screen (viewing distance: 57 cm). Eye position was monitored at 120 Hz through the pupil position (Iscan, MA, United States). The fourth monkey (C) was scanned only under sedation.
All stimuli were displayed on a CRT monitor (Vision Research Graphics, equipped with P46 phosphor) operating at 120Hz.
Stereo test. The stimulus set of the stereo experiment consisted of random-dot stereograms in which depth was defined by horizontal disparity (dot size 0.08 deg, dot density 50%, vertical size 5.5 deg) presented on a grey background [70]. All stimuli were generated using Matlab (MathWorks) and were gamma-corrected. The stimuli in the search test consisted of three types of smoothly curved depth profiles (1, one-half, or one-fourth vertical sinusoidal cycle) together with their antiphase counterparts obtained by interchanging the monocular images between the eyes (disparity amplitude within the surface: 0.5 deg), control stimuli (the monocular images presented to both eyes simultaneously), and flat surfaces at different disparities. Each of the six depth profiles was combined with one of four different circumferential shapes and appeared at two different positions in depth (mean disparity + or—0.5 deg), creating a set of 48 curved surfaces. Ferroelectric liquid crystal shutters (Displaytech) each operating at 60 Hz were used to generate dichoptic presentation. The shutters were synchronized with the vertical retrace of the display monitor. There was no measurable cross-talk between the two eyes [21]. After 200 ms of fixation, the stimulus was presented at the fixation point for 1 s.
In the search test, all stimuli (stereo and control, curved and flat) were presented randomly interleaved at the center of the display and at the fixation plane during passive fixation. Single or multi-unit activity was recorded, and if a site was visually responsive, we isolated single neurons online and tested these neurons in more detail for higher-order disparity selectivity (i.e., selectivity for gradients of disparity) in the position-in-depth test [5]. In this test the stimulus (a combination of a depth profile and a circumferential shape) evoking the highest response in the search test was selected together with its antiphase counterpart, and presented at five different positions in depth ranging from-0.5 degree (near) to +0.5 degree (far) disparity in equal steps.
Object test. Previous studies [23,24,58] have characterized pAIP based on the presence of selective visual responses to images of objects presented foveally during passive fixation. The same stimuli as in [23] were used to confirm the presence of object-selective responses in pAIP in three animals (M, K, and C). The stimulus set for the object test consisted of 21 two‐dimensional (2-D) area‐equalized static images of natural and artificial objects, including faces, hands, fruits, branches, and several artificial graspable objects. The presence of object-selective SUA or MUA responses was assessed using a one-way ANOVA (p < 0.05).
Grasping test. In the visually guided grasping test, a bar attached to a plate was positioned in the monkey’s view. The animal had to rest his right hand on a sensing device in complete darkness for a variable time (inter‐trial interval ITI 3,000–5,000 ms), after which a light inside the object was illuminated, whereupon the monkey had to fixate the object (keeping its gaze inside a ±2.5‐degree fixation window). After a 500 ms fixation period, an audible go‐signal was given for initiating the grasping movement, which consisted of reaching, grasping, and pulling the object on the plate (holding time: 500–900 ms)[24].
Saccade test. In the visually guided saccade task, monkeys had to maintain fixation within a window of 2 × 2 visual degrees around a small green spot in the center of the display for a fixed duration of 450 ms, after which a single green saccade target appeared at one of ten possible positions on the screen, spaced 15 (horizontal) or 11 (vertical) degrees apart. After a variable time, the green fixation spot dimmed, indicating to the animal to saccade towards the target location. The presence of spatially selective saccadic SUA or MUA responses was confirmed using a one-way ANOVA with factor target position (p < 0.001 for all target-selective cells).
Functional images were acquired with a 3.0 T full-body scanner (TIM Trio; Siemens), using a gradient-echo T2*-weighted echo-planar imaging (EPI) sequence (40 horizontal slices; TR: 2s; TE: 16 ms; 1.25 mm3 isotropic voxels) with a custom-built eight-channel phased-array receive coil, and a saddle-shaped, radial transmit-only surface coil [79]. Before each scanning session, a contrast agent, monocrystalline iron oxide nanoparticle (MION) (Feraheme: AMAG pharmaceuticals; Rienso: Takeda) was injected into the femoral/saphenous vein (7–11 mg/kg) [78].
To verify the stimulation positions, structural MR images (0.6 mm resolution) were acquired in every sedated scan session (prior to the start of the fMRI experiment) while the electrode was located at the exact stimulation site inside a standard recording grid (Crist Instruments, Hagerstown, MD, US). In the few sessions in which the latter could not be achieved, we inserted glass capillaries filled with a 2% copper sulphate solution into the grid at several positions, acquired structural MR images (0.6 mm resolution) and reconstructed the electrode penetrations using SPM 5 (Statistical Parametric Mapping).
In every scanning session, a Platinum/Iridium electrode (impedance 50–200 kΩ in situ, FHC, Bowdoinham, ME) was inserted in the grid through glass capillaries serving as guide tubes (Plastics One Inc, Kent, United Kingdom; FHC, Bowdoinham, ME, US). A platinum wire served as ground. The electrical microstimulation (EM) signal was produced using an eight-channel digital stimulator (DS8000, World Precision Instruments) in combination with a current isolator (DLS100, World Precision Instruments). During stimulation blocks, a single EM train was applied in every trial.
In awake scanning sessions, the animals were either fixating a spot on a screen (Fix) or performing memory-guided saccades (Sacc) towards ten different positions contralateral to the stimulated hemisphere. Briefly, during the memory-guided saccade task a saccade target was flashed for 200 ms on the screen, and the animals had to maintain fixation (300–1,500 ms) until the dimming of the fixation point instructed an eye movement to the remembered target location. During the baseline fixation task (Fix0), only a central fixation point was displayed on the screen, while during the control fixation task (Fix1), one distractor (identical to the saccade target in the Sacc task) was shown on the screen with the same position and timing parameters as the saccade target in the memory saccade task. The color of the fixation point indicated to the animals to either maintain fixation or to make saccades. In this study, the data collected during all three tasks were combined. The three tasks were presented to the animals in blocks, and EM was administered during all three tasks, thus creating six types of blocks which were alternated in one run in pseudo-random order. We alternated between stimulation and no-stimulation blocks (each lasting 40 s), with each run lasting 245 pulses (490 s).
Stimulation trains in awake scan sessions lasted 500 ms and were composed of biphasic square-wave pulses (repetition rate 200 Hz; amplitude 200 μA). Note that pilot experiments showed that a current amplitude of less than 200 μA did not evoke increased fMRI-activations. Each pulse consisted of 190 μs of positive and 190 μs of negative voltage, with 0.1 ms between the two pulses (total pulse duration: 0.48 ms). During sedated scanning sessions, a trial-by-trial stimulation protocol was used similar to the awake sessions (one EM train every 3 s, approximately). EM trains in sedated sessions lasted 250 ms with an amplitude of 1 mA, while other EM-parameters remained similar (200 Hz, 0.48 ms pulse duration). The timing of the EM pulses during the fMRI experiment was computer controlled. Note that pilot experiments showed that a current amplitude of 200 μA (= current strength during awake sessions) during sedated sessions only caused increased fMRI-activations around the tip of the electrode.
During sedated scan sessions, a 0.5/0.5 cc mixture of ketamine (Ketalar; Pfizer) and medetomidine (Domitor; Orion) was administered every 45 min. The animals were video-controlled during sedation, and body temperature was maintained using a heating pad.
Off-line image reconstruction was conducted to overcome problems inherent to monkey body motion at 3T. Details about the image reconstruction protocol have been given elsewhere [79]. Briefly, the raw EPI images were corrected for lowest-order off-resonance effects and aligned with respect to the gradient-recalled-echo reference images before performing a SENSE (sensitivity encoding) image reconstruction [80]. Corrections for higher-order distortions were performed using a non-rigid slice-by-slice distortion correction.
Data were analyzed using statistical parametric mapping (SPM5) and BrainMatch software, using a fixed-effect GLM. Realignment parameters were included as covariates of no interest to remove brain motion artifacts. Spatial preprocessing consisted of realignment and rigid coregistration with a template anatomy (M12) [11]. To compensate for echo-planar distortions in the images as well as inter-individual anatomical differences, the functional images were warped to the template anatomy using non-rigid matching BrainMatch software [81]. The algorithm computes a dense deformation field by the composition of small displacements minimizing a local correlation criterion. Regularization of the deformation field is obtained by low-pass filtering. The functional volumes were then resliced to 1 mm3 isotropic and smoothed with an isotropic Gaussian kernel (full width at half maximum: 1.5 mm). Single subject and group analyses were performed, and the level of significance was set at p < 0.001, uncorrected for multiple comparisons. For display purposes, SPM T-maps were presented on coronal or flattened representations of the M12 anatomical template, using xjView toolbox (http://www.alivelearn.net/xjview) and Caret software (version 5.64; http://brainvis.wustl.edu/wiki/index.php/Caret:About), respectively.
The exact locations and extents of the fMRI-activations were verified on the animal’s own EPI-images. Percent signal change was calculated in regions of interest (ROIs), and statistical significance was tested using MarsBaR (version 0.41.1). We considered a set of 32 ROIs for early visual areas and the ROIs of all brain areas connected to AIP [27], which included premotor, prefrontal, parietal, temporal, and visual ROIs (F5a, F5p, F5c, 45A, 45B, 46v, FEF, AIP, LIP, MIP, CIP, PIP, PFG, STP, OT, PITv, PITd, TE, TEr, FST, MSTv, MT, S2, V1, V2, V3A, V3, V4, V4A, V4T, V6A, V6). Moreover, we also included an additional set of ROIs of frontal areas that are not connected with AIP: F1, F2, F3, F4, F6, and F7. Note that the no-stimulation condition served as the baseline. The significance threshold for one-tailed t-tests was set at p = 0.05, corrected for multiple comparisons (32 t-tests calculated; p = 0.05/32 = 0.0016). Standard fMRI analysis methods were used, as described in previous studies [30,52]. All regions of interest were described previously [11,30,62].
To quantify the similarity between the awake and sedated states and between animals, a Pearson correlation was calculated between the percentage of significant voxels (t-value > 3.1: p < 0.001 uncorrected) per ROI in each state (awake-sedated) or in each animal, across the set of 32 ROIs of all early visual areas and all areas connected to AIP. The significance of the correlations between animals was calculated using a permutation test, in which the 32 calculated percentages of significantly (p < 0.001 uncorrected) activated voxels were randomly assigned (5,000 times) to the 32 ROIs, after which the correlations between corresponding ROIs were calculated. P-values were calculated as the proportion of correlations exceeding the actual correlation between corresponding ROIs. Moreover, to confirm the consistency of the activations across animals and states, a conjunction analysis was performed on the data of all animals (at p < 0.05 uncorrected for each animal).
|
10.1371/journal.pgen.1003869 | Dominant Mutations in S. cerevisiae PMS1 Identify the Mlh1-Pms1 Endonuclease Active Site and an Exonuclease 1-Independent Mismatch Repair Pathway | Lynch syndrome (hereditary nonpolypsis colorectal cancer or HNPCC) is a common cancer predisposition syndrome. Predisposition to cancer in this syndrome results from increased accumulation of mutations due to defective mismatch repair (MMR) caused by a mutation in one of the mismatch repair genes MLH1, MSH2, MSH6 or PMS2/scPMS1. To better understand the function of Mlh1-Pms1 in MMR, we used Saccharomyces cerevisiae to identify six pms1 mutations (pms1-G683E, pms1-C817R, pms1-C848S, pms1-H850R, pms1-H703A and pms1-E707A) that were weakly dominant in wild-type cells, which surprisingly caused a strong MMR defect when present on low copy plasmids in an exo1Δ mutant. Molecular modeling showed these mutations caused amino acid substitutions in the metal coordination pocket of the Pms1 endonuclease active site and biochemical studies showed that they inactivated the endonuclease activity. This model of Mlh1-Pms1 suggested that the Mlh1-FERC motif contributes to the endonuclease active site. Consistent with this, the mlh1-E767stp mutation caused both MMR and endonuclease defects similar to those caused by the dominant pms1 mutations whereas mutations affecting the predicted metal coordinating residue Mlh1-C769 had no effect. These studies establish that the Mlh1-Pms1 endonuclease is required for MMR in a previously uncharacterized Exo1-independent MMR pathway.
| Lynch syndrome (hereditary nonpolypsis colorectal cancer or HNPCC) is a common cancer predisposition syndrome. Predisposition to cancer in this syndrome results from increased accumulation of mutations due to defective mismatch repair (MMR) caused by a mutation in one of the mismatch repair genes MLH1, MSH2, MSH6 or PMS2/scPMS1. In addition to these genes, various DNA replication factors and the excision factor EXO1 function in the repair of damaged DNA by the MMR pathway. Although EXO1 is considered to be the major repair nuclease functioning in mismatch repair, the relatively low mutation rates caused by an exo1 deletion suggest otherwise. Here we used genetics, microscopy and protein biochemistry to analyze the model organism Saccharomyces cerevisiae to further characterize a poorly understood mismatch repair pathway that functions in the absence of EXO1 that is highly dependent on the Mlh1-Pms1 complex. Surprisingly, we found that the highly conserved metal binding site that is critical for the endonuclease activity of the Mlh1-Pms1 heterodimer is required for MMR in the absence of Exo1 to a much greater extent than in the presence of Exo1. Thus, this work establishes that there are at least two different polynucleotide excision pathways that function in MMR.
| DNA mismatch repair (MMR) acts to repair the potentially mutagenic misincorporation errors that occur during normal DNA replication and the absence of MMR results in increased rates of accumulating mutations. Consequently, defects in human MMR genes cause the hereditary cancer susceptibility syndrome HNPCC (hereditary nonpolypsis colorectal cancer, otherwise known as Lynch syndrome) [1], [2], and loss of MMR function also appears to underlie the development of some sporadic cancers [3]–[7]. MMR also repairs mispaired bases that occur in recombination intermediates as well as prevents inappropriate recombination between DNAs with imperfect homology where recombination could result in genome rearrangements [8]–[10].
The mechanism of MMR has been extensively characterized in both E. coli and different eukaryotic systems, with E. coli MMR being the best characterized [11]–[14]. In E. coli MMR, mismatches are recognized by the MutS homodimer [15], [16]. Mispair bound MutS then recruits the MutL homodimer [17]. This recruitment leads to activation of the MutH endonuclease, which introduces single strand breaks, called nicks, at unmethylated GATC sites in the newly replicated and hemimethylated DNA strand [18]. Next, a combination of the UvrD helicase and one of four single stranded DNA specific exonucleases excise the nicked strand past the mispair and the resulting singled-stranded gap is filled in by DNA polymerase III, single strand DNA binding protein and DNA ligase [14], [19].
In eukaryotes mispairs are recognized by either Msh2-Msh6 or Msh2-Msh3, two partially redundant heterodimers of MutS family member proteins [12], [20], [21]. Mispair bound Msh2-Msh6 and Msh2-Msh3 recruit the MutL related complex, called Mlh1-Pms1 in S. cerevisiae and Mlh1-Pms2 in human and mouse [11], [12], [22]–[24]. The Pms1/Pms2 subunit of the Mlh1-Pms1/Pms2 complex is known to contain an endonuclease active site, suggesting that Mlh1-Pms1/Pms2 may be analogous to a combination of both E. coli MutL and MutH [25], [26]. Exo1, a DNA exonuclease from the Rad2/XPG family, has been implicated in the excision step of eukaryotic MMR; however, mutations in S. cerevisiae and mouse EXO1 only result in partial MMR defects, suggesting the existence of additional excision mechanisms [27]–[29]. Genetic and biochemical studies have also implicated DNA polymerase δ, RPA, RFC and PCNA in MMR [12], [30]–[37] and have suggested that several of these proteins including PCNA and RFC may function both prior to excision and in the resynthesis steps of MMR [25], [26], [33], [38].
MMR is spatially and temporally coupled to replication in vivo [38], [39], providing a mechanism to bring MMR proteins into the proximity of newly formed mispairs. DNA replication generates nicks in the nascent DNA strands that may be involved in MMR [30], [34], [40], [41], consistent with the observation that discontinuous lagging strand MMR is more dependent on excision catalyzed by Exo1 than leading strand MMR [38], [42]. Furthermore, preexisting nicks in DNA target the Mlh1-Pms1/Pms2 endonuclease to the nicked strand in vitro [43]. However, these results raise two unresolved questions: Why is it necessary to target additional nicks to an already nicked DNA strand, and if preexisting nicks can support MMR in vitro then why is Mlh1-Pms1 absolutely required for MMR in vivo? Part of the answer to the apparent contradictions implied by these experimental results could be the presence of multiple MMR pathways in which the same MMR proteins have differing roles. Consistent with this, biochemical studies have identified two types of excision mechanisms that may function in MMR, excision by Exo1 [44] and strand displacement synthesis toward the mispair by DNA polymerase δ potentially coupled with flap cleavage [45]. Both mechanisms could act at either a pre-existing 5′ nick or a 5′ nick introduced by the Mlh1-Pms1/Pms2 endonuclease. Genetic studies have also identified Exo1-dependent and -independent MMR pathways [27]. The Exo1-independent pathway requires the PCNA-Msh2-Msh6 interaction and the Pol32 subunit of DNA polymerase δ and is inactivated by separation-of-function mutations affecting Mlh1, Pms1, Msh2, Msh3, and PCNA that do not affect Exo1-dependent MMR [27], [38]. How these mutations specifically affect in the Exo1-independent MMR pathway and how this pathway excises the nascent DNA strand is unclear.
To better understand the role of both Mlh1-Pms1 and Exo1 in MMR, we performed a genetic screen for dominant mutations in the PMS1 gene. We identified pms1 null missense mutations that caused weakly dominant MMR defects when present in a wild-type S. cerevisiae strain on a single-copy plasmid. Interestingly, these mutations caused much stronger MMR defects when present on a single-copy plasmid in an exo1Δ mutant. Analysis of these mutations using the structure of the C-terminal domains of Mlh1-Pms1 [46] predicted that three amino acids altered by these mutations were metal ligands in the Mlh1-Pms1 nuclease active site and the fourth was a residue adjacent to the metal binding site. Biochemical analysis of mutant proteins confirmed that both pms1 and mlh1 mutations affecting the predicted active site eliminated or significantly reduced the RFC-PCNA dependent nuclease activity of Mlh1-Pms1 and in vivo imaging showed that the same mutations resulted in accumulation of Mlh1-Pms1-4GFP foci consistent with failure to execute a downstream step in MMR. These results both define the nuclease active site of the Mlh1-Pms1 endonuclease and thoroughly characterize the role of this endonuclease activity in a previously uncharacterized Exo1-independent MMR sub-pathway.
To gain insight into the role of PMS1 in mismatch repair, we sought to generate novel pms1 mutations that cause a dominant MMR defect. First we mutagenized the PMS1 gene by PCR amplification and gap-repaired the resulting DNA fragments into the low copy number ARS CEN pRS316 plasmid by co-transformation into a S. cerevisiae strain with a wild-type PMS1 gene. A total of 38,000 transformants were screened for increased reversion of the lys2-10A frameshift mutation. This screen identified 211 transformants potentially containing dominant pms1 mutations. Rescreening these 211 transformants for increased reversion of the hom3-10 frameshift mutation identified 8 transformants with mutator phenotypes. The pms1 mutation-bearing plasmids were isolated from these 8 transformants, and the PMS1 gene from each plasmid was sequenced. Site directed mutagenesis and sub-cloning were used to construct PMS1 plasmids containing single point mutations. These mutant plasmids were retested in the three mutator assays confirming four dominant pms1 mutations resulting from amino acid substitutions in the C-terminal domain of Pms1: G683E, C817R, C848S, and H850R.
The four amino acid substitutions resulting from the dominant mutations altered highly conserved amino acids (Figure 1A). Initial analysis of these amino acid substitutions using a homology model based on the C-terminal domains of endonuclease-proficient and zinc-binding Neisseria gonorrhoeae and Bacillus subtilis MutL homologs [47], [48] indicated that they affect the active site of the Pms1 endonuclease. Mapping of the amino acid substitutions onto the newly available structure of the C-terminal domains of Mlh1-Pms1 (Figure 1B,C) confirmed that the C817R, C848S, and H850R amino acid substitutions each eliminated one of the 5 metal ligands in the Pms1 endonuclease active site; all 5 ligands are conserved in eukaryotic ScPms1/HsPMS2 proteins and in the N. gonorrhoeae and B. subtilis MutL homologs [47], [48]. The fourth amino acid substitution, G683E, mapped to a conserved position adjacent to sites of metal coordination and could sterically disrupt the site or locally perturb the structure. We also constructed mutations resulting in the amino acid substitutions H703A and E707A to eliminate the remaining two predicted metal ligands not identified in our screen.
Fluctuation analysis was performed to evaluate the mutator effects of the dominant pms1 mutants. Mutation rates were measured using the CanR forward mutation assay and the hom3-10 and lys2-10A frameshift reversion assays [21], [27] when the pms1 dominant mutations were present on low copy plasmids in a strain with a wild-type PMS1 gene (Table 1). None of the mutant plasmids caused more than a 2-fold increased mutation rate in the CanR assay, which has a relatively high background mutation rate in wild-type cells and a low sensitivity for detecting MMR defects. In contrast, the mutations on the low copy plasmid caused between a 2- and 102-fold increase in mutation rate in the hom3-10 and lys2-10A assays compared to introduction of a wild-type copy of PMS1 on a low copy plasmid. Introduction of these mutations onto a high copy 2-micron plasmid resulted in much higher rates (Table S1), but still did not cause mutation rates that were as high as caused by deletion of PMS1 (Compare Table S1 to Table S2). The pms1 mutant plasmids were also unable complement a pms1Δ strain above that seen for the vector control (Table S2). Taken together, the fluctuation analysis indicated that the dominant pms1 mutations were null PMS1 alleles that caused a copy-number dependent dominant mutator phenotype. This dominant mutator phenotype was stronger for mutations causing amino acid substitutions of metal ligands as compared to the pms1-G683E mutation.
Previous studies have shown that S. cerevisiae Mlh1-Pms1 has a metal-dependent endonuclease activity that can be stimulated by RFC and PCNA [26]. To determine if the dominant pms1 mutations affecting metal ligating amino acids disrupt the endonuclease function of Mlh1-Pms1, we expressed and purified the S. cerevisiae wild-type Mlh1-Pms1 complex and the mutant Mlh1-Pms1-G683E, Mlh1-Pms1-C817R, Mlh1-Pms1-C848S, and Mlh1-Pms1-H850R complexes and assayed the ability of these complexes to nick supercoiled pRS425 plasmid DNA with or without accessory factors RFC-Δ1N and PCNA (Figure 2A). Wild-type Mlh1-Pms1 alone showed little endonuclease activity. However, addition of PCNA and RFC-Δ1N to reactions containing wild-type Mlh1-Pms1 resulted in a 20-fold increase in endonuclease activity resulting in cleavage of nearly half of the original substrate DNA. The newly identified Mlh1-Pms1 mutant proteins and the previously studied Mlh1-Pms1-E707K mutant protein did not exhibit any PCNA and RFC-Δ1N stimulated endonuclease activity, with the exception of the Mlh1-Pms1-H850R mutant protein (Figure 2A). An explanation of the ability of the Mlh1-Pms1-H850R mutant protein to nick supercoiled DNA is provided in the “Discussion”. These results support the idea that loss of metal coordination by Pms1 inhibits the endonuclease activity of Mlh1-Pms1.
Exo1 is a 5′-3′ nuclease that functions in MMR in vitro by resecting DNA from a preexisting nick to a point past the mispair [28], [44], [49]. Loss of Exo1 function in vivo by deletion of EXO1 or missense mutations inactivating the Exo1 active site only results in a weak mutator phenotype and hence only partial loss of MMR (Table 1) [27]–[29], [50]. To determine the consequences of eliminating the two known nucleases involved in S. cerevisiae MMR, we tested the effect of introducing plasmids containing the dominant pms1 mutations into an exo1Δ mutant strain containing a wild-type PMS1 gene. Relative to the effects in a wild-type strain, introduction of the dominant pms1 mutations on a low-copy plasmid into the exo1Δ mutant strain increased the mutation rate to a much greater extent (Table 1). The pms1-G683E mutation again caused the weakest mutator phenotype, whereas the other five pms1 mutations that affect the metal ligands caused relatively high mutation rates ranging from a 57- to a 210-fold increased mutation rate depending on the mutation and the assay. These results show that the dominant pms1 mutations cause a greater MMR defect in an exo1Δ mutant strain than in a wild-type strain, suggesting that Exo1 and the Mlh1-Pms1 nuclease have a redundant function in MMR.
The four C-terminal residues of Mlh1 are almost completely conserved as the amino acids FERC in fungal and animal species (Figure 1D). Furthermore, Mlh1 has no C-terminal extension beyond the FERC residues in almost all sequenced eukaryotic organisms except nematodes (FERCG[T/S]) whereas the ScPms1/HsPMS2 family of proteins have variable length C-terminal extensions (Figure 1D). Consistent with a special role for the C-terminus of Mlh1, C-terminal fusions to the S. cerevisiae Mlh1 are non-functional for MMR in vivo, whereas C-terminal fusions to S. cerevisiae Pms1 do not affect function [38]. A structure of this highly conserved Mlh1 C-terminus (Figure 1B,C) revealed that this region might be appropriately positioned to play a role at the endonuclease active site of Pms1, with the C-terminal Mlh1 cysteine potentially acting as a metal ligand. We therefore generated the mlh1-E767stp, mlh1-C769stp, mlh1-C769A and mlh1-C769S mutations to probe the role of these highly conserved residues in endonuclease activity in vitro and MMR in vivo.
Unlike the pms1 mutations affecting metal ligands, none of the mutations affecting the C of the conserved FERC motif of Mlh1 that is predicted to be a metal ligand including the mlh1-C769A, mlh1-C769S, and mlh1-C769stp mutations caused a dominant mutator phenotype when present on an ARS CEN plasmid in a wild-type strain or an exo1Δ strain (Table 2). The mlh1-C769A, mlh1-C769S, and mlh1-C769stp mutants fully complimented the MMR defect of an mlh1Δ strain (Table S2) consistent with previously published results for the mlh1-C769A mutation but not the mlh1-C769S mutation [51] or the mlh1-C769stp mutation [46]. In this regard, it should be noted that our studies used a broader series of mutator assays, including more sensitive assays, than previous studies of mutations affecting Mlh1-C767. In contrast, the mlh1-E767stp mutant plasmid failed to complement the MMR defect of an mlh1Δ strain and resulted in a null phenotype (Table S2). Furthermore, the mlh1-E767stp mutation on an ARS CEN plasmid caused a weak dominant mutator phenotype when present in a wild-type strain and a stronger dominant mutator phenotype when present in an exo1Δ strain (Table 2), although not to the extent as that caused by the PMS1 metal ligand mutations.
We also tested the effect of the mlh1-C769stp and mlh1-E767stp mutations in the endonuclease assay (Figure 2B). The mutant Mlh1-Pms1 protein lacking only the Mlh1 C-terminal cysteine that did not cause an MMR defect in vivo nicked supercoiled DNA to the same extent as the wild-type Mlh1-Pms1 protein. In contrast, the mutant Mlh1-Pms1 protein resulting from the mlh1-E767stp mutation was significantly defective for nicking supercoiled plasmid DNA, which parallels the effect of this mutation on MMR in vivo. These results support the idea that the C-terminus of Mlh1 functions in the endonuclease active site although if the terminal cysteine coordinates bound metal then this role is not required for MMR or endonuclease activity (Figure 1C).
Mutations affecting the active sites of leading and lagging strand DNA polymerases Pol ε, pol2-M644G, and Pol δ, pol3-L612M, have been identified that preferentially introduce misincorporation errors in their respective strand during DNA replication [52], [53]. In a wild-type background, these lesions are then efficiently corrected by MMR. This strand-biased misincorporation can be used to determine strand preferences for MMR [52], [54], [55]. Here we probed mutants containing these polymerase active site mutations with ARS CEN plasmids encoding endonuclease defective pms1 and mlh1 mutations to investigate whether the Mlh1-Pms1 endonuclease preferentially functions in leading or lagging strand MMR. We found that ARS CEN PMS1 plasmids containing the pms1-C817R, pms1-C848S or pms1-H850R mutations caused a statistically similar synergistic increase in mutation rate when present in strains containing mutations affecting either DNA polymerase (Table 3). Of nine pairwise comparisons between pol2-M644G and pol3-L612M mutants containing the same pms1 mutation on a low copy plasmid using three different mutation rate assays, seven were not different (p-value>0.05, Mann-Whitney test). For the two comparisons that showed a difference, one showed a modestly higher rate in the pol2-M644G strain while the other showed a modestly higher rate in the pol3-L612M strain (p-value<0.05, Mann-Whitney test). This degree of similarity between the pol2-M644G and pol3-L612M mutants in these comparisons is in marked contrast to the effect of an exo1Δ mutation that caused a 9-fold higher increase in mutation rate in a pol3-L612M mutant compared to a pol2-M644G mutant [38]. Overall, these results suggest that Mlh1-Pms1 functions similarly on both the leading and lagging strands during MMR.
We have previously demonstrated that Mlh1-Pms1 foci are an intermediate in MMR and that blocking MMR downstream of Mlh1-Pms1 recruitment results in increased levels of Mlh1-Pms1 foci [38]. To test the effect of the Mlh1-Pms1 active site mutations on the levels of Pms1 foci, different mlh1 and pms1 mutations were introduced into the relevant endogenous locus in a strain in which the single wild-type copy of PMS1 was functionally tagged with four tandem copies of GFP. Normally, Mlh1-Pms1-4GFP foci are present in approximately 10% of logarithmically growing wild-type cells whereas all of the mutations that affected endonuclease function in vitro that were tested caused increased Mlh1-Pms1-4GFP foci formation (Figure 3). These included the metal coordination pms1 mutations pms1-E707K, which was previously tested [38], as well as pms1-C817R, pms1-C848S and pms1-H850R, the pms1-G683E mutation and the mlh1-E767stp endonuclease active site mutation, which increased the proportion of cells containing Mlh1-Pms1-4GFP foci to between 64 and 96%. In contrast, the endonuclease and MMR proficient mlh1-C769stp mutation did not alter the levels of Mlh1-Pms1-4GFP foci compared to wild-type strain. These results are consistent with the idea that mispairs are recognized normally in these Mlh1-Pms1 endonuclease active site mutants and that there is proper loading of the mutant Mlh1-Pms1 on DNA but instead there is a mispair processing defect resulting in decreased turnover of the mutant Mlh1-Pms1 from the DNA [38].
In this study, we used the highly sensitive lys2-10A frameshift mutation reversion assay to screen mutagenized low copy PMS1 plasmids for pms1 mutations that caused a dominant mutator phenotype in the presence of a single-copy of the wild-type PMS1 gene. We identified four null mutations that caused a weak dominant phenotype under these conditions. These mutations all caused amino acid substitutions in or near the region predicted to contain the Pms1 endonuclease active site and three of the amino acid substitutions including Pms1-C817R, Pms1-C848S, and Pms1-H850R affected predicted metal binding motifs while Pms1-G683 was in close proximity to the metal coordination site. The weak effect of these mutations explains why a prior study of a predicted Pms1 active site mutation did not observe a dominant effect [56]. A model of the endonuclease active site predicted Pms1 amino acids H703 and E707 as well as Mlh1-C769 to also be a part of the Mlh1-Pms1 active site [46]. Consistent with this model, the pms1-H703A and pms1-H707A mutations were found to cause the same phenotypes as the other PMS1 metal ligand mutations. In contrast, no mutation affecting Mlh1-C769 caused either a dominant mutator phenotype or affected Mlh1 function whereas the mlh1-E767stp mutation caused phenotypes that were similar to those caused by the PMS1 metal ligand mutations. Remarkably, all of the pms1 mutations caused a stronger dominant mutator phenotype when present in an exo1Δ strain on a low copy plasmid. The mlh1-E767stp also caused an increased dominant mutator affect in the exo1Δ strain, although not to the extent seen with the pms1 mutations. The phenotype of these mutants is similar to previously described separation-of-function mutations in MSH2, MSH3, MSH6, MLH1, PMS1, POL30 and POL32 that cause strong defects in Exo1-independent MMR but little if any defect in MMR when Exo1 is functional [27].
The asymmetry of the endonuclease active site is consistent with proposed roles of Mlh1-Pms1 in nicking double-stranded DNA during MMR; however, this asymmetry is not present in homodimeric bacterial MutL homologs with endonuclease function [47], [48]. The similarity of the eukaryotic Mlh1-Pms1 and bacterial MutL homologs suggests that MutL-DNA complexes may be functionally asymmetric so that only one active site is positioned to cleave DNA, analogous to the functional asymmetry during mispair recognition by bacterial MutS homodimers [15], [16]. The asymmetry of the eukaryotic MutL complexes, however, allows specialization of each subunit. The highly conserved C-terminus of Mlh1 is positioned in the Mlh1-Pms1 structure in a way that suggests that the Mlh1-ScPms1/HsPMS2 and Mlh1-Mlh3 complexes have a composite endonuclease active site. Potential roles for residues in the Mlh1 C-terminus include coordinating DNA phosphates, promoting nucleophilic attack by a water molecule or stabilization of the Pms1 active site. Consistent with this, the mlh1-E767stp mutant plasmid did not complement the mutator phenotype caused by a deletion of MLH1, and the mlh1-E767stp mutation resulted in the accumulation of Mlh1-Pms1-4GFP foci and reduced endonuclease activity similar to mutations in PMS1 affecting the endonuclease active site. It was surprising that mutation or deletion of the highly conserved C-terminal cysteine did not cause a MMR defect in vivo or reduce endonuclease activity in vitro given the possibility that this residue might coordinate with metals in the endonuclease active site. It is possible that conservation of the C-terminal cysteine may reflect other roles for Mlh1, potentially including crossover resolution during meiosis [57]–[59].
Analysis of the genetically identified and structure-based mutations in the MLH1 and PMS1 genes revealed that disruption of the metal binding sites leads to disruption of the Mlh1-Pms1 endonuclease activity and arrest of MMR repair at a step following recruitment of Mlh1-Pms1 into microscopically-observable foci. In a Mn2+-, RFC-, and PCNA-dependent endonuclease assay, amino acid substitutions affecting all of the metal ligands caused defects in the Mlh1-Pms1 endonuclease activity, which confirms and extends previous studies of the human Mlh1-Pms2 E705K amino acid substitution [25], [26]. The pms1-H850R and the mlh1-E767stp mutations resulted in proteins with partial defects in the in vitro endonuclease assay, but caused complete MMR defects in vivo. The partial endonuclease defect caused by these mutations may reflect the fact that the in vivo metal ion in eukaryotic and bacterial homologs is Zn2+, which has tetrahedral coordination geometry, whereas the metal that promotes the in vitro assay is Mn2+, which prefers an octahedral coordination geometry [25], [48], [60]–[63]. All of the pms1 mutations that affect predicted metal binding ligands, as well as mlh1-E767stp, caused complete MMR defects in vivo, consistent with the observation that metal ligand defects in human Mlh1-Pms2 inactivate MMR in vitro [26], [63]. These mutations also caused an accumulation of Mlh1-Pms1-4GFP foci indicating the step at which these mutations disrupt MMR is after loading of Mlh1-Pms1 by Msh2-Msh6, suggesting that loss of endonuclease activity leads to a turnover defect of Mlh1-Pms1 during MMR.
A striking property of the dominant pms1 and mlh1 mutations is that when present on a low copy plasmid they cause a much greater defect in Exo1-independent MMR compared to MMR when Exo1 is functional even though Mlh1-Pms1 appears to be absolutely required for all MMR. This phenotype is similar to the phenotype caused by the previously described separation-of-function mutations in genes like MSH2, MSH3, MSH6, MLH1, PMS1, POL30 and POL32 that result in strong defects in Exo1-independent MMR but little if any defect in MMR when Exo1 is functional [27]. A hypothesis that could explain these observations is that Mlh1-Pms1 has two roles in MMR, one involving activation of the Mlh1-Pms1 endonuclease and one where Mlh1-Pms1 plays a role in the recruitment of downstream MMR factors. It was previously shown that Exo1 interacts with Mlh1 and that this interaction is required for Exo1 to function in MMR [50]. This suggests the possibility that mispair recognition by Msh2-Msh6 or Msh2-Msh3 recruits Mlh1-Pms1 which then targets Exo1 to DNA where it could promote excision at pre-existing nicks in the DNA, consistent with the observation that lagging strand MMR is more Exo1 dependent than leading strand MMR [38], [42]. Such a reaction would be expected to be relatively insensitive to inhibition by competition with an endonuclease inactive but structurally normal form of Mlh1-Pms1 that would still bind Exo1 and target it to the site of MMR. In contrast, MMR in the absence of Exo1 might be completely dependent on the Mlh1-Pms1 endonuclease activity. This reaction would be expected to be competed for and interfered with by the presence of endonuclease inactive but structurally normal form of Mlh1-Pms1. A model that summarizes these concepts is presented in Figure 4.
S. cerevisiae cells were grown in YEPD (1% yeast extract, 2% Bacto peptone and 2% dextrose with or without 2% Bacto agar) or SD (0.67% yeast nitrogen base and 2% dextrose with or without 2% Bacto agar) medium. SD medium was supplemented with the appropriate dropout mix of amino acids (USA Biological). The S. cerevisiae strains used in genetic experiments were derived from an S288c parental strain and the strains used for protein purification were derived from RDKY1293 or RDKY8053 (listed in Table S3). All strains were constructed using standard gene disruption and transformation procedures. E. coli strains were propagated in LB media (0.5% yeast extract, 1% tryptone, 0.5% NaCl, 50 µg/ml thymine with or without 2% Bacto agar) containing 100 µg/ml ampicillin as required.
All plasmids (listed in Table S4) were maintained in E. coli TOP 10F′. A pRS316 Ampr URA3 ARS-CEN PMS1 plasmid pRDK1667 was constructed by recombination in vivo. Briefly, PMS1 was amplified from S. cerevisiae S288c chromosomal DNA using the primers 5′ACGACGGCCAGTGAATTGTAATACGACTCACTATAGGGCGAATTGGAGCTattgccaaacaggcaaagac that contains 50 bp of homology to pRS316 upstream of the multiple cloning site followed by 20 bp of homology to the PMS1 genes 707 bp upstream of the promoter starting at chromosome XIV coordinate 472684 and 5′TTAACCCTCACTAAAGGGAACAAAAGCTGGGTACCGGGCCCCCCCTCGAGgcatacaagaaacaacgcga that contains 50 bp of homology to pRS316 encompassing the XhoI, DraII, ApaI and KpnI sites of the multiple cloning site followed by 20 bp at the 3′ end with homology to the PMS1 genes 302 bp downstream of the stop codon from chromosome XIV coordinate 476314. The PCR product was mixed with an equimolar amount of pRS316 that had been linearized by digestion with SmaI and co-transformed into the wild-type S. cerevisiae strain RDKY3590. The transformants were selected on SC-uracil drop out plates, DNA was isolated from individual transformants, rescued by transformation into E. coli and sequenced. The plasmid selected for further use has a silent C3055T mutation. A pRS426 Ampr URA3 2-micron PMS1 plasmid, pRDK1689, was constructed by subcloning the XhoI to StuI (StuI cuts in the URA3 gene) PMS1 fragment from pRDK1667 into the XhoI to StuI backbone of pRS426. The pRS316 Ampr URA3 ARS-CEN MLH1 plasmid pRDK1338 was from our laboratory collection and contains a SacI to XhoI MLH1 fragment inserted between the SacI and XhoI sites of pRS316. The MLH1 fragment starts at the native SacI site 2281 bp upstream of the MLH1 ATG and ends at an XhoI site inserted by PCR 121 bp downstream of the MLH1 stop codon. The 2-micron Mlh1 and Pms1 over-expression plasmids pRDK573 TRP1 GAL10-MLH1 and pRDK1099 LEU2 GAL10-PMS1-FLAG have been described previously [64]. Mutations were made in these plasmids using standard site-directed mutagenesis methods or by subcloning from a mutant plasmid and the resulting plasmids were verified by DNA sequencing. The PMS1 mutant alleles E707K, H850R, C848S, G683E and C817R were introduced at the chromosomal locus using standard pop in/out techniques employing the integrative plasmids listed in Table S4. These integration plasmids were generated by subcloning the XhoI-StuI fragment containing the pms1 mutant sequence from their respective pRS316-pms1 mutant series plasmids into the XhoI-StuI sites of the pRS306 backbone and were linearized with BlpI prior to transformation for integration into the strains of interest. The MLH1 mutant alleles E767stp and C769stp were introduced at the chromosomal locus using standard gene disruption employing an HPH disruption cassette generated by PCR such the upstream homology targeted the C-terminus of MLH1 and contained the mutations needed to introduce the E767stp and C769stp alleles. All of the chromosomal pms1 and mlh1 mutations were verified by sequencing the entire PMS1 or MLH1 gene as relevant, which also ensured that no additional mutations were introduced during strain construction.
Mutagenesis of the PMS1 gene by PCR was performed essentially as previously described [65] with the following modifications. The primers used for PCR were those described above for amplification of PMS1. Ten PCR reactions were performed using Klentaq DNA polymerase and a PMS1 gene containing plasmid pRDK433 from our laboratory collection as a template. The PCR reactions were pooled, aliquots of DNA were mixed with an equimolar amount of pRS316 that had been linearized by digestion with SmaI and co-transformed into the wild-type S. cerevisiae strain RDKY3590. The transformants were plated on SC-uracil drop out plates to select for transformants, which were then replica plated onto SC-uracil-lysine drop out plates to screen for colonies that had increased rates of reversion of the lys2-A10 frameshift mutation. Candidate mutator mutants were retrieved from the uracil drop out plates, restreaked on SC-uracil drop out plates, patched in duplicate onto uracil drop out plates and replica plated onto threonine-uracil drop out plates to screen for patches that had increased rates of reversion of the hom3-10 frameshift mutation. Plasmid DNA was isolated from each mutator mutant, transformed into E. coli TOP10F′ and sequenced. Individual mutations identified were then transferred to a new pRS316 PMS1 plasmid pRDK1667 by either sub-cloning using appropriate restriction endonuclease cleavage sites or by site-directed mutagenesis and retested essentially as described for the initial screen above.
S. cerevisiae Mlh1-Pms1 was purified from 2.2 L of culture of the overproduction strain RDKY7608 (RDKY1293 containing the 2-micron plasmids pRDK573 TRP1 GAL10-MLH1 and pRDK1099 LEU2 GAL10pr-PMS1-FLAG) (Table S3) according to a previously published procedure [64], except with the following 6 modifications: (1) Cell growth and induction of Mlh1-Pms1 expression utilized a published lactate to galactose shift protocol according to previously published methods [66]; (2) 2 mM β-mercaptoethanol was substituted for the 1 mM DTT in the buffers used to run the Heparin and FLAG antibody columns whereas all other buffers contained 1 mM DTT; (3) After washing the Heparin column with Buffer A containing 200 mM NaCl, the proteins were eluted using a single step of 1 M NaCl in Buffer A; (4) The pooled Heparin column fractions were diluted with Buffer A to obtain a final NaCl concentration of 500 mM prior to being subjected to 3 cycles of binding and elution from the FLAG antibody column; (5) The SP Sepharose column fractions were diluted with Buffer A to a final NaCl concentration of 200 mM prior to being loaded onto a 1 ml HiTrap Q column (GE Healthcare) followed by elution with a 100 mM to 1 M linear NaCl gradient run in Buffer A; and (6) The HiTrap Q column fractions containing the Mlh1-Pms1 were concentrated and desalted using a Centraprep (Ultracel 30K) spin column. The resulting Mlh1-Pms1 was contained in 0.5 ml of Buffer A +100 mM NaCl, and was frozen in liquid nitrogen and stored at −80 C. The Mlh1-Pms1-E707K, Mlh1-Pms1-C817R, Mlh1-Pms1-C848S and Mlh1-Pms1-H850R proteins were purified using the overproduction strains RDKY7696, RDKY7756, RDKY7759 and RDKY7793 (Table S3). The Mlh1-C769stp-Pms1 and Mlh1-E767stp-Pms1 proteins were purified using the overproduction strains RDKY8055 and RDKY8057 for which the RDKY8053 host strain (Table S3) was a derivative of RDKY1293 containing a deletion of the MLH1 gene. Because mlh1-E767stp allele does not compliment the MLH1 deletion in the host, after the protein expression period, DNA was isolated from the culture, 20 independent MLH1 and PMS1 plasmids were rescued by transformation into E. coli and sequenced to ensure that no mutations had occurred in the expression plasmids. S. cerevisiae PCNA and RFC-Δ1N were purified exactly as described in published procedures [66]–[68]. All of the protein preparations used in these studies were greater than 98% pure as analyzed by SDS-PAGE.
Mismatch-independent endonuclease assays were performed as a modification of one used previously [26]. 40 µL reactions containing 1 mM MnSO4, 20 mM Tris pH 7.5, 0.5 mM ATP 0.2 mg/mL bovine serum albumin (BSA), 2 mM DTT and 100 ng pRS425 were incubated at 30°C for 30 minutes. Reactions were terminated by incubation at 55°C following introduction of SDS, EDTA, glycerol and proteinase K at concentrations of 0.1%, 14 mM, 8% and 0.5 ug/ml respectively. Mlh1-Pms1, PCNA, or RFC-Δ1N were diluted to the appropriate working concentrations with a buffer comprised of 10% glycerol, 200 mM NaCl, 2 mM DTT and 20 mM Tris pH 7.5. Following termination of the reaction the samples were electrophoresed on a 0.8% agarose gel, the gel was stained with ethidium bromide, extensively destained and then the bands were quantified using a BioRad ChemiDoc XP imaging system. Serial dilutions of XhoI linearized pRS425 ranging from 10–100 ng were used as a concentration standard for quantification.
Mutation rates were determined by fluctuation analysis. A single colony was used to inoculate a culture that was then diluted and used for transformation with a selectable plasmid carrying the desired allele and transformed colonies were selected by growth for 3 days at 30°C on SC-uracil dropout plates. 7 independent colonies were used to inoculate individual overnight cultures containing 10 ml of SC-uracil dropout media. Following cell growth, appropriate dilutions of the cultures were plated onto SC –uracil, –uracil-lysine, -uracil-threonine, and -uracil-arginine+canavanine dropout plates. The resulting colonies counted after growth at 30°C for 3 days and the average mutation rate was calculated for each strain as described previously [21], [27]. Each experiment was performed independently up to 4 times.
Site-directed mutagenesis was guided by a molecular model of the C-terminal domains of Mlh1-Pms1. The C-terminus of S. cerevisiae Mlh1 was modeled using Phyre, xfit, and CNS [69]–[71] starting from the crystal structure of the human Mlh1 C-terminal domain (PDB id 3rbn). The C-terminus of S. cerevisiae Pms1 was similarly modeled using the crystal structures of N. gonorrhoeae [PDB id 3ncv; [47]] and B. subtilis [PDB ids 3gab, 3kdg, 3kdk; [48]] MutL homologs. Subsequently, the amino acid substitutions studied were mapped onto the newly available structure of the C-terminal domains of S. cerevisiae Mlh1-Pms1 [PDB id 4e4w; [46]].
For microscopy studies, the C-terminus of each PMS1 protein of interest was fluorescently tagged by targeting a 4GFP tag to the chromosomal locus so that the native promoter was intact and expression remained unaffected. Previous analysis of the tagged PMS1 gene demonstrated that the 4GFP tag did not affect the biological activity of Pms1 [38]. Exponentially growing cultures were washed and resuspended in water, placed on minimal media agar pads, covered with a coverslip, and imaged on a Deltavision (Applied Precision) microscope with an Olympus 100× 1.35NA objective. Fourteen 0.5 µm z sections were acquired and deconvolved with softWoRx software. Further image processing, including maximum intensity projections and intensity measurements were performed using ImageJ.
|
10.1371/journal.pgen.1007574 | Enhanced uptake of potassium or glycine betaine or export of cyclic-di-AMP restores osmoresistance in a high cyclic-di-AMP Lactococcus lactis mutant | The broadly conserved bacterial signalling molecule cyclic-di-adenosine monophosphate (c-di-AMP) controls osmoresistance via its regulation of potassium (K+) and compatible solute uptake. High levels of c-di-AMP resulting from inactivation of c-di-AMP phosphodiesterase activity leads to poor growth of bacteria under high osmotic conditions. To better understand how bacteria can adjust in response to excessive c-di-AMP levels and to identify signals that feed into the c-di-AMP network, we characterised genes identified in a screen for osmoresistant suppressor mutants of the high c-di-AMP Lactococcus ΔgdpP strain. Mutations were identified which increased the uptake of osmoprotectants, including gain-of-function mutations in a Kup family K+ importer (KupB) and inactivation of the glycine betaine transporter transcriptional repressor BusR. The KupB mutations increased the intracellular K+ level while BusR inactivation increased the glycine betaine level. In addition, BusR was found to directly bind c-di-AMP and repress expression of the glycine betaine transporter in response to elevated c-di-AMP. Interestingly, overactive KupB activity or loss of BusR triggered c-di-AMP accumulation, suggesting turgor pressure changes act as a signal for this second messenger. In another group of suppressors, overexpression of an operon encoding an EmrB family multidrug resistance protein allowed cells to lower their intracellular level of c-di-AMP through active export. Lastly evidence is provided that c-di-AMP levels in several bacteria are rapidly responsive to environmental osmolarity changes. Taken together, this work provides evidence for a model in which high c-di-AMP containing cells are dehydrated due to lower K+ and compatible solute levels and that this osmoregulation system is able to sense and respond to cellular water stress.
| Second messengers relay signals received from the environment to intracellular targets that adjust cellular physiology. One widespread bacterial cyclic-dinucleotide signalling molecule, cyclic-di-AMP (c-di-AMP) has been shown to regulate a range of cellular processes via binding to protein and riboswitch targets, with most identified thus far being linked to osmoregulation functions. C-di-AMP levels need to be carefully tuned under different environmental conditions to allow optimal growth. Here we show that a Lactococcus lactis GdpP phosphodiesterase mutant with a high intracellular pool of c-di-AMP is able to grow under hyperosmotic conditions after acquiring mutations which increase osmolyte (potassium [K+] or compatible solute) uptake or by actively exporting c-di-AMP. Interestingly, elevated K+ or glycine betaine uptake triggered accumulation of c-di-AMP and environmental osmolarity changes were also found to significantly impact c-di-AMP levels in various bacteria. These results support a model in which c-di-AMP negatively impacts osmoresistance through inhibition of the import of osmoprotectants and this system can sense both cellular and environmental changes causing water stress.
| In order to survive and grow, bacteria must be able to sense and respond to a multitude of environmental conditions. Changes in external osmolarity can cause cellular water loss or gain due to uncontrolled osmotic movement across the semipermeable cytoplasmic membrane. Cells can adapt to these changes in order to maintain appropriate cellular volume and solute concentration for metabolism as well as turgor pressure to drive expansion for cell division [1]. In response to an osmotic upshift (hyperosmotic stress), bacteria import potassium ions (K+) which is followed by a secondary response involving uptake or synthesis of compatible solutes such as glycine betaine, carnitine and proline [2]. This allows the cell to limit the loss of water and maintain turgor. During osmotic downshift (hypoosmotic stress), bacteria release K+ and compatible solutes from the cell in order to limit water influx causing cell swelling and in severe cases, lysis. The speed at which cells need to detect and respond to the external osmolarity change is critical and therefore the early responses in many cases involves posttranslational modulation of existing transporter activity, since the synthesis of new proteins can take too long [3]. These transporters include membrane stretch-activated mechanosensitive channels that activate during an osmotic downshift and intracellular ionic strength or K+ activated compatible solute uptake systems that function during an osmotic upshift [4].
The signalling molecule cyclic-di-AMP (c-di-AMP) found in many Gram-positive bacteria and some Gram-negative bacteria has been recently demonstrated to play a significant role in regulating K+ and compatible solute (carnitine) import systems either via direct binding to transporter protein complexes (Ktr and OpuCA), a regulatory sensor kinase (KdpD) or riboswitch upstream of the transporter genes (ktr and kimA) [5–9]. A high level of c-di-AMP has been found to repress K+ and carnitine uptake, thereby inhibiting growth under hyperosmotic conditions [5, 6, 10, 11]. Conversely, low/absent c-di-AMP results in cells which have likely uncontrolled K+ and compatible solute uptake, resulting in viability only under conditions where external K+ and compatible solute concentrations are low or when the cells are osmotically stabilised by the addition of high salt [9, 12].
Whilst the role of c-di-AMP in osmoregulation is becoming evident, the signals which trigger changes in the c-di-AMP pool are still poorly understood [13, 14]. C-di-AMP is synthesised from two ATP molecules by diadenylate cyclase (DAC) enzymes and degraded by phosphodiesterase (PDE) enzymes. In most Gram-positive bacteria, only one DAC exists, named CdaA or DacA, and it is localised to the membrane via three transmembrane domains [15]. Most bacteria also contain one PDE, called GdpP, while others contain an additional PDE called PgpH [16] and both of these PDEs are membrane localised [17]. It is through these enzymes, and possibly active c-di-AMP export, that the c-di-AMP pool is regulated. The intracellular level of c-di-AMP is under strict control since both low and high levels have been shown to be detrimental to growth in several Firmicutes [18, 19]. C-di-AMP levels in bacteria have been found to be responsive to growth phase, acid stress, growth media nitrogen source, (p)ppGpp induced by mupirocin, mutations in the peptidoglycan biosynthesis enzyme GlmM or YbbR protein and inactivation of the LiaFSR membrane stress response system [11, 20–23]. Upregulated CdaA expression and a 2-fold higher level of c-di-AMP was observed in Bacillus subtilis grown in defined media with 5mM K+ compared to cells grown with 0.1mM K+ [9]. Recently, it was found that inactivation of the gating component (CabP) of the Trk family K+ transporter in Streptococcus pneumoniae reduced the c-di-AMP level [24]. These results suggest that K+ may act as a signal for c-di-AMP level modulation directly or indirectly.
In this study, we sought to better understand the cause of high c-di-AMP toxicity and to identify mechanisms by which this toxicity can be averted in bacteria. We used a high c-di-AMP containing Lactococcus lactis ΔgdpP mutant which is hypersensitive to elevated salt in growth media. This strain can accrue suppressor mutations which allows it to grow under such an osmotically stressful condition and previous work has identified changes in the c-di-AMP synthase CdaA and binding partner GlmM which both lowered the c-di-AMP level [11]. Here we expanded the screen to saturation and identified and characterised three different suppressor mutant groups. Two groups rescue osmoresistance through mutations which elevate osmolyte (K+ or glycine betaine) uptake while one group activated the expression of a c-di-AMP export system involving a multidrug resistance protein (MDR) which lowered intracellular c-di-AMP. In addition we show that c-di-AMP accumulates in response to elevated K+ and glycine betaine uptake, and in several bacteria the c-di-AMP level is rapidly responsive to environmental osmolarity changes, thus allowing it to sense and respond to water stress.
To identify regulators of the c-di-AMP pool, we previously employed a genetic screen to obtain osmoresistant suppressors of the high c-di-AMP Lc. lactis ΔgdpP strain OS2 [11]. In the current study, a further 212 osmoresistant suppressor mutants were obtained and analysed. Numerous cdaA mutations were identified (n = 184) as well as several restorative gdpP mutations (n = 8). Twenty suppressors which contained identical cdaA and gdpP sequences as the parent strain were subjected to whole genome sequencing and the mutation locations are shown in Table 1. The 20 suppressors contained mutations in 3 functionally linked gene groups. The first group included 4 independent mutants which possessed changes in the K+ uptake transporter KupB (Llmg0588). The second group included 11 independent mutants which contained changes in the Eep/PptAB hydrophobic peptide processing and export system (S1A Fig). Eep (Llmg2413) is a transmembrane metalloprotease homolog of RseP [25] and PptAB (Llmg2271-2270) are ATPase and permease homologs of EcsAB [26]. Interestingly one suppressor contained a mutation in pptB and a mutation in busR which encodes a transcriptional repressor of the glycine betaine transporter [27]. The third group included 2 independent mutants which contain deletions in an intergenic region between the ribosomal 50S protein encoding rplL and the MarR family transcriptional regulator encoding rmaX.
Osmoresistance of representative strains from these groups as well as a cdaA mutant are shown in Fig 1A. As found in previous work [11], mutations lowering the c-di-AMP level restored osmoresistance in ΔgdpP and this was found to be the case for the eep, pptB, rpilLΔterm209 and cdaA mutants (Fig 1B). A pptB gene disruption was constructed in the ΔgdpP background strain which, as expected, resulted in restoration of osmoresistance (S1B Fig) and a lowering of c-di-AMP (S1C Fig). Strikingly, the kupB mutants were found to have significantly elevated c-di-AMP or at least equivalent levels to the ΔgdpP parent (Fig 1B). This is the only group of osmoresistant suppressors we have identified thus far which do not have lower c-di-AMP levels relative to the ΔgdpP parent. Therefore, they have developed osmoresistance independently of c-di-AMP pool modulation. Importantly suppressors which contain only a single mutation in kupB (kupBA618V and kupBR508G) had significantly elevated c-di-AMP (Fig 1B), suggesting that changes in K+ uptake triggers c-di-AMP accumulation. In the two other kupB suppressors, one additional mutation is present in each (ftsXP249L or pptAS55fs) which likely caused a lowering of the c-di-AMP level (Fig 1B).
KupB contains twelve transmembrane spanning domains and a C-terminal 227 amino acid intracellular domain (Fig 1C). Among the 4 identified KupB missense mutations in the suppressor mutants, one is present in an internal loop region between membrane-spanning domains while three are in the C-terminal intracellular domain. KupB is homologous to a large family of Kup proteins present in several Gram-positive bacteria, Gram-negative bacteria and plants (Fig 1D) where their roles in K+ uptake and osmoregulation is well established [28]. Kup proteins possess highly similar transmembrane domains, but differ in their C-terminal intracellular domain (Fig 1D) which has been proposed to regulate K+ uptake [29]. Mutations identified in our suppressor screen restore osmoresistance, suggesting that they are gain of function mutations and might result in increased K+ uptake most likely through modification of channel regulation.
To determine if the KupB mutations resulted in a gain-of-function activity we overexpressed KupBA618V and wild-type KupB using its native promoter in ΔgdpP and measured their effect on osmoresistance. It was found that overexpression of wild-type KupB increased the osmoresistance of ΔgdpP, suggesting that an increase in kupB copy number can increase K+ uptake (Fig 2A). Overexpression of KupBA618V in ΔgdpP restored osmoresistance to a higher level than both ΔgdpP overexpressing wild-type KupB and the ΔgdpPkupBA618V suppressor mutant (Fig 2A). We next determined the intracellular K+ levels and found that ΔgdpP has lower K+ compared to both the wild-type and the ΔgdpPkupBA618V suppressor mutant (Fig 2B). Overexpression of KupBA618V in ΔgdpP resulted in 43% higher K+ level than ΔgdpP (Fig 2B). Together these results confirm that the A618V mutation in KupB is a gain-of-function mutation which increases osmoresistance by increasing K+ uptake.
One possibility was that these mutations prevent c-di-AMP from binding KupB and inhibiting K+ uptake, since it has been shown to bind to and/or regulate K+ transporters in other bacteria such as Ktr, Kdp and KimA. To determine if c-di-AMP binds to KupB we carried out the differential radial capillary action of ligand assay (DRaCALA) using the KupB intracellular C-terminal as fusions to hexa-His or MBP tags. This domain could be expressed but both fusion proteins were insoluble and testing of whole cell lysates with radiolabelled c-di-AMP did not reveal any binding (S2 Fig). Previous work has shown that insoluble proteins in whole cell lysates can bind cyclic dinucleotides [8, 30]. Therefore at this stage the regulation of KupB by c-di-AMP or other signals is not known. Interestingly using quantitative reverse transcriptase PCR (qRT-PCR), it was found that transcription of kupB was slightly higher (3-fold; P = 0.02; Student’s t test) in ΔgdpP compared to wild-type. This may be in response to low intracellular K+ levels present in the high c-di-AMP mutant ΔgdpP.
The c-di-AMP levels in L. lactis ΔgdpP strains overexpressing wild-type KupB or KupBA618V were determined next. Overexpression of wild-type KupB triggered higher c-di-AMP in ΔgdpP, while the c-di-AMP level in ΔgdpP overexpressing KupBA618V was even higher (Fig 2C). The level of c-di-AMP in these ΔgdpP background strains directly correlated with their osmoresistance level (Fig 2A), with the highest c-di-AMP strains being the most osmoresistant. This result is in stark contrast to previous work, where high c-di-AMP ΔgdpP mutants were found to be osmosensitive due to reduced activity or expression of K+ and/or compatible solute transporters [5, 7, 8, 11]. In the situation here, it appears that increased intracellular K+ due to enhanced K+ import activity not only results in the restoration of osmoresistance in ΔgdpP, but also that intracellular K+ serves either directly or indirectly as a signal for the cell to increase the c-di-AMP level. Indeed the level of K+ in ΔgdpP, ΔgdpPkupBA618V and ΔgdpP-pGh-kupBA618V strains directly correlated with the level of c-di-AMP (Fig 2B and 2C). An increase in the c-di-AMP level can occur via either increased c-di-AMP synthesis and/or decreased c-di-AMP hydrolysis. However it appears that control is mediated by increased c-di-AMP synthesis by CdaA here, as gdpP is defective in these strains. Using qRT-PCR, transcription of cdaA in ΔgdpPkupBA618V was unchanged compared with the parent ΔgdpP (P = 0.96; Student’s t test), which suggests that increased c-di-AMP levels are not due to higher cdaA expression.
Next we examined if elevated K+ uptake could trigger greater c-di-AMP accumulation in strains other than ΔgdpP. We overexpressed KupB and KupBA618V in wild-type Lc. lactis and related Lactobacillus reuteri which both contain GdpP. In Lc. lactis, the level of c-di-AMP was 7.5-fold higher when KupBA618V was overexpressed compared to cells overexpressing KupB or the wild-type (Fig 2D). In wild-type Lb. reuteri, overexpression of KupB and KupBA618V resulted in 1.9-fold and 6.7 fold c-di-AMP level increases, respectively (Fig 2D). Therefore the c-di-AMP pool size correlates directly with K+ uptake activity in at least two different bacterial genera irrespective of the presence of GdpP.
One suppressor mutant contained an in-frame deletion in the busR gene and an inactivating mutation in the pptB gene (S189fs) (Table 1). This busR deletion removed 42 residues (amino acids 42–83) from the N-terminal GntR HTH domain (Fig 3A). BusR is a transcriptional repressor of the BusAA-AB glycine betaine transporter in Lc. lactis [27]. To determine if the 42 amino acid deletion reduces the ability of BusR to repress the busAA promoter, we expressed the wild-type and mutant BusR in E. coli in conjunction with a lacZ fusion to the promoter of busAA. It was found that the busAA promoter activity increased moderately upon NaCl addition, similarly to that described before [31] (S3A Fig). The wild-type BusR repressed expression of busAA more strongly than the mutated BusR under all conditions with or without additional NaCl (S3A Fig) suggesting that the 42 amino acid deletion is a loss-of-function mutation. BusR is a member of the family of GntR regulators and homologs are present in a subset of Firmicutes including Clostridium spp. and certain lactic acid bacteria (S3B Fig). These proteins contain a C-terminal TrkA_C domain (Pfam02080) which is also found in gating components of the Trk family of K+ transporters that bind c-di-AMP [6, 7]. To determine if BusR binds c-di-AMP we expressed both full length and the TrkA_C domain in E. coli and carried out DRaCALA on whole cell lysates. Both full length BusR and the BusR TrkA_C domain bound c-di-AMP (Fig 3B) and the purified TrkA_C domain was found to interact with c-di-AMP with a Kd of ~10μM (Fig 3C). We determined the expression of busAA-AB in strains with varying c-di-AMP levels using a plasmid containing the busAA-AB promoter fused to a lacZ reporter. It was found that expression of busAA-AB was absent in the high c-di-AMP strain ΔgdpP compared to wild-type (Fig 3D). All osmoresistant suppressors were found to have greater busAA-AB expression as compared to ΔgdpP (Fig 3D). Suppressors which have lower c-di-AMP including those with mutations in cdaA, rpilLΔterm85 and glmM had restored busAA-AB expression to varying levels (Fig 3D). The suppressor mutant ΔgdpP pptB189fs busRΔ126 containing an inactive BusR showed very high busAA-AB expression, as expected (Fig 3D). Glycine betaine levels in different strains were measured and it was found that ΔgdpP contained ~10-fold less than wild-type (Fig 3E). Inactivation of busR in the wild-type or ΔgdpP background resulted in elevated glycine betaine levels compared to their parent strains (Fig 3E). Glycine betaine levels in osmoresistant suppressor mutants of ΔgdpP (cdaAD123Y, rpilLΔterm85, pptB189fsbusRΔ126) were all significantly higher (14- to 37-fold) compared with ΔgdpP (Fig 3E). These results suggest that BusAA-AB expression and glycine betaine levels are reduced in strains with high c-di-AMP due to this signalling molecule binding to BusR and enhancing its repression.
Next we determined if strong repression of glycine betaine transporter expression by BusR contributes to the osmosensitive phenotype of ΔgdpP. Inactivation of BusR in ΔgdpP restored osmoresistance, indicating that glycine betaine transport allows the ΔgdpP to regain normal turgor pressure under high osmolarity conditions (Fig 3F). Lastly we investigated if increased glycine betaine uptake caused by loss of BusR triggers c-di-AMP accumulation. It was found that the level of c-di-AMP was indeed 3-fold higher in ΔbusR compared to WT (Fig 3G). Taken together these results provide evidence that uncontrolled glycine betaine uptake is sensed by the cell, which in turn responds by elevating the c-di-AMP level in order to enhance BusR mediated repression of busAA-AB.
Two independent osmoresistant mutants containing overlapping 209 and 85bp deletions in an intergenic region between rplL and rmaX were obtained (Figs 4A and S4A). This deleted region includes two inverted repeats with free energies of -8.1 and -11.7 kcal/mol (S4A and S4B Fig). Either both or one are likely to act as transcription terminators for the upstream rplJ-rplL operon. The two different deletion mutants showed equivalent salt resistance and the deletion events were verified by PCR using primers flanking the deletion (S4C and S4D Fig). The operon upstream of the deletion events contains two ribosomal genes rplJ (50S ribosomal protein L10) and rplL (50S ribosomal protein L7/L12) (Fig 4A). Downstream of the deletion is a three gene operon composed of rmaX (MarR transcriptional regulator), llmg1210 (MDR of the EmrB family) and llmg1211 (predicted membrane protein of unknown function). Due to the deletion of a putative terminator, we hypothesised that transcription of the downstream operon may now be increased due to extension of transcripts from the likely highly expressed ribosomal genes. Using qRT-PCR, it was found that RNA transcripts of rmaX, llmg1210 and llmg1211 were several hundred fold higher in the two intergenic deletion suppressor mutants relative to the parent ΔgdpP (Fig 4B). As expected, expression of the rplJ and llmg1212 were not elevated in the deletion suppressor mutants (Fig 4B).
Previous work has demonstrated that c-di-AMP is actively secreted by MDR proteins of the MFS superfamily in L. monocytogenes and B. subtilis [17, 32–36]. Llmg1210 was found to share 45% amino acid identity with MdrT (Lmo2588) and 36% identity with MdrM (Lmo1617) (S5 Fig), both of which have been shown to export c-di-AMP in L. monocytogenes [32]. We hypothesised that higher MDR expression leads to greater c-di-AMP export resulting in a lowering of the intracellular c-di-AMP. The 2 different deletion suppressor mutants contained reduced intracellular c-di-AMP levels compared to ΔgdpP, which were equivalent to that of eep and pptB suppressor mutants (Fig 4C). However extracellular c-di-AMP levels were found to be disproportionately high in the deletion mutants, between 7 and 45 fold higher compared to the eep and pptB suppressor mutants (Fig 4C). To determine which gene(s) within the overexpressed operon is required for c-di-AMP export and osmoresistance, we overexpressed each using the predicted strong rplJ promoter on a plasmid in the ΔgdpP strain. It was found that overexpression of rmaX or llmg1211 did not lower the intracellular c-di-AMP level or restore osmoresistance (Fig 4D and 4E). We were able to clone llmg1210 downstream of the rplJ promoter in E. coli as a host, however upon introduction in Lc. lactis ΔgdpP, mutations in llmg1210 occurred in several independent trials, suggesting that overexpression of this protein by itself is toxic. When we cloned llmg1210 combined with llmg1211 downstream of the rplJ promoter, the plasmid was stable in ΔgdpP and a reduction of intracellular c-di-AMP occurred along with rescue of osmoresistance (Fig 4D and 4E).
To further confirm the role of this overexpressed operon in c-di-AMP export, we generated a series of chromosomally integrated mutants throughout the operon in the osmoresistant intergenic deletion suppressor strain ΔgdpPrplLtermΔ85 (S6A Fig). In this strain, the rmaX-llmg1210-llmg1211 operon is highly expressed. It was found that inactivation of expression of the entire operon by plasmid insertion elevated intracellular c-di-AMP and eliminated osmoresistance (S6B and S6C Fig). Interestingly the c-di-AMP level was significantly higher than that in ΔgdpP suggesting that the rmaX operon may be expressed and function to export c-di-AMP in ΔgdpP without the deletion event. Insertion of a plasmid allowing overexpression of rmaX only also resulted in similarly high c-di-AMP level and osmosensitivity. Plasmid insertion allowing overexpression of rmaX-llmg1210 from the genome lowered the c-di-AMP level compared to the two strains with plasmid insertions upstream (p < 0.001), however this was not sufficiently low enough to restore osmoresistance (S6B and S6C Fig). Only the plasmid insertion following llmg1211, which allows expression of all three genes, lowered the intracellular c-di-AMP level enough to restore osmoresistance in this deletion suppressor mutant (S6B and S6C Fig). Together these results suggest that the MDR llmg1210 is a c-di-AMP export protein, but requires llmg1211 for full activity and/or stability.
From the findings above and from other work, it is clear that c-di-AMP is a major regulator of osmoresistance. Therefore the c-di-AMP pool size would be predicted to be responsive to and inversely proportional to the external osmolarity allowing appropriate regulation of K+ and compatible solute transporters to control cell turgor. Therefore cells experiencing high turgor in low osmolarity environments would elevate their c-di-AMP pool, while cells experiencing low turgor in high osmolarity environments would deplete their c-di-AMP pool. Difficulties exist when trying to test this hypothesis however, since wild-type bacteria generally have low levels of c-di-AMP and changes in environmental osmolarity imposed during growth will likely lead to other physiological and gene expression changes which may have indirect influences on the c-di-AMP pool size. In order to separate direct and indirect effects, a simple new method was developed whereby the c-di-AMP level could be monitored in non-growing cells in low and high osmolarity conditions (Fig 5A). In an initial attempt to stimulate c-di-AMP accumulation, we suspended washed Lc. lactis WT cells in a low osmolarity buffer to stimulate increased turgor pressure. However, the c-di-AMP level remained the same as that found in cells grown in culture media and extracted with acetonitrile-methanol (Fig 5B). We hypothesised that washed cells have depleted ATP and therefore no immediate precursor for c-di-AMP synthesis. This is the case for bacterial ATP binding cassette (ABC) solute uptake systems which need to be energised by glucose addition to cells during uptake assays [37, 38]. It was found that energizing the cells by the addition of glucose resulted in a ~10-fold increase in c-di-AMP (Fig 5B). The addition of the closely related analog deoxyglucose, which is unable to initiate glycolysis, did not trigger c-di-AMP accumulation (Fig 5B). Several Gram-positive bacteria with a single DAC domain protein were analysed using this assay and all rapidly increased their c-di-AMP pools under these low osmolarity conditions which would trigger high turgor pressure (Fig 5C–5F).
We next examined the effect of an osmotic increase on c-di-AMP levels in several bacteria. These conditions would trigger lower turgor pressure. Ten minutes after glucose addition, either water, NaCl or KCl (0.1 M or 0.3 M final concentration) was added and cells were harvested after a further 10 minutes. The addition of NaCl or KCl resulted in rapid depletion or a block in c-di-AMP synthesis relative to water treated cells for all bacteria (Fig 5C–5F). We also tested the effect of non-ionic solutes (sorbitol and sucrose) on c-di-AMP levels in L. monocytogenes and found that they also stopped c-di-AMP synthesis or at higher levels triggered rapid c-di-AMP degradation (Fig 5G). These results demonstrate that environmental osmolarity changes trigger rapid c-di-AMP level fluctuation.
To determine the roles of CdaA and GdpP in c-di-AMP control in the energised cell assay, we examined an osmoresistant suppressor mutant of Lc. lactis ΔgdpP which has a partially defective cdaA (T273fs mutation). The c-di-AMP level did not vary over the 20 minute time course (Fig 5H). This demonstrates that CdaA and GdpP are the main controllers of c-di-AMP pool modulation in this assay.
The primary role of a second messenger is to transduce a signal(s) from the environment to effectors within the cell, resulting in a physiological response and ultimately adaptation. It is becoming apparent that a major role of c-di-AMP is in osmoregulation due to its control of osmolyte (K+ and compatible solute) transporters and observed osmosensitive and osmoresistant phenotypes of high and low c-di-AMP mutants, respectively. The results presented here reinforce this notion, with suppressor mutations which elevate K+ or glycine betaine transport rescuing osmoresistance in a high c-di-AMP mutant of Lc. lactis. Conversely in suppressor screens with mutants devoid of c-di-AMP, mutations inactivating osmolyte (peptides and glycine betaine) uptake [12, 39, 40] or increasing osmolyte (K+) export [9] have been found. Osmolyte import and export affects cellular turgor pressure and a recent proposition is that the central role of c-di-AMP is in the regulation of turgor pressure [41], which is supported by our work presented here.
The largest number of binding effectors under the control of c-di-AMP known at present are involved in K+ uptake [6, 7, 9]. KupB identified in this work has not been linked to c-di-AMP signalling pathways. It is a member of the Kup/HAK/KT family of K+ importer proteins (Pfam02705) which are widely distributed in bacteria, fungi and plants [28]. In bacteria, Kup homologs are most common in Proteobacteria (634 species), followed by Actinobacteria (91 species) in the Pfam database. Sixty-one species, mainly lactic acid bacteria, within the Firmicutes also contain Kup homologs. In E. coli, Kup is the major K+ importer under hyperosmotic conditions at low pH and likely functions as an H+—K+ symporter [42, 43]. Little is known regarding the regulation of Kup family proteins. Deletion of the intracellular the C-terminal domain of Kup in E. coli significantly reduced K+ transport activity [29] which suggests that this domain has a regulatory function. In Firmicutes, K+ transporters KtrAB, KtrCD, KdpABCD and KimA are regulated by c-di-AMP either via direct binding to the transporter or a two-component sensor kinase and/or through gene expression changes. In Lc. lactis MG1363, homologs of Ktr, Kdp and KimA are absent, and we did not find any evidence for c-di-AMP-mediated regulation of KupB. In lactoccoci, the kup gene has likely undergone a duplication event since immediately upstream there is a highly similar gene encoding KupA (>70% amino acid identity to KupB) in the majority of Lc. lactis strains. However, in strain MG1363 this is a pseudogene carrying a stop codon at amino acid 254. It was found that deletion of kupB did not affect the growth of Lc. lactis MG1363 in chemically defined media with lower K+ concentrations (S7 Fig), so it appears that there is another K+ transport system distinct from other characterised bacterial K+ transporters yet to be identified.
C-di-AMP is a significant regulator of compatible solute uptake through binding of the cystathionine-β-synthase (CBS) domain in the OpuCA carnitine transporters in L. monocytogenes and S. aureus [5, 8]. In our screen we identified a deletion event in busR which encodes the repressor of the glycine betaine transporter BusAA-AB [27]. It was found in previous work that busAA-AB expression was reduced in ΔgdpP [10], however the mechanism was not clear at that time. In this work, BusR was found to bind c-di-AMP and rescue osmoresistance in ΔgdpP, which has also been recently reported for Streptococcus agalactiae BusR [39]. It is likely that this is a conserved function in other Gram-positive bacteria that contain BusR orthologs with the same GntR family domain structure. C-di-AMP binding takes place via the TrkA_C domain with a Kd of 10 μM, which is higher than other c-di-AMP binding proteins analysed using DRaCALA. These include L. monocytogenes pyruvate carboxylase (8 μM), CbpA (2.2 μM), CbpB (1.8 μM), PstA (1.3 μM), OpuCA (1.2 μM) and PgpH (0.3–0.4 μM); S. aureus OpuCA (2.5 μM) and KdpD (2 μM); and Enterococcus faecalis OpuCA (6 μM) [5, 8, 16, 44, 45]. The homologous RCK_C domain of KtrA in S. aureus has a Kd of 0.4 μM [6]. Therefore it is possible that the affinities for proteins towards c-di-AMP have evolved to transduce the signal at different threshold concentrations of this nucleotide. We found that glycine betaine levels were significantly lower in the high c-di-AMP ΔgdpP mutant, which were restored by inactivation of busR or suppressor mutations which lower the c-di-AMP level. Low glycine betaine levels due to enhanced repression of busAA-AB by c-di-AMP bound BusR is a likely important contributor to osmosensitive phenotype of ΔgdpP. BusR promoter binding has also been found to be influenced by ionic strength in vitro [46], so it is possible that it senses multiple signals within the cell in order to respond to osmotic challenges.
Whilst the kupB and busR mutations resulted in restoration of osmolyte uptake to allow growth of ΔgdpP under high osmolarity, another way to achieve this is for the cell to simply reduce its intracellular c-di-AMP level. During this screen many destructive mutations in cdaA and restorative mutations in gdpP were observed which lowered the c-di-AMP level. Also in previous work a mutation in glmM was found to downregulate CdaA activity and lower the c-di-AMP level [11]. Two independent intergenic deletion mutations identified in the current work provide a straightforward way to lower the intracellular c-di-AMP pool, which is export. By removing a transcription terminator from a highly expressed ribosomal operon upstream, the suppressor mutants evolved to have a large increase in MDR gene expression. We were unable to demonstrate the MDR Llmg1210 alone was responsible for c-di-AMP export, since Llmg1211 co-expression was needed to stabilise the construct. Llmg1211 (DUF4811) has no homology to proteins with known function, but contains 2 N-terminal transmembrane domains and is located adjacent to Llmg1210 MDR homologs in other species suggesting they have a functional linkage. Interestingly L. lactis MG1363 contains another gene encoding a DUF4811 protein (Llmg1625), which is also located in an operon with MDR and MarR regulator genes upstream. C-di-AMP export has been observed in several pathogens where it triggers an IFN-β innate immune response [33, 47, 48]. In L. monocytogenes, overexpression of MDRs were observed following inactivation of their cognate transcriptional repressors, which led to elevated c-di-AMP export [32, 33]. A strain of L. monocytogenes (LO28) with a naturally occurring mutant tetR exported high levels of c-di-AMP because of strong expression of MdrT [34]. The amino acid identity (45%) between Llmg1210 and MdrT supports its likely function as a c-di-AMP exporter in Lc. lactis, however MdrT has also been shown to act as a bile exporter in L. monocytogenes [49] suggesting it may exhibit broad substrate specificity. Our results demonstrate that active c-di-AMP export is a mechanism by which the cell can modulate intracellular c-di-AMP levels significantly enough to impact cellular physiology, in this case osmoresistance.
Several other mutations were identified in the osmoresistance suppressor screen which triggered a lowering of the c-di-AMP level in ΔgdpP. The most common changes were in a peptide cleavage and export system which is broadly conserved in Gram-positive bacteria. Interestingly, loss of function mutations in Eep (RseP) and PptAB (EcsAB) genes were found during a screen for acid resistant suppressor mutants from a ybbR deletion strain of S. aureus [22]. Like that seen in our study, these mutations lowered the c-di-AMP compared to the parent strain. The connection between peptide processing/export and c-di-AMP level regulation is not clear at present. In B. subtilis, inactivation of EcsAB results in a defect in intramembrane cleavage activity by the Eep ortholog RsaP and it was suggested that peptides not cleared from the membrane may inhibit RsaP activity [50]. Peptides exported by these systems are known to function in cell-to-cell communication as pheromones following re-importation into a neighbouring cell [25, 51]. We tested the ability of spent supernatants from wild-type and ΔgdpP Lc. lactis to induce salt sensitivity in the eep or pptAB suppressor strains, but no effect was observed. Recent work in L. monocytogenes revealed that mutations in the Opp peptide uptake system can restore growth in a DacA (CdaA) mutant [12]. Peptides can function as osmolytes and were shown to be toxic in cells lacking c-di-AMP likely due to an uncontrollable increase in intracellular osmotic pressure [12]. Obtaining osmoresistant suppressor mutants unable to export peptides in ΔgdpP aligns with the hypothesis that elevated osmolyte (peptide) levels within a cell with high c-di-AMP can restore normal turgor pressure and allow growth on high salt agar. One additional mutation was also observed in each of two kupB suppressor mutants. These were in ftsX and pptA and at least for the latter, it is most likely this change which caused a lowering of the c-di-AMP level, similar to other single ppt mutants studied here. It remains to be determined if and how changes in FtsX, which regulates cell wall peptidoglycan hydrolase activity [52] affects c-di-AMP levels.
This study has identified both cellular and external stimuli which trigger significant variations in the c-di-AMP level in bacteria. It was found that enhanced K+ uptake due to gain-of-function mutations in the KupB transporter or simply overexpression of wild-type KupB resulted in elevated c-di-AMP in Lc. lactis and Lb. reuteri. This result is in agreement with recent work showing inactivation of the Trk K+ transporter gating component CabP which is predicted to lower K+ uptake, triggered a lower c-di-AMP level in S. pneumoniae [24]. Our results show that CdaA is activated as a result of higher K+ uptake during growth (since gdpP is inactivated in ΔgdpP). We also found that inactivation of BusR results in higher c-di-AMP suggesting that elevated glycine betaine uptake triggers c-di-AMP accumulation. In the cell suspension assay, under low osmotic conditions, c-di-AMP accumulation was observed under low osmotic conditions, but the addition of ionic or non-ionic solutes triggered a halt in synthesis or increased degradation of c-di-AMP. It therefore appears that the c-di-AMP level is modulated in response to turgor pressure changes as a result of water migration. Uncontrolled K+ or glycine betaine uptake during growth or low osmolarity conditions will result in water entering the cells causing high turgor pressure. In these cases, c-di-AMP accumulation would occur, allowing the cell to subsequently block uptake of K+ and compatible solutes. This then limits excessive cellular hydration and potentially cell lysis. Upon entry into a high osmolarity environment, cells require greater K+ and compatible solutes in order to prevent cellular dehydration and a fall in turgor pressure, so they therefore lower their c-di-AMP level to achieve this. This feedback loop ensures that the cell can quickly sense turgor pressure, and if need be, change the cell’s physiology to regulate water migration through the c-di-AMP signalling receptor network. These environmental changes are directly received at the protein/enzyme level, since the assay involves non-growing cells. Thus, the enzymes involved in c-di-AMP synthesis and degradation could therefore be considered as osmosensors in addition to their roles as osmoregulators.
The mechanisms underlying coordinated synthesis and/or hydrolysis of c-di-AMP in response to increased K+ and glycine betaine uptake, or external osmolarity, are currently unknown. Several possibilities appear possible. Both CdaA and GdpP contain 3 and 2 transmembrane domains which may allow sensing of membrane stretching or curvature which is likely to change under varying turgor pressures. CdaA forms a complex with and is regulated by the membrane bound extracellullar protein CdaR and the peptidoglycan biosynthesis enzyme GlmM in several bacteria [11, 19, 21, 23]. These protein-protein interactions (or individual activities) may be affected by changes in turgor pressure or ionic strength within the cell and affect c-di-AMP synthesis. In L. lactis, cdaR is a non-functional pseudogene, however c-di-AMP levels in this strain are responsive to external osmolarity changes, suggesting that this protein is not essential for sensing. Ultimately, identification of the osmo-signal sensing mechanism of the c-di-AMP system will be of significant interest as it is likely to be conserved across many bacteria.
Lc. lactis strains (S1 Table) were grown at 30°C in M17 media (Difco, USA) supplemented with 0.5% (w/v) glucose (GM17). Lb. reuteri BR11 and Lb. plantarum 299v were grown in deMan Rogosa Sharpe (MRS) media (Oxoid, UK) at 37°C either anaerobically on agar or in static liquid cultures. L. monocytogenes ATCC 19112 and S. aureus IPOOM14235 were grown in Heart Infusion (HI) media (Oxoid, UK) at 37°C with aeration at 150 rpm. E. coli NEB-5α containing pRV300 derivatives were grown in Luria-Bertani (LB) broth containing 100 μg/ml ampicillin at 37°C with aeration at 230 rpm. E. coli NEB-5α containing pGh9 derivatives were grown in HI media (Oxoid, UK) containing 150 μg/ml erythromycin at 30°C with aeration at 230 rpm. Osmoresistant suppressor mutants of ΔgdpP strain OS2 were isolated and confirmed as described before [11]. Sanger sequencing of cdaA and gdpP from the suppressors and whole genome sequencing of 20 mutants using the HiSeq2000 platform was carried out at Macrogen (Seoul, South Korea). SNP analysis was carried out using Geneious 8.1.8. (Biomatters Ltd, New Zealand) as described previously [11].
Plasmids and primers used in this study are shown in S2 and S3 Tables. Lc. lactis was transformed as described previously [10]. Insertional inactivated mutants were made using pRV300 and gene overexpression was done using pGh9. Lc. lactis transformants were grown at 30°C in the presence of 3 μg/ml erythromycin. Lb. reuteri was transformed as described previously [37] and plasmid containing cells were maintained using 10μg/ml erythromycin. For pGh9-kupBA618V transformations into Lc. lactis 0.1–0.2M NaCl was added to the agar as a precaution to prevent mutations occurring. Wild-type or 42 amino acid deleted busR and downstream busAA promoter were cloned into pTCV-lac and introduced into E. coli T7 Express LysY (New England Biolabs) with selection using kanamycin (50 μg/ml). For β-galactosidase activity assays, strains were grown overnight in LB broth without NaCl (10 g/L tryptone; 5 g/L Yeast extract) and then diluted 1:100 in the same fresh LB medium and grown at 30°C, aeration 220 rpm to early log phase (OD600 ~0.25) where 0, 0.1, 0.2, 0.3 or 0.4 M NaCl was added. Following further incubation to OD600 ~0.6, cells were quantified for β-galactosidase as described previously (Miller, 1972), except chloroform and 0.1% SDS were used to permeabilize cells instead of toluene. The busAA promoter (252 bp) was also cloned into pTCV-lac and introduced into different Lc. lactis strains. Promoter activity (β-galactosidase activity) in different strains were compared following growth on GM17 0.1 M NaCl agar supplemented with 3 μg/ml erythromycin and 80 μg X-gal at 30°C for 2 days followed by storage at 4°C for 2 days for colour development. It should be noted that we have found that the ΔgdpP strain can undergo mutations restoring normal c-di-AMP levels during prolonged subculture even under normal growth conditions and caution should be taken when working with high c-di-AMP strains [53]. The ΔgdpP strains were generated with minimal sub-culturing and the cdaA and gdpP genes were checked for mutations and c-di-AMP levels were checked following mutant construction.
Strains were grown to mid-log phase (OD600 ~0.7), then pelleted by centrifugation (5,000 x g for 10 mins), cells then washed 2 times in 1/10 KPM buffer, then re-suspended in 1.5 ml ice-cold extraction buffer (40:40:20 methanol:acetonitrile:ddH2O v/v mix). Lysis was carried out according to that described previously [11]. For determining both extracellular and intracellular c-di-AMP levels, cells were grown in minimal media. First, overnight Lc. lactis cultures were diluted 1:100 into 15ml GM17 broth and incubated at 30°C till OD600 ∼ 0.7. Cells were pelleted by centrifugation at 5000 × g (Beckman Coulter, USA) for 10 min at 4°C, then washed 2 times and re-suspended in 15ml minimal media D6046 (Sigma-Aldrich, St. Louis, MO) supplemented with KH2PO4 3.6mg/ml, K2HPO4 7.3mg/ml, histidine 0.13mg/ml, arginine 0.72 mg/ml, leucine 1mg/ml, valine 0.6mg/ml, glucose 0.5%, potassium acetate 0.9mg/ml, MOPS 13mg/ml, guanine 0.05mg/ml, xanthine 0.05mg/ml, FeSO4 0.10mg/ml, ZnSO4 0.1mg/ml, adenine 0.2 mg/ml and incubated a further 3 hours. Cultures were centrifuged to separate supernatant and cells and supernatants were subsequently filtered (0.22 μm pore size) and used directly to quantify extracellular c-di-AMP. Cells were resuspended in 0.5 ml ice-cold extraction buffer and lysed as described previously [11]. C-di-AMP quantification using ultra performance liquid chromatography-coupled tandem mass spectrometry (UPLC-MS/MS) was carried out as described previously (11) using a different column (HSS PFP 1.8μm, 2.1 x 100 mm) and a BEH C18 VanGuard pre-column protector. Eluent A was composed of 0.1% of formic acid in water while eluent B was 100% acetonitrile. C-di-AMP was detected using electrospray ionization in a negative ion mode at m/z 657.5 → 328.26 and the internal standard 625.52 → 312.26 with collision energy being 30V and 28V, respectively.
Overnight cultures were diluted 1:100 into GM17 broth with erythromycin if required and incubated until late log phase (OD600 ~ 1). Cultures (50 ml) were centrifuged at 5,000 x g (Beckman Coulter, USA) for 10 mins at 4°C and the supernatants were discarded. A second centrifugation was carried out and all media residue was removed by pipette. Cells were subsequently digested with 500μl of 15% HNO3 at 95°C for 1 hour. After digestion, the mixture was centrifuged at 5000 x g for 30 minutes at 4°C. Thereafter, the supernatant was collected to measure K+ content using Vista-Pro, CCD Simultaneous ICP-OES (Varian Inc., USA). The argon gas was ionized and used to create plasma at 7,000–10,000°C and the emission wavelength of 766.491 nm was used for measuring K+. Mean ± SEM were calculated based on three biological replicates.
Full length BusR and the BusR C-terminal RCK_C domain were codon optimised for expression in E. coli and cloned into pRSETA as fusions to His-6 tags (Geneart, Germany). The C-terminal intracellular fragment of KupB was also cloned into pRSETA and pMAL-p5X. The resulting constructs were transformed into E. coli BL21 for expression. Briefly, bacterial cultures were grown at 37°C in LB broth with ampicillin (100 μg/mL) to OD ~ 0.7, then induced with 0.5 mM IPTG at 37°C for 3 hours. Bacteria were pelleted, resuspended in lysis buffer (30mM K2HPO4 pH 8, 300mM NaCl, 1mM PMSF), and lysed by sonication. For purification of the BusR C-terminal domain, after centrifugation the cell lysate was collected and applied to a Ni-NTA resin (Thermo Fisher), washed several times with wash buffer (30mM K2HPO4 pH 8, 300mM NaCl), and eluted with elution buffer (wash buffer + 300mM imidazole). The elute from the Ni-NTA resin was exchanged into nucleotide binding buffer (50mM Tris HCl pH 7.5, 150mM NaCl, 20mM MgCl2) using a PD10 desalting column (GE Healthcare). Radio-labeled c-di-AMP was prepared from 32P-ATP (Perkin-Elmer), and DRaCALA was performed as previously described [16]. Briefly, proteins were incubated with 32P-c-di-AMP for 10 minutes at room temperature, then spotted on a nitrocellulose membrane. Radioactivity was visualized with a Phospho-Imager and a Typhoon imaging system (GE Healthcare).
Lc. lactis were grown overnight in GM17 at 30°C, subcultured 1:100 into 30 ml fresh GM17 and incubated at 30°C till OD600 ~ 0.9 (mid-log phase). Cells were harvested by centrifugation at 5200 x g for 10 min, and washed twice with 2ml 1/10 KPM buffer [0.01M K2HPO4 adjusted to pH 6.5 with H3PO4 and 1mM MgSO4.7H2O]. Cells were resuspended in 0.3 ml 1/10 KPM buffer and 1.2 ml extraction buffer (40% methanol:40% acetonitrile:20% ddH2O v/v). Samples were mixed with 0.5 ml equivalent of 0.1 mm zirconia/silica beads and disrupted using a Precellys 24 homogenizer (Bertin Technologies) three times for 30s each, with 1 min cooling on ice in between. Glass beads were separated by centrifugation at 17000 x g for 5 min. The supernatant was dried under liquid nitrogen before resuspended in 0.5 ml MilliQ water before filtered (0.22 mm pore size). Glycine betaine level was measured using UPLC/MS/MS with HSS PFP column (1.8μm, 2.1 x 100 mm; Waters) and a BEH C18 VanGuard pre-column protector. The method uses was the same with c-di-AMP quantification method described above, with some modifications. Glycine betaine was detected using electrospray ionization in positive ion mode at m/z 118 → 58 with the collision energy being 25eV. The levels were calculated based on a standard curve prepared with glycine betaine from Sigma-Aldrich.
Overnight cultures were diluted 1:100 in GM17 and incubated until OD600 ~ 0.6. To 500μL of culture, 1ml RNA protect reagent (Qiagen, Hilden, Germany) was added and then tubes were vortexed for 5s and held for 5 minutes at room temperature. Cells were harvested by centrifugation (5000 x g for 10min). RNA was extracted using the RNeasy minikit (Qiagen, Hilden, Germany) with some modifications as previously described [10]. cDNA was synthesized using SuperScript III First-Strand Synthesis SuperMix (Invitrogen, Carlsbad, CA). Platinum SYBR green quantitative PCR (qPCR) SuperMix-UDG (Invitrogen, Carlsbad, CA) was used for qPCR using the Rotor-gene Q qPCR machine (Qiagen) with primers described in S3 Table. Test genes rplJ, llmg1209, llmg1210, llmg1211, llmg1212, cdaA and kupB and the reference gene tufA were amplified along with no reverse transcriptase and no template controls. Data was analyzed using the comparative CT method [54] from 3 biological replicates.
Overnight cultures were diluted 1:100 in fresh media and incubated until OD600~0.7 (mid-log phase). Then, 30 ml of the mid-log culture was aliquoted into different tubes for various conditions (30 ml of culture was used for a single c-di-AMP measurement for each condition or time-point replicate). Next, cells were pelleted by centrifugation at 5,000 x g (Beckman Coulter, USA) for 10 min at 4°C and washed twice with buffer (1/10 KPM [0.01M K2HPO4 adjusted to pH 6.5 with H3PO4 and 1mM MgSO4.7H2O]). Cells were resuspended in 1.5 ml buffer and the cells were energised by adding 20mM D-glucose and incubated at 30°C (Lc. lactis) or 37°C (Lb. plantarum, L. monocytogenes and S. aureus) for 10 min to allow ATP production [37, 38]. D-deoxyglucose (20mM) was also tested as it de-energises cells and blocks ATP synthesis. Samples were either taken at this point for the time-course experiments or additional 90 μl ddH2O or NaCl (0.1M or 0.3M final concentration) or sorbitol (0.2M or 0.6M) or sucrose (0.2M or 0.6M) was added. At the 20 min time point the assay was stopped. To prevent any changes in external osmolarity or effects of centrifugation, at the harvest time point the cell suspensions were mixed directly with 0.5 ml equivalent of 0.1 mm zirconia/silica beads (Daintree Scientific, Australia) and were immediately lysed using a Precellys24 homogeniser (Bertin Instruments, France) three times for 30 seconds at 6,000 rpm with 1 min cooling on ice in between. Beads were separated by centrifugation at 16,873 x g (Eppendorf, Germany) for 5 min. The supernatant was mixed with methanol/acetonitrile producing a final v/v ratio of 40:40:20 methanol:acetonitrile:supernatant. The sample was centrifuged at 16,873 x g for 5 min and the supernatant was air-dried with nitrogen at 40°C before being resuspended in 0.5 ml ddH2O and filtered (0.22 μm pore size). Levels of c-di-AMP were determined with UPLC-MS/MS as above.
Overnight cultures of WT Lc. lactis and ΔkupB were diluted 1:100 in fresh GM17 media and grown OD600~0.7. Two millilitres of culture was centrifuged at 11000 x g for 3 mins and the cell pellets were washed 2 times in chemically defined media ZMB1 [55] with some modifications as follows. The potassium salts K2SO4 and KI were omitted and potassium phosphate buffers were replaced with NaH2PO4 (2.736 g/L) and Na2HPO4 (5.23 g/L). The washed cell pellets were re-suspended in 2 ml of modified ZMB1 and 5 μL of the cell suspension was inoculated into 5 ml of media with various concentrations of KCl (0.1 mM, 0.5 mM, 1 mM, 5 mM, 10 mM, 20 mM or 50 mM). The cultures were incubated at 30°C for 19 hours, at which point the OD600 was determined.
|
10.1371/journal.pntd.0002419 | Update on the Mapping of Prevalence and Intensity of Infection for Soil-Transmitted Helminth Infections in Latin America and the Caribbean: A Call for Action | It is estimated that in Latin America and the Caribbean (LAC) at least 13.9 million preschool age and 35.4 million school age children are at risk of infections by soil-transmitted helminths (STH): Ascaris lumbricoides, Trichuris trichiura and hookworms (Necator americanus and Ancylostoma duodenale). Although infections caused by this group of parasites are associated with chronic deleterious effects on nutrition and growth, iron and vitamin A status and cognitive development in children, few countries in the LAC Region have implemented nationwide surveys on prevalence and intensity of infection. The aim of this study was to identify gaps on the mapping of prevalence and intensity of STH infections based on data published between 2000 and 2010 in LAC, and to call for including mapping as part of action plans against these infections. A total of 335 published data points for STH prevalence were found for 18 countries (11.9% data points for preschool age children, 56.7% for school age children and 31.3% for children from 1 to 14 years of age). We found that 62.7% of data points showed prevalence levels above 20%. Data on the intensity of infection were found for seven countries. The analysis also highlights that there is still an important lack of data on prevalence and intensity of infection to determine the burden of disease based on epidemiological surveys, particularly among preschool age children. This situation is a challenge for LAC given that adequate planning of interventions such as deworming requires information on prevalence to determine the frequency of needed anthelmintic drug administration and to conduct monitoring and evaluation of progress in drug coverage.
| Soil-transmitted helminths (STH) are part of the group of neglected infectious diseases (NID) in Latin America and the Caribbean (LAC), and are associated with several adverse chronic effects on child health. Although control interventions such as periodic administration of anthelmintic drugs, health education, improved access to safe water and sanitation, among others, are acknowledged to be an important means to reduce morbidity and to achieve control, epidemiological information on prevalence status is lacking at the lowest sub-national administrative levels (municipalities, districts or provinces) in many countries thus hindering decision making regarding not only the treatment, but also the monitoring of progress in deworming coverage, the assessment of epidemiological impact on parasite prevalence and load and, therefore, the achievement of the overall public health goals. Epidemiological surveys can be expensive and require time and effort for their implementation, which could explain the low number of studies published with data on prevalence and intensity of infection in the Americas. The use of alternative methodologies, for instance those based on geographical information systems and remote sensing technologies, or of sentinel surveillance in schools may help countries in the task of collecting information and support the implementation of integrated control programs against STH.
| Helminth infections impose a great and often silent burden of morbidity and mortality on poor populations in developing countries. The most common helminth infections are caused by soil-transmitted helminths (STH): roundworms (Ascaris lumbricoides), whipworms (Trichuris trichiura), and hookworms (Necator americanus and Ancylostoma duodenale). Worldwide estimates suggest that A. lumbricoides infects 1.221 billion people, T. trichiura, 795 million, and hookworms, 740 million. Infections occur most frequently in the Americas, China and East Asia, and Sub-Saharan Africa [1].
STH are amongst the most prevalent pathogenic organisms on the planet, estimated to infect almost one-sixth of the global population with the highest rates among school-age children (SAC) who are frequently infected with two or more species at a time [2]. Stunting usually occurs between 6 months and 2 years of age, overlapping with the period in which STH begin to emerge [3]. STH infection primarily affects physical and cognitive development [4]. A. lumbricoides can cause malnutrition; hookworms damage the intestinal mucosa leading to bleeding, loss of iron and anemia, and infections by T. trichiura cause chronic reduction of food intake [5]. During pregnancy, mild or severe infections by hookworms can cause anemia in the mother and damage to the fetus, leading to low birth weight [6]. In areas where helminths are common, deworming activities can be done once or twice a year among the population at risk (those with no access to improved sanitation facilities), including deworming for pregnant women after the first trimester. Deworming during pregnancy reduces severe maternal anemia, increases birth weight and reduces infant mortality [7]. Thus, regular treatment against helminth infections produces both immediate and long-term benefits, contributing significantly to improving the growth and cognitive development of affected individuals, especially children.
In 2001, the World Health Assembly adopted Resolution WHA/54.19 [8] urging all Member States where STH are endemic to attain “a minimum target of regular administration of chemotherapy to at least 75% and up to 100% of all SAC at risk of morbidity by 2010”. On October 2009, the Directing Council of the Pan American Health Organization (PAHO) approved Resolution CD49.R19 [9] stating the commitment of PAHO's Member States to eliminate or reduce neglected diseases, among them STH, to levels such that they are no longer considered public health problems by the year 2015, and hence help to achieve the Millennium Development Goals. In PAHO's Resolution, STH and schistosomiasis were classified as diseases whose prevalence can be drastically reduced with available cost-effective interventions. Regarding STH, the following goal to be reached by 2015 was defined: reducing prevalence among SAC in high risk areas (prevalence >50%) to less than 20% as measured by quantitative egg count in feces. The Resolution also mentioned several interventions to reach the STH control goal, especially those related with improved access to safe water and basic sanitation, preventive chemotherapy and health education through inter-sectoral collaboration.
According to WHO/PAHO estimates, in LAC there were 13.9 million preschool-age children (PSAC) and 35.4 million SAC in need of preventive chemotherapy for STH in 2012. These estimates have been calculated based mainly on the percentage of people without access to improved sanitation facilities, differentiated by rural and urban areas, due to the fact that STH prevalence and intensity of infection in LAC are not well mapped [10].
The purpose of this paper is to present the status of the mapping of prevalence and intensity of STH infection and to identify information gaps in LAC for the 2000–2010 period based on a literature search. This study is a call for action in LAC to address the existing information gaps in order to prioritize integrated interventions for STH control based on solid evidence, increase efforts towards reduction of the morbidity caused by these parasites, and reach the targets in the WHO and PAHO resolutions.
A wide literature search was conducted to collect data on STH prevalence and intensity of infection (Ascaris lumbricoides, Trichuris trichiura and hookworms) among preschool (1–4 years of age) and school-age (5–14 years of age) children for the 2000–2010 period at the lowest subnational administrative levels in LAC countries (Checklist S1). The decision to include studies published between 2000 and 2010 was taken arbitrarily by the authors considering this 10-year period to have sufficient updated information on the status of STH mapping in the Region to establish the current information availability, reflecting the level of interest on this issue in the Region. A database was built including information from 236 studies, of which 120 met the inclusion criteria [11]–[130]. There are 45 countries and territories in LAC with at least 13,591 units at second subnational level that may be districts, municipalities or provinces depending on the geopolitical structure of each country or territory. Once the information was collected, a preliminary report was published in PAHO's website [131]. That document is the main information source for the analysis presented here.
The scientific literature search was done through the online databases of PubMed (including MEDLINE), LILACS (including SciELO), BIREME and Cochrane. Additionally, we carried out a search of information published in the websites of health ministries, NGOs and FBOs reporting data on deworming activities to PAHO and WHO from 2005 to 2010, as well as information published on the websites of PAHO country offices in LAC. The online database search was done using MeSH terms to facilitate an ample retrieval of published information on STH prevalence and intensity of infection in LAC. The following MeSH terms and subheadings were used for searches on PubMed: ((“Helminthiasis”[Mesh] OR (“Helminthiasis/epidemiology”[Mesh] OR “Helminthiasis/parasitology”[Mesh] OR “Helminthiasis/statistics and numerical data”[Mesh]))) AND (“Child”[Mesh] OR (“Child/epidemiology”[Mesh] OR “Child/statistics and numerical data”[Mesh])). The following terms were also used for searches on PubMed to recover more papers: 1) ‘Prevalence intestinal parasites child’ restricted by country, sub-regions in LAC (Central American Isthmus, Latin Caribbean, Andean area, Southern Cone, Non-Latin Caribbean) and publication year of study; 2) ‘Soil transmitted helminths prevalence’ restricted by country, sub-regions in LAC (Central American Isthmus, Latin Caribbean, Andean area, Southern Cone, Non-Latin Caribbean) and publication year of study; and 3) ‘Ascaris lumbricoides or Trichuris trichiura or hookworms or Necator americanus or Ancylostoma duodenale’ by country, sub-regions in LAC (Central American Isthmus, Latin Caribbean, Andean area, Southern Cone, Non-Latin Caribbean) and publication year of study. Additionally, for the search of papers in BIREME and LILACS databases, the following terms were used: 1) “Helminthiasis” AND “prevalence” by country and LAC; and 2) “Intestinal parasites” AND “prevalence” by country and LAC.
The following were the inclusion criteria: 1) studies with data on prevalence and intensity of infection published from 2000 to 2010 including the geographical location of the study so as to enable the identification of the municipal or local level; and 2) studies including STH prevalence data disaggregated by age groups. The following were the exclusion criteria: 1) studies undertaken before 1995; 2) reports with data on intensity of infection not using WHO classification parameters (mild, moderate and high), and 3) studies with no data on their geographical location or including data on prevalence and intensity of infection of limited use for the present analysis (e.g. studies reporting prevalence data only for adults, or studies in children with eosinophilia). Given the limited number of data published on STH prevalence and intensity of infection in the Region, no restriction was established regarding sample size, type of study (community-based or among school children, base line or intervention monitoring) or laboratory diagnostic methods used, as the aim was not to verify the existence of accurate data for the two indicators in the Region, but to explore existing information and gaps on mapping (Figure S1).
After reviewing the articles and reports, information was extracted from those that met the inclusion criteria and then a database was built with the following variables: name of country; name of geographical location of study; name of geographical locations within the first and second subnational levels corresponding to the geographical location of the study (in those cases where no information on this regard appeared, the publication year was recorded), sample size, STH prevalence, prevalence of infection by species, intensity of infection and age group of study subjects. Each value of prevalence and intensity of infection registered on the database was denominated a data point. In those studies that included data for several geographical locations at the lowest subnational level (e.g., municipalities), data were registered for each of these locals and, therefore, we were able to extract more than one prevalence or intensity of infection data point from several studies.
A descriptive analysis of the number of studies including data on prevalence and intensity of infection published from 2000 to 2010 was done by country, age group and prevalence and intensity of infection range. Although the search was restricted to studies published from 2000 to 2010, some authors included results from surveys conducted before the study publication date, and for this reason our analysis included data only from studies carried out from 1995 onwards. Besides the analysis of frequency and proportion distributions, the geographic locations of prevalence and intensity of infection data points for preschool and school age children were mapped, as this was useful to visualize gaps in data publishing. The database was made with MS Excel 2010 and the analysis with Tableau 7.0.
A total of 236 publications were found, of which 120 met the selection criteria established for the study; the publications corresponded to 18 countries: Brazil (39 publications, 32.5%), Argentina (11, 9.2%), Colombia (10, 8.3%), Venezuela, Mexico, Ecuador (9, 7.5% each), Peru (7, 5.8%), Cuba (5, 4.2%), Honduras, Bolivia (4, 3.3% each), Guatemala (3, 2.5%), Haiti, Costa Rica, Belize (2, 1.7% in each country), Saint Lucia, Paraguay, Nicaragua and Guyana (1, 0.8% each). All publications were articles published in scientific journals. Although some documents were found in the websites of health ministries, NGOs and FBOs, none of them included information meeting the inclusion criteria. The studies recovered by the Cochrane database had no information on their geographic location and, therefore, they were not considered.
A total of 335 data points on STH prevalence were registered and analyzed for 18 countries out of which 12.0% were for preschool-age children (40 prevalence data points), 56.7% for school-age children (190 prevalence data points) and the remaining 31.3% (105 prevalence data points) were registered for children under 15 years of age because authors did not differentiate by age groups, although they did state in the methodology that the study had been undertaken among children under 15 years of age. Another 687 data points on prevalence by STH species were also extracted, of which 40.2% corresponded to A. lumbricoides, 35.5% to T. trichiura, 19.5% to hookworms (undifferentiated by species), 2.9% to A. duodenale, and 1.9% to N. americanus. Additionally, 151 data points were extracted for intensity of infection.
The main characteristics of the 335 data points for STH infection prevalence were the following: 1) 27.8% of prevalence data points showed values over 50%, 34.9% between >20 and 50%, and 37.3% below 20%; 2) 63% of prevalence data points corresponded to data published by four countries (Argentina, Brazil, Honduras and Mexico); 3) 35% or more of STH prevalence data from Belize, Ecuador, Guatemala, Honduras, Mexico and Venezuela showed values over 50%. Prevalence ranges found in each country are shown in Table 1. The geographic locations of STH prevalence data points for PSAC and SAC by prevalence ranges are shown in Figures 1 and 2.
A total of 151 infection intensity data points from seven countries (Argentina, Bolivia, Brazil, Colombia, Ecuador, Honduras and Venezuela) were registered and analyzed. The following were their main characteristics: 1) The largest number of data points was found for Honduras (41.7%), Bolivia and Brazil (15.9% each); 2) 56.9% of infection intensity data corresponded to SAC, 17.8% to PSAC and 25.1% to children under 15 years of age; 3) 37.1% showed light, 34.4%, moderate and 28.5% heavy intensity infections (Table 2); 4) 65.1% of the data points on heavy intensity infections corresponded to A. lumbricoides (28 data points), and 5) no data on intensity of infection published after 2005 were found.
The most recent year in which prevalence data points were found varied among countries. Regarding prevalence data points for SAC, and taking as the most recent data those from studies conducted after 2005, we observed that 10 countries (out of 16 with data points) had studies conducted from 2005 to 2010. As for preschool age children, six countries (out of 11 with data points) had studies, while nine countries had studies for children under 15 years of age (out of 16 with data points).
A total of 335 data points on STH prevalence from 18 LAC countries were found in the 120 articles included in the present study that were then registered and analyzed; 12.0% of the data were for PSAC, 56.7% for SAC and 31.3% for children under 15 years of age. We found that 60% of the prevalence data points showed prevalence levels above 20%. If these data were to be used for decision making regarding deworming in those geographic areas, it could be concluded that in more than half of such places deworming would be necessary for all children under 15 years of age (excluding children under 1 year of age) at least once a year. We also registered and analyzed 151 infection intensity data points from seven countries, of which 37.1% showed light, 34.4%, moderate and 28.5%, heavy intensity infections; no data published on intensity of infections were found after 2005. Our analysis suggests that there is a significant gap in data published on STH prevalence and intensity of infection in LAC, especially for PSAC.
No restriction was established regarding sample size, sample randomization, type of study (community-based or among SAC, baseline or post-intervention monitoring) or laboratory diagnostic methods used, and, therefore, the study does not allow for making any inferences. Nevertheless, this information is useful to recognize existing data gaps on prevalence and intensity of infection for STH, independently of the quality of the few data found, and emphasizes the need to make a call for action aimed at collecting evidence-based information required to define needed deworming activities for public health purposes in LAC countries. The results also emphasize the need to promote the development of studies that are epidemiologically robust, with a certain standardization of both the epidemiological and laboratory methods, following WHO guidelines, that will provide reliable information on the epidemiological situation of STH and allow for comparability between studies done in different geographical regions and at different points in time. Another limitation of the study could be the fact that the data from some reports published by health ministries and organizations (NGOs and FBOs), as well as from undergraduate and graduate theses and other type of documents containing information on studies about prevalence and intensity of infection could have been omitted due to the methodology used for data recovery that focused mainly on published indexed data accessible on the web. Due to the aforementioned limitations and the bias that this could represent for this study, inferences or use of these data to estimate prevalence and intensity of infection for STH and its distribution in LAC, or for making decisions on deworming activities to be implemented in any area of the 18 countries included in the analysis should not be done. Countries are encouraged to conduct their own scientifically sound studies of prevalence and intensity of STH infection, where data gaps are seen.
During the period from January 2000 to June 2010 data on STH prevalence and intensity of infection at the local level were recovered only for 18 LAC countries (Argentina, Belize, Bolivia, Brazil, Colombia, Costa Rica, Cuba, Ecuador, Guatemala, Guyana, Haiti, Honduras, Mexico, Nicaragua, Paraguay, Peru, Saint Lucia and Venezuela) despite the fact that it is estimated that at least 30 countries in the Americas are endemic for STH infections according to a document published by PAHO in 2010 [131]. In this report, LAC countries were classified into four groups based on the estimated number of PSAC and SAC at risk of STH infections, as well as on the presence of other neglected infectious diseases (NIDs) targeted for preventive chemotherapy as one of the tools to reach their control and elimination. This classification of countries in four groups was done from analysis of quantitative and qualitative variables. The group of quantitative variables included sanitation coverage, population at risk for STH, deworming coverage on PSAC and SAC population, population at risk, prevalence and treatment coverage for onchocerciasis, schistosomiasis, blinding trachoma and lymphatic filariasis. The qualitative variables included inter-programmatic actions already in implementation in the countries, partners supporting deworming, needs of technical cooperation, progress on mapping, and opportunities for integrated actions.
These four groups included 33 countries as follows: the first group comprised 11 countries (Bolivia, Brazil, Ecuador, Guatemala, Guyana, Haiti, Mexico, Peru, Dominican Republic, Saint Lucia and Suriname) concentrating 66.8% of PSAC and 67.4% of SAC at risk of STH infections in the Region. The second group included six countries (Belize, Colombia, El Salvador, Honduras, Panama and Venezuela) with 26.8% of PSAC and 26.1% of SAC at risk. The third group comprised three countries (Argentina, Paraguay and Nicaragua) with 5.4% of children at risk from both age groups, and the fourth group of 13 countries registering 1% of children at risk from both age groups. If deworming activities were focused in the 17 countries in groups 1 and 2, 94% of PSAC and SAC at risk of STH infections in the Region would be reached and protected [132].
Of the 335 data points from the 18 LAC countries analyzed in the study, 62.7% of data points showed STH infection prevalence above 20% (including both ranges, 20–50% and >50%). If these data were representative of the geographic areas to which they correspond, and if a country was interested in implementing deworming using these data without waiting for data of a survey of prevalence to make a decision on number of treatment rounds needed, it would be necessary to conduct deworming at least once a year for all children aged between 1 and 15 years there. According to information provided by PAHO's Regional NID Program, by 2011 a total of eight countries (Dominican Republic, Guyana, Haiti, Mexico, Belize, Honduras, Nicaragua and Cuba) had deworming programs at national and subnational levels while four other countries had started updating their mapping of prevalence and intensity of infection at national level (Brazil, Ecuador, El Salvador and Honduras), while Suriname had finished a national survey in 2010. As a result of the advocacy activities waged by PAHO and other regional partners, and based on meetings held with national health authorities, it is expected that by 2013 at least 11 countries from groups 1, 2 and 3 prioritized by PAHO (Bolivia, Dominican Republic, Guatemala, Guyana, Haiti, Peru, Colombia, Panama, Argentina, Paraguay and Venezuela) will update their mapping of STH prevalence and intensity of infection, and that four countries will conduct mapping to evaluate the impact of their national deworming programs (Belize, Mexico, Nicaragua and Saint Lucia) (Table 3).
Although the methodologies recommended by PAHO/WHO to estimate STH prevalence and intensity of infection indicate that surveys among SAC are enough to establish prevalence and intensity of infection levels in a community [133], the amount of published prevalence data found in the present study for PSAC (12.0%) is very low compared with the data recovered for SAC (56.7%). The lack of information regarding PSAC is a challenge for the Region where, according to estimates made for 2011 based on the WHO-recommended algorithm [10], one fifth of PSAC and SAC were at risk of STH infections (around 13.9 and 35.4 million of a total of 61 and 155 million PSAC and SAC, respectively).
Although this indicator of children at risk of infection is an indirect way of identifying areas and population groups requiring massive deworming interventions, it does not take into account differences within countries regarding ecological zones and deworming activities that may be underway. Therefore, mapping (be it through population-based surveys, sentinel surveillance or the use of methods based on geographic information systems and remote sensing technologies that combine epidemiological, demographic, climate, social and economic variables to estimate geographic areas and populations at risk of several diseases) should be promoted in all LAC countries to identify the administrative units at second subnational level (districts, municipalities or provinces depending on the geopolitical structure of each country or territory) and focus on the implementation of deworming interventions among population groups at risk due to their living conditions. Such mapping should be promoted not only as a baseline to start deworming and other interventions in each country, but also as part of the monitoring and evaluation of progress towards STH control goals, as well as to modulate deworming activities (increasing, sustaining or reducing them).
According to data from the PAHO Regional NID Program, the countries that have conducted national mass deworming programs maintaining coverage above 75% for more than five years are Mexico (two or three rounds per year for SAC), Nicaragua (one round per year for PSAC and SAC), Dominican Republic (one round per year for SAC) and Honduras (a round per year for SAC). However, it is worth noting that despite the limitations already mentioned, our study found data published on prevalence levels above 20% and even above 50%. This may indicate that it is necessary to insist on mapping differentiated by ecological zones and population groups living in at-risk conditions that perpetuate STH transmission cycles given that national averages, and sometimes even subnational averages, can mask the real local situation.
The low amount of data published on STH intensity of infection in LAC during the period under study (151 data points from seven countries) is also noteworthy, especially the lack of data published for this indicator after 2006 at the lowest subnational administrative levels. It may be that this indicator was not estimated in the studies found, or that it was estimated but not included in the publications. However, it has been clearly documented that moderate to heavy intensity infections have significant implications for children's physical and cognitive development, as well as for severe or even fatal complications among children under 15 years of age, especially school age children [4]–[6]. Our study also found that in 63% of the data published intensity of infection ranged from moderate to heavy. It is important to suggest to health authorities in LAC countries that when massive and sustained deworming is implemented, the indicator showing the fastest change towards reduction is intensity of infection and that this has a remarkably positive impact on the well-being of children [134]. The reduction in infection prevalence will be slower, as it depends on the effect of other social determinants such as access to safe water, basic sanitation, nutrition, use of footwear and housing improvement (e.g., eliminating dirt floors, adding ventilation, lighting), amongst others.
In 2011, PAHO published an updated mapping of NIDs in LAC based on secondary sources of information [135]. The study included data on STH prevalence at first subnational level (departments, states or provinces depending on each country's characteristics) for several countries and the aim was to have an approximate idea of which areas would need to implement deworming activities at least once a year. The study also concluded that there were information gaps for several countries. Through this new search of scientific literature, PAHO's Regional NID Program wanted to analyze data availability at the lowest subnational administrative levels and learn about the status of mapping.
As indicated by WHO, mapping based on secondary sources (i.e., surveys published in recent years, official reports on prevalence surveys, undergraduate and graduate papers on STH prevalence) are useful as a first approach when the epidemiological status of a given geographic area is unknown. However, mapping quality varies according to the quality of the data published in the different secondary sources and it does not replace the power and strength of population-based studies whose aim is to establish baseline epidemiological data for STH or evaluate the impact of deworming programs, particularly those conducted at large scale and sustained through time [133].
Despite the limitations of this study, the dataset has been useful for further analysis. PAHO's Regional Neglected Infectious Diseases Program joined efforts with the Department of Epidemiology and Public Health at the Swiss Tropical and Public Health Institute to develop ecological niche-based models and Bayesian geostatistical models to predict the disease distribution at municipality level including STH. This dataset was shared to contribute to a manuscript on geostatistical meta-analysis of STH in South America whose results were published in 2013 and the data were included on the Global Database for Mapping, Control, and Surveillance of Neglected Tropical Diseases (GNTD) [136]. The dataset was shared also with the Inter-American Development Bank (IADB) as part of a joint initiative to carry out analysis on correlation between STH prevalence and some determinants as access to water and sanitation, education and housing features whose results were published in 2013 [137].
Even though by the time that this study was completed data for STH in LAC were not available on the Global Atlas for Helminth Infections (GAHI) nor on the GNTD, countries in LAC could benefit from data already available by 2013, and thus have useful data to support control activities for STH. Additionally, PAHO encourages countries in LAC to share information with the aforementioned initiatives that contribute for a better understanding of distribution of prevalence of STH and thus support countries in LAC to move forward the agenda for the control and elimination of neglected infectious diseases, in general, and in particular of STH.
It is necessary and urgent to update the mapping of STH prevalence and intensity of infection in several LAC countries in order to make better evidence-based decisions regarding deworming activities. The data available for PSAC are insufficient to know the real situation of STH prevalence and intensity of infection in many countries in the LAC region, and although more data for SAC population were found, these data are only from a limited number of countries and some second administrative levels.
It is also necessary to prioritize the operational research agenda within governments and interested groups in order to develop STH mapping activities. Mapping is necessary to know where are the populations at risk and which age groups are at risk of STH infection, as well as the municipalities where authorities need to focus their efforts (if these data are available), and also for monitoring and evaluation purposes to know the impact of interventions, not only deworming, but also nutrition, education, and environmental interventions and integrated actions to reduce child morbidity and mortality and improve child development.
Without enough accurate and specific data about STH prevalence and intensity of infection by age group (PSAC and SAC) in LAC, it will be difficult to identify with certainty the main needs, resources to be assigned and places where integrated actions, including deworming, must be focused in order to reach the regional and global deworming goals for children at risk and reduce the prevalence and intensity of infections by STH.
|
10.1371/journal.pbio.1002127 | Colour As a Signal for Entraining the Mammalian Circadian Clock | Twilight is characterised by changes in both quantity (“irradiance”) and quality (“colour”) of light. Animals use the variation in irradiance to adjust their internal circadian clocks, aligning their behaviour and physiology with the solar cycle. However, it is currently unknown whether changes in colour also contribute to this entrainment process. Using environmental measurements, we show here that mammalian blue–yellow colour discrimination provides a more reliable method of tracking twilight progression than simply measuring irradiance. We next use electrophysiological recordings to demonstrate that neurons in the mouse suprachiasmatic circadian clock display the cone-dependent spectral opponency required to make use of this information. Thus, our data show that some clock neurons are highly sensitive to changes in spectral composition occurring over twilight and that this input dictates their response to changes in irradiance. Finally, using mice housed under photoperiods with simulated dawn/dusk transitions, we confirm that spectral changes occurring during twilight are required for appropriate circadian alignment under natural conditions. Together, these data reveal a new sensory mechanism for telling time of day that would be available to any mammalian species capable of chromatic vision.
| Animals use an internal brain clock to keep track of time and adjust their behaviour in anticipation of the coming day or night. To be useful, however, this clock must be synchronised to external time. Assessing external time is typically thought to rely on measuring large changes in ambient light intensity that occur over dawn/dusk. The colour of light also changes over these twilight transitions, but it is currently unknown whether such changes in colour are important for synchronising biological clocks to the solar cycle. Here we show that the mammalian blue–yellow colour discrimination axis provides a more reliable indication of twilight progression than a system solely measuring changes in light intensity. We go on to use electrical recordings from the brain clock to reveal the presence of many neurons that can track changes in blue–yellow colour occurring during natural twilight. Finally, using mice housed under lighting regimes with simulated dawn/dusk transitions, we show that changes in colour are required for appropriate biological timing with respect to the solar cycle. In sum, our data reveal a new sensory mechanism for estimating time of day that should be available to all mammals capable of chromatic vision, including humans.
| The ability to predict and adapt to recurring events in the environment is fundamental to survival. Organisms across the living world achieve this using endogenous circadian clocks [1–3]. However, if such clocks are to fulfil their ethological function they need to be regularly reset to local time. This is achieved by sensory inputs that report changes in the physical environment providing a useful proxy for time of day. By far the best characterised of these input pathways is that recording the diurnal change in the overall quantity of light reaching the earth’s surface (irradiance). In the case of mammals, a dedicated retino-hypothalamic projection brings this visual information to the brain’s “master” clock in the suprachiasmatic nuclei (SCN) [4–7].
The retino-hypothalamic projection is formed by a unique class of retinal ganglion cells (RGCs), which are intrinsically photosensitive thanks to their expression of melanopsin [4]. Although these so-called ipRGCs can therefore entrain the clock even in the absence of the conventional rod and cone photoreceptors, all photoreceptor classes are capable of influencing the clock in intact animals [8–13]. This arrangement has previously been considered only insofar as it allows the clock to respond to changes in irradiance. However, the inclusion of cones in this pathway allows for the possibility that the clock could also receive information about changes in the spectral composition of light (colour) [14]. There has previously been speculation that such colour signals could provide a reliable method of telling time of day [3,15] but, to date, there has been no direct test of that possibility in mammals.
The ability to discriminate colour relies on comparing the relative activation of photopigments with divergent spectral sensitivities. In mammals, this task is achieved via differential processing of cone photoreceptor signals in the retina [16]. At least 90% of mammalian species are believed capable of this form of colour discrimination [17] which, with the exception of Old World primates, allows for dichromatic vision. Thus, most mammals express just two distinct classes of cone opsin, one maximally sensitive to short wavelengths (ultraviolet–blue) and a second with peak sensitivity to longer wavelengths (green–red) [18]. Here we show that, in mice, this primordial colour discrimination axis (equivalent to human blue–yellow colour vision) is an influential regulator of SCN activity, essential for appropriate circadian timing relative to the natural solar cycle.
We first set out to determine whether changes in spectral composition associated with the earth’s rotation could provide reliable information about solar angle that the circadian clock could use to estimate time of day. To this end, we obtained high resolution measurements of natural variations in spectral irradiance across multiple days (Manchester, August–October 2005, n = 36 d).
As expected, these measurements revealed highly predictable changes in both irradiance and spectral composition as a function of solar angle (Fig 1A and 1B). In particular, we observed a progressive enrichment of short-wavelength light across negative solar angles: a result of the increasing amount of ozone absorption and consequent Chappuis band filtering of green–yellow light when the sun was below the horizon [14].
We next calculated the extent to which this change in spectral composition was detectable to the mammalian visual system. Taking the mouse as a representative species, we employed previously validated approaches to quantify the relative excitation of its ultraviolet and medium wavelength sensitive (UVS/MWS) cone opsins [12,19,20]. This analysis revealed a robust change in the ratio of excitation between the two pigments (Fig 1C) that would constitute substantial changes in apparent colour along the blue–yellow axis. These changes were restricted to the twilight transition, with the UVS:MWS ratio fairly invariant throughout the day, indicating that measuring the change in colour could provide a useful method of tracking the progression of dawn and dusk.
To ascertain how reliably this blue–yellow colour signal alone could be used to estimate phase of twilight, relative to simple measures of irradiance, we next compared the day-to-day variability of colour and irradiance measurements across our dataset (Fig 1D). Surprisingly, we found that colour was in fact more predictive of sun position across twilight (-7 to 0° below horizon) than was irradiance (78.5 ± 0.1% versus 75.8 ± 0.1% of variance explained by solar angle; mean ± SD). Accordingly, for any fixed solar angle, the range of observed colour values was considerably more tightly clustered than those for irradiance. These observations most likely reflect the fact that cloud cover can change overall brightness quite dramatically, but exerts only relatively minor effects on spectral composition.
Importantly, then, measuring colour could provide a more reliable estimate of the approach of night or day than measuring irradiance. However, while the mammalian circadian clock is certainly known to respond to diurnal variations in irradiance [8–13], there has been no investigation of whether the SCN also receives colour signals. Accordingly, we next asked whether the central clock showed electrophysiological responses to changes in colour by recording extracellular activity in the mouse SCN.
In order to identify colour-sensitive cells, we set out to generate test stimuli which differentially modulated the UVS and MWS mouse cone opsins. While, in principle, producing such stimuli is straightforward, the close spectral sensitivity of mouse opsins makes it difficult to achieve this aim without concomitant changes in the activation of rods and/or melanopsin. To circumvent this problem, we employed a well-validated transgenic model in which the native mouse MWS opsin is replaced by the human long-wavelength sensitive (LWS) opsin (Opn1mwR; [8,12,19]). Cones in these animals develop and function normally, with LWS opsin expression entirely recapitulating that of the native MWS opsin [21]. Importantly, however, the resultant shift in cone spectral sensitivity in Opn1mwR mice facilitates the generation of stimuli that provide selective modulation of individual opsin classes [22].
Using this Opn1mwR model, we first established a background lighting condition (using a three-primary LED system), whose spectral composition recreated a wild-type mouse’s experience of natural daylight (S1A Fig). We next designed a set of manipulations of this background spectrum that allowed us to modulate excitation of one or both cone opsins without any concomitant change in rod or melanopsin activation. Under these conditions, we were then able to unambiguously distinguish colour-sensitive neurons based on the following criteria: (1) the presence of larger responses to chromatic versus achromatic changes in cone excitation and (2) responses of opposite sign to selective activation of UVS and LWS opsin in isolation (i.e., excitatory/ON versus inhibitory/OFF).
To achieve the largest possible change in colour, we started by selectively modulating UVS and LWS cone opsin excitation in antiphase (“colour”; S1B Fig). We then compared responses to this stimulus with those elicited by one in which the change in UVS and LWS opsin activation occurred in unison (“brightness”; S1B Fig). Any spectrally opponent cells should be more responsive to the “colour” as opposed to “brightness” condition. We found that 17/43 SCN units (from 15 mice) that responded to these stimuli showed a significant preference for the pair in which UVS and LWS activation was modulated in antiphase (Fig 2A and 2B; paired t test, p<0.05, n = 17). As there were an additional 26 visually responsive units that did not respond to either of these analytical stimuli (paired t tests, p>0.05), these data indicate that at least one quarter of light-responsive SCN neurones show chromatic opponency.
Interestingly, we found that cone inputs exerted a much more powerful influence over the firing activity of cells exhibiting a preference for chromatic stimuli relative to than achromatic cells (Fig 2B; absolute change for responses of chromatic cells = 8.1 ± 2.3 spikes/s versus 1.9 ± 0.5 spikes/s for achromatic cells, n = 17 and 26 respectively; t test: p<0.01). Moreover, we found that the spiking activity for the majority (13/17) of colour-sensitive SCN neurons was highest during the stimulus phase biased towards UVS opsin activation (Fig 2A), and that these cells exhibited especially robust and sustained changes in firing (Fig 2A and 2B).
Our data above therefore indicate that cone inputs constitute a dominant influence on the firing activity of colour-sensitive SCN neurons and that most of these cells exhibit blue-ON/yellow-OFF colour opponency. We confirmed this by selectively modulating brightness for each of these cone opsins independently (stimulus shown in S2A Fig); as expected, these cells reliably increased firing in response to selective increases in UVS opsin activation and decreased firing following increases in LWS opsin activation (Fig 2C). Conversely, the remaining colour-sensitive cells exhibited the opposite preference (yellow-ON/blue-OFF; Fig 2C).
An aspect of mouse retinal organisation that poses a challenge to colour vision is that most cones in this species co-express UVS and MWS opsin [23]. The exceptions are rare “primordial S-cones” that only express UVS opsin [24] and peripheral cones that may express either pigment alone [25]. One might expect that chromatic opponency would rely on comparisons between these rare single pigment cones. If this were the case, then responses to LWS- and UVS-specific stimuli of defined contrast should be insensitive to changes in the spectral composition of the background light. In fact, we found that this was not the case (S2A–S2C Fig), indicating involvement of the more common opsin co-expressing cones in the chromatic responses of SCN neurons.
By contrast with chromatic SCN neurons, none of the cells identified as achromatic exhibited any overt OFF response to selective activation of either UVS or LWS cone opsin. Instead, these achromatic cells exhibited pure ON responses to stimuli targeting one or both opsin classes, such that on average the population showed little bias towards UVS/LWS opsin-driven responses under background spectra resembling natural daylight (Fig 2C). Adjusting the background spectra to equalise basal activation of the two cone opsins skewed responses in favour of UVS opsin, however (S2D Fig), consistent with previous suggestions that ipRGCs are relatively enriched in the UVS opsin-biased dorsal retina [26].
We next asked whether colour opponent cells also received irradiance information from the melanopsin-expressing ipRGCs that dominate retinal input to the SCN [4,6,7]. To this end, we used changes in spectral composition to selectively modulate melanopsin excitation (see Methods; 14/15 mice above tested with these stimuli). When presented with large steps in melanopsin excitation (92% Michelson contrast) generated in this way, “blue”-ON cells showed slow and sustained increases in firing (Fig 3A; peak response = 3.2 ± 0.8 spikes/s above baseline; paired t test, p<0.01, n = 13), as previously described for melanopsin-driven responses [8,27,28]. The behaviour of the rare “yellow” ON cells to this stimulus was variable (Fig 3B; n = 4), while colour-insensitive cells showed the expected excitatory response (Fig 3A; peak response = 1.8 ± 0.2 spikes/s above baseline; paired t test, p<0.01, n = 23). These data therefore reveal that both chromatic and achromatic cells have access to melanopsin-dependent information about irradiance.
Of note, for the smaller changes in opsin excitation applied above (70% Michelson; ~0.75 log units), we found that the inclusion of melanopsin contrast had little impact on the integrated cellular response. Thus for both chromatic and achromatic populations, responses evoked by spectrally neutral increases in irradiance (“energy”) were very similar in magnitude to those observed where changes in irradiance were restricted to just cone opsins (Fig 3A; subtraction: energy − “brightness”). This was true even for steady-state components of the SCN response (last 1 s of step)—we found no significant difference in responses to two conditions (paired t test; p>0.05 for both blue-ON and achromatic populations). Thus SCN responses to relatively modest changes in light intensity and/or spectral composition are, in fact, dominated by those originating with cones.
How then do chromatic and irradiance responses interact to encode time of day under more natural conditions? To address this question, we produced stimuli that recreated, for Opn1mwR mice, the change in irradiance and colour experienced by wild-type (green cone) mice across the twilight to daylight transition (Fig 4A). We presented these as discrete light steps from darkness, to simulate the challenge in telling time of day faced by a rodent emerging from a subterranean burrow to sample the light environment. Due to their scarcity, we were unable to determine the behaviour of yellow-ON cells under these conditions. However, blue ON cells reliably exhibited a near linear increase in firing rate as a function of simulated solar angle (Fig 4B and 4C; n = 9 from 7 mice), indicating that their sensitivity is well suited to track changes in colour/irradiance occurring across the twilight to daylight transition. Interestingly, the range of solar angles to which these neurons responded was substantially greater than that for achromatic cells recorded in the same set of mice (Fig 4D and 4E; see also S3 Fig; n = 8) indicating that they may be an especially important source of temporal information for the clock around twilight.
To determine the extent to which this ability of blue-ON cells to encode solar angle relied upon their chromatic opponency, we next presented stimuli that recreated the natural change in irradiance over twilight but in which colour was invariant. Two versions of these stimuli were produced, in which colour was fixed either to that at the lowest solar angle for which data was available (“night”) or to that recorded in daylight (“day”; Fig 4A). Whereas achromatic cells were unable to distinguish between these two stimulus sets (Fig 4D and 4E; F-test, p = 0.72), the relationship between solar angle and blue-ON cell firing rate was consistently disrupted under these conditions (Fig 4B and 4C; F-test, p = 0.009; see also S3A Fig). Thus, firing was reliably higher for “night” and lower for “day” conditions than appropriate for that time of day. These effects are consistent with the blue-ON nature of the chromatic units and confirm that these cells employ a combination of colour and irradiance signals in order to encode time of day.
These electrophysiological recordings indicate a significant fraction of neurons in the SCN convey information about changes in spectral composition occurring during natural twilight. We hypothesised, therefore, that by improving the SCN’s ability to estimate solar angle, activation of the colour mechanism would influence the phasing of circadian rhythms under natural conditions. To determine whether this was indeed the case, we scaled up our twilight stimuli to produce an artificial sky that could be presented to freely moving mice over many days in their home cage. We aimed then to compare the phase of circadian rhythms (assayed using body temperature telemetry) under exposure to lighting conditions that recreated natural changes in irradiance across dawn/dusk transitions, with or without the associated alterations in colour (S4 Fig; “irradiance only” twilight replicated “night” spectral composition). To maximise our ability to detect changes in phasing under these conditions, we modelled the temporal profile of these photoperiods on the extended twilight of a northern-latitude summer (S4C Fig). To allow us to readily separate irradiance and colour elements, we undertook these experiments in Opn1mwR mice. Importantly, however, we designed the stimuli to recreate the change in colour across twilight that is experienced by normal, wild-type mice.
We found that the inclusion of colour significantly altered the phase of circadian entrainment. Peak body temperature occurred consistently later when irradiance and colour elements of twilight were included compared to the irradiance signal alone (Fig 5A; 31 ± 8 min; paired t test, p = 0.003; n = 10). This distinction was absent in mice lacking cone phototransduction (Cnga3-/-, [29,30]; Fig 5B; 6 ± 9 min; paired t test, p = 0.51; n = 9), confirming that it originated with cone-dependent colour coding, rather than any differences in the pattern of rod/melanopsin activation between the two photoperiods.
As further confirmation that these differences in body temperature cycles reflected an action on the timing of central clock output, we also monitored SCN firing rate rhythms in a subset of mice via ex vivo multielectrode array recordings. We and others have previously shown that the distribution of daily electrical activity patterns among individual SCN neurons encodes photoperiod duration, resulting in broad phase distributions under summer days [31,32]. Consistent with this work, peak multiunit firing (sampled across small groups of neurons) in the ex vivo SCN of twilight-housed mice was widely distributed across recording epochs corresponding to projected day. Importantly, in line with our body temperature data, this distribution was centred around the middle of the projected day for Opn1mwR mice exposed to “natural” twilight (Fig 6A; n = 124 SCN electrodes from seven slices) but shifted substantially earlier when mice were housed under twilight that lacked changes in colour (Fig 6B; p<0.001, bootstrap percentiles; n = 170 SCN electrodes from six slices). A similarly early phase of peak SCN electrical activity was also observed in slices prepared from Cnga3-/- individuals housed under natural twilight (S5 Fig; p<0.001 versus Opn1mwR, bootstrap percentiles), confirming that the cone-dependent colour signal is indeed required for appropriate biological alignment with twilight. We also found that, across the three groups, Opn1mwR mice exposed to “irradiance-only” twilight exhibited a significantly broader distribution of SCN phasing (Brown-Forsythe test, p = 0.01), suggesting that the inappropriate cone signals under this photoperiod partially impair SCN synchrony.
Here we demonstrate that the mammalian clock has access to information about not just the amount but also the spectral composition of ambient illumination, in the form of a cone-dependent colour opponent input that reports blue–yellow colour. The idea that chromatic signals associated with twilight might provide important cues for circadian photoentrainment has been proposed previously [3,15]. However, the significant technical challenges inherent in distinguishing the influence of changes in colour versus brightness have left the specific role of colour untested, until now.
Our work thus represents the first demonstration that colour-opponent signals influence the circadian clock in any mammalian species. It is clear, from the long history of housing animals under artificial lighting, that colour signals are not necessary for circadian entrainment per se. However, our data indicates that most mammals could use colour [18,33–35] to provide additional information about sun position, above that available from simply measuring irradiance. Our entrainment experiments likely underestimate the importance of that colour signal under field conditions as they lack the daily variation in cloud cover that makes irradiance-alone a less reliable indicator of time of day. Nevertheless, even under these conditions, we find a significant impact of the twilight spectral change on the phasing of entrained rhythms. This reveals that spectral opponency contributes to the most fundamental function of the entrainment mechanism, ensuring correct timing of physiological and behavioural rhythms.
Given the nature of the change in spectral composition, we might expect that it would be available to any species capable of comparing the activity of short with middle/longer wavelengths. Our own subjective experience is that the event is detectable to humans, and a chromatic opponency equivalent to that described here could account for previous reports of subadditivity for polychromatic illumination in human melatonin suppression [36,37]. It is also noteworthy that the majority of mammalian species have retained the short and mid-/long-wavelength cone opsins required to detect changes in spectral composition associated with twilight (for a detailed discussion of the exceptions to this rule see [17,18]). Similarly, earlier studies have identified the capacity for blue–yellow colour discrimination in the pineal/parietal organs of a number of non-mammalian vertebrates, including reptiles, amphibians, and fish (for review see [38]). By directly influencing melatonin secretion, chromatic signals are thus presumably also a key component of the neural mechanisms responsible for appropriate alignment of non-mammalian physiology relative to dawn and dusk. Alongside our present data, it appears then that the use of colour as an indicator of time of day is an evolutionarily conserved strategy, perhaps even representing the original purpose of colour vision.
The specific sensory properties of the circadian photoentrainment mechanism in mammals have long remained a subject of debate [2]. SCN neurons are known to receive input from all major classes of retinal photoreceptor [8–13]. However, since “cone-only” mice do not reliably entrain to conventional light–dark cycles, current models posit that photoentrainment is primary driven by a combination of rod and melanopsin inputs [11,12]. By contrast, the proposed role of cones has been to allow the clock to track relatively high frequency changes in light—a signal that does not appear to play much role in circadian entrainment under conditions most commonly employed in the laboratory (but see [12]). Our data thus establish an important new role for cones in photoentrainment, one which would not be apparent under standard laboratory conditions but will act as an essential regulator of biological timing in more natural settings.
Insofar as most retinal input to the clock is provided by ipRGCs [4,6,7] the appearance of colour opponency in this subset of retinal ganglion cells would provide a simple explanation for the chromatic responses of SCN neurons observed here. Colour opponency has not yet been documented in mouse ipRGCs [39], but has been reported in primates [40] (although it is unknown whether any of these cells project to the SCN). Interestingly, the dominant form of spectral opponency we observe here in the mouse SCN (blue-ON/yellow-OFF) is opposite to that reported for primate ipRGCs and, most recently, for chromatic response of the pupil in humans [41]. While this would, by no means, rule out a role for chromatic influences on the human circadian system, it is also currently unclear whether such yellow-ON/blue-OFF responses are a characteristic feature of all primate ipRGCs. Indeed, such behaviour certainly appears inconsistent with the sensory properties of human melatonin regulation, which seems to exhibit a short- rather than long-wavelength bias [42].
Of course, alternative possibilities to that outlined above are that colour information reaches SCN neurons via the small number of non-melanopsin-expressing RGC inputs or is generated by a mechanism distinct from the conventional retinal colour processing circuitry. Previous work indicates that asymmetries in the gradient of cone opsin expression in the mouse retina could impose an indirect form of chromatic bias for stimuli larger than the cells’ receptive field [43]. Alternatively, opponent responses in the SCN may be generated centrally, e.g., via local processing or indirect visual input from the intergeniculate leaflet. Indeed, based on our identification of a rare yellow-ON cell exhibiting inhibitory responses to melanopsin contrast, we speculate that central processing could contribute to at least some of the responses reported here.
Regardless of their biological origin, chromatic signals provide the SCN with additional information about solar angle, above that available from measuring brightness alone, allowing the clock to appropriately time its output under natural photoperiods. Based on the widespread capacity for colour vision among mammals (and the previous identification of colour opponent ipRGCs in primates), we suggest related mechanisms are likely to be broadly applicable across many mammalian species.
All animal use was in accordance with the Animals (Scientific Procedures) Act of 1986 (United Kingdom). Electrophysiological experiments were performed under urethane anaesthesia; other procedures were conducted under isfluorane anaesthesia. Unless otherwise stated, animals used in this study (homozygous Opn1mwR and Cnga3-/- mice) were housed under a 12-h dark/light cycle at a temperature of 22°C with food and water available ad libitum.
Spectral irradiance measurements (280–700 nm, 0.5 nm bins) were collected in Manchester, UK (Lat.: 53.47, Long.: -2.23, Elevation 76 m) every minute across the solar cycle using a METCON diode array spectroradiometer contained within a temperature stabilised weatherproof housing. The global entrance optics was levelled and mounted at a rooftop monitoring site, providing a horizon relatively clear of obstructions, the entrance optics being connected to the spectrometer by way of a 600 μm diameter 5 m long optical fibre. Instrument calibrations were carried out with reference to spectral irradiance standards, traceable to NIST (National Institute of Standards and Technology, United States). Instrument dark counts were observed to be spectrally flat and were removed by subtracting the mean value for wavelengths <290 nm (where no ground-level solar signal is present).
Data analysed were spectral irradiance measurements collected between 31 August and 14 October 2005 (41 d). Due to gaps in the data collection record, we were able to extract from these 71 complete dawn/dusk transitions (from 36 d). No attempt was made to select data on the basis of weather condition although the period was broadly representative in comparison to relevant climatological averages.
For each twilight transition, we first calculated the average spectral irradiance profile as a function of solar angle relative to the Horizon (0.5° bins, 2–5 measurements/bin). We restricted this analysis to solar angles greater than 7° below the horizon, since our detector was specifically optimised to obtain measurements across light intensities encountered through civil twilight to daytime (making night-time measurements less reliable). We next converted these spectral irradiance profiles into effective photon fluxes as experienced by mouse opsin proteins, using established and validated procedures [19,20,22] based on Govardovski visual pigment templates [44] and published values for mouse lens transmission [45]. Calculations presented in the manuscript were based on the following peak sensitivity (λmax): UVS cone opsin-365 nm, Melanopsin-480 nm, Rhodopsin-498 nm, MWS cone opsin 511 nm.
The resulting series” of photon flux versus solar angle values for each opsin were then analysed individually or in combination (additive or as ratios). Specifically, we calculated the percentage of variance for the dataset in question that was explained by sun position, using the following calculation (with N representing the total number of data points, K the number of dawn/dusk observations and P the number of solar angle bins):
Var|θ=100K∑h=1P(X¯h−X¯)2∑i=1N(Xi−X¯)2
Since there was no apparent difference in photon flux versus solar angle profiles obtained during dawn or dusk transitions, we pooled these data for the above analysis, treating each as an independent observation.
For comparisons of colour versus irradiance based estimates of solar angle (Fig 1D and associated text), irradiance was defined as effective photon flux at UVS+MWS cone opsins. Values obtained using other mouse opsins (singly or in combination) produced essentially identical results. For the aforementioned comparisons, estimates of mean and standard deviation for Var|θ were obtained based on bootstrap replicates (every possible combination of 69 out of the total 71 dawn/dusk observations). Similar analysis to those described above, but performed using only observations taken at either dawn or dusk also produced essentially identical results. Calculations of ““blue–yellow”“colour index (Fig 1C and 1D) were based on the ratio of MWS:SWS cone opsin activation ([MWS+LWS]/SWS for human visual system).
Urethane (1.55 g/kg) anaesthetised adult (60–120 d) male Opn1mwR mice were prepared for stereotaxic surgery as previously described [8]. Recording probes (Buszaki 32L; Neuronexus, MI, US) consisting of four shanks (spaced 200 μm), each with eight closely spaced recordings sites in diamond formation (intersite distance 20–34 μm) were coated with fluorescent dye (CM-DiI; Invitrogen, Paisley, UK) and then inserted into the brain 1 mm lateral and 0.4 mm caudal to bregma at an angle of 9° relative to the dorsal-ventral axis. Electrodes were then lowered to the level of the SCN using a fluid-filled micromanipulator (MO-10, Narishige International Ltd., London, UK).
After allowing 30 min for neural activity to stabilise following probe insertion, wideband neural signals were acquired using a Recorder64 system (Plexon, TX, US), amplified (x3000) and digitized at 40 kHz. Action potentials were discriminated from these signals offline as “virtual”-tetrode waveforms using custom MATLAB (The Mathworks Inc., MA, US) scripts and sorted manually using commercial principle components based software (Offline sorter, Plexon, TX, US) as described previously [46].
Surgical procedures were completed 1–2 h before the end of the home cage light phase, such that electrophysiological recordings spanned the late projected day-early projected night, an epoch when the SCN light response is most sensitive. Cells were initially characterised as light responsive on the basis of responses to bright mono and polychromatic light steps (10–30 s dur.; intensity >1014 photons/cm2/s). Once visual responsiveness was confirmed, experimental stimuli were applied as described below. Following the experiment, accurate electrode placement was confirmed histologically as described previously [8]. Projected anatomical locations of light response units reported in this study are presented in S6 Fig.
All visual stimuli were delivered in a darkened chamber from a custom built source (Cairn Research Ltd, Kent, UK) consisting of independently controlled UV, blue and amber LEDs (λmax: 365, 460, and 600 nm respectively). Light was combined by a series of dichroic mirrors and focused onto a 5 mm diameter piece of opal diffusing glass (Edmund Optics Inc., York, UK) positioned <1 mm from the eye (contralateral to the recording probe for SCN recordings). LED intensity was controlled by a PC running LabView 8.6 (National instruments).
Light measurements were performed using a calibrated spectroradiometer (Bentham instruments, Reading, UK). LED intensity was initially calibrated (using the principles described above) to recreate for Opn1mwR individuals the effective rod, cone and melanopsin excitation experienced by a wild-type (green cone) mouse visual system under typical natural daylight (average values from our environmental data above at a solar angle 3° above the horizon; S1A Fig). We also carefully calibrated differential modulations in the intensity of each LED to produce stimuli that independently varied in apparent brightness for one or both cone opsin classes (either in unison or antiphase) with no apparent change in rod or melanopsin excitation (S1B Fig). In each case, brightness for the stimulated opsin was varied by ±70%, to produce an overall 4.7-fold increase in intensity of between “bright” and “dim” phases of the stimulus. Transitions between the two stimulus phases occurred smoothly over 50ms (half sinusoid profile). We also applied stimuli that selectively modulated melanopsin excitation (±92%), without changing effective cone excitation. These later also, in principle, modulated apparent brightness for rod photoreceptors (±84%), however we think a rod contribution to the resulting responses unlikely owing to the high background light levels (14.9 rod effective photons/cm2/s) and our previous work suggesting that rods have little influence on acute electrophysiological light responses in the SCN [8]. Indeed, similar stimuli evoke very little response in the lateral geniculate nuclei of melanopsin knockout animals [22].
In a subset of experiments (7/15) we also applied a second set of stimuli designed to recreate various stages of twilight, using our calculations of the effective photon fluxes experienced by mouse opsins at solar angles between -7 and 3° relative to the horizon. These were applied as light steps (30 s) from darkness in random sequence with an interstimulus interval of 2 min. To confirm whether elements of the resulting responses were dependent on spectral composition, these stimuli were interspersed with two additional stimulus sets which were identical except that irradiance for the UVS opsin was fixed at a constant ratio relative to LWS (mimicking either day or night spectral composition).
For behavioural experiments, we used similar principles to generate photoperiods that smoothly recreated our measured changes in twilight illumination, with (“natural”) or without the associated change in spectral composition (irradiance-only: spectra fixed to mimic “night”). Stimuli were generated by an array of three violet (400 nm) and three amber (590 nm) high-power LEDs (LED Engin Inc., San Jose CA, US) placed behind a polypropylene diffusing screen covering the top of the cage. The combination of multiple LEDs allowed a larger range of brightness (from dark up to approximately 25 W/m2 for the violet and 10 W/m2 for the amber). Intensity of each LED was independently controlled by a voltage controlled driver (Thorlabs Inc., Newton NJ, US). The light intensity modulation signals were provided by a PC running Labview through a voltage output module (National Instruments), and followed a temporal profile that recreated the sun’s progression during a northern latitude summer (calculations based on Stockholm, Sweden; Lat: 59, Long: 18, Elevation 76 m, 20 June 2013; total twilight duration = 2.3 h).
To determine the impact of twilight spectral changes on mouse entrainment, female Opn1mwR and Cnga3-/- mice (housed under an 18:6 light–dark [LD] cycle) were first implanted with iButton temperature loggers (Maxim, DS1922L-F5#). To reduce weight and size, these were dehoused and encapsulated in a 20% Poly(ethylene-co-vinyl acetate) and 80% paraffin mixture as described by Lovegrove [47]. For implantation, mice were anaesthetised with isoflurane (1%–5% in O2) and the temperature logger implanted into the peritoneal cavity. Following surgery, animals were given a 0.03 mg/kg subcutaneous dose of buprenorphine and allowed to recover for at least 9 d in 18:6 LD before the start of the experiment. The timing of lights off under this cycle was designated as Zeitgeber time (ZT) 12 and the timing of experimental photoperiods were set to align their midnight (ZT15) with this square wave LD cycle.
Following recovery, group housed mice (five per cage) were transferred to the natural twilight photoperiod. The cage environment contained an opaque plastic hide, allowing the animals to choose their own light sampling regime. After 14 d, mice were then returned to 18:6 LD for a further 14 d and finally transferred to the “irradiance-only” twilight photoperiod.
At the end of the experiment, mice were culled by cervical dislocation and temperature loggers recovered. Temperature data (recorded in 30 min time bins) was processed by upsampling to 5 min resolution (cubic spline interpolation), Gaussian smoothing (SD = 45 min), and normalisation as a fraction of daily temperature range. Phase of entrainment was estimated as the timing of peak body temperature from that individual’s daily average profile (calculated from the last 9 d in each photoperiod).
Opn1mwR and Cnga3-/- mice were housed under twilight stimuli of either “natural” or “night” composition (as described above) for at least 14 d prior to experiments. Mice were removed from the home cage 30–60 min after the end of the dawn transition (~ZT19) and culled by cervical dislocation followed by decapitation. The brain was then rapidly removed, mounted onto a metal stage and cut using a 7000 smz vibrating microtome (Campden Instrument, UK) in ice-cold (~4°C) sucrose-based slicing solution composed of (in mM): sucrose (189); D-glucose (10); NaHCO3 (26); KCl (3); MgSO4 (5); CaCl2 (0.1); NaH2PO4 (1.25); oxygenated with 95% O2/5% CO2 mixture. Coronal brain slices containing the SCN (350 μm) were then immediately transferred into a petri dish containing oxygenated artificial cerebrospinal fluid (aCSF) composed of (in mM): NaCl (124); KCl (3); NaHCO3 (24); NaH2PO4 (1.25); MgSO4 (1); glucose (10); CaCl2 (2); slices were then left to rest at room temperature (22 ± 1°C).
Approximately 30 min after slice preparation, slices were placed, recording side down, onto 6x10 perforated multielectrode arrays (pMEAs; Multichannel Systems, MCS, Germany). Slices were visualised under the microscope and photos were taken with a GXCAM-1.3 camera (GX Optical, UK) in order to confirm appropriate slice placement over pMEA electrode sites. Slices were held in place by both the suction via the MEA perforations and a harp slice grid (ALA Scientific Instruments Inc., US). The pMEA recording chamber was continuously perfused with pre-warmed oxygenated aCSF (34 ± 1°C) to both slice surfaces at a rate of 2.5–3 ml/min. Neural signals were acquired as time-stamped action potential waveforms using a USB-ME64 system and a MEA1060UP-BC amplifier (MCS, Germany). Signals were sampled at 12.5 kHz, High pass filtered at 200 Hz (second order Butterworth) with a threshold of usually at -16.5 μV. Recordings were maintained for a total duration of 30 h.
At the end of each recording, slices were treated with bath applications of 20μM NMDA to confirm maintained cell responsiveness, followed by 1 μM TTX to confirm acquired signals exclusively reflected Na+-dependent action potentials. All drugs were purchased from Tocris (UK), kept as stock solutions at -20°C (dissolved in dH2O), and were diluted to their respective final concentrations directly in pre-warmed, oxygenated aCSF; all drugs were bath applied for 5 min.
Multiunit action potential firing rates detected at electrodes located within the SCN region were then selected for further analysis. Data were subsequently binned (60 s) and smoothed via boxcar averaging (width: 2 h) to determine the timing of peak activity. Channels where peak firing did not decay by >50% within ±12 h, or where peak firing was less than 0.2 spikes/s were excluded from this analysis, such that on average 22 ± 2 SCN electrodes were analysed for each experiment. Based on peak firing rates observed (mean ± SEM: 6.9 ± 0.5 spikes/s) we estimate these typically represent recordings from less than four neurons.
To assess for significant differences in the timing of population activity under our different experimental conditions, we drew 1,000 samples of 100 randomly selected neurons from each condition (Opn1mwR “natural”, Opn1mwR “night”, Cnga3-/- “natural”). By calculating the circular mean phase for each sample, we thus obtained estimates of the probability that the observed population means differed by chance.
|
10.1371/journal.pgen.1005352 | Genome-Wide Association and Trans-ethnic Meta-Analysis for Advanced Diabetic Kidney Disease: Family Investigation of Nephropathy and Diabetes (FIND) | Diabetic kidney disease (DKD) is the most common etiology of chronic kidney disease (CKD) in the industrialized world and accounts for much of the excess mortality in patients with diabetes mellitus. Approximately 45% of U.S. patients with incident end-stage kidney disease (ESKD) have DKD. Independent of glycemic control, DKD aggregates in families and has higher incidence rates in African, Mexican, and American Indian ancestral groups relative to European populations. The Family Investigation of Nephropathy and Diabetes (FIND) performed a genome-wide association study (GWAS) contrasting 6,197 unrelated individuals with advanced DKD with healthy and diabetic individuals lacking nephropathy of European American, African American, Mexican American, or American Indian ancestry. A large-scale replication and trans-ethnic meta-analysis included 7,539 additional European American, African American and American Indian DKD cases and non-nephropathy controls. Within ethnic group meta-analysis of discovery GWAS and replication set results identified genome-wide significant evidence for association between DKD and rs12523822 on chromosome 6q25.2 in American Indians (P = 5.74x10-9). The strongest signal of association in the trans-ethnic meta-analysis was with a SNP in strong linkage disequilibrium with rs12523822 (rs955333; P = 1.31x10-8), with directionally consistent results across ethnic groups. These 6q25.2 SNPs are located between the SCAF8 and CNKSR3 genes, a region with DKD relevant changes in gene expression and an eQTL with IPCEF1, a gene co-translated with CNKSR3. Several other SNPs demonstrated suggestive evidence of association with DKD, within and across populations. These data identify a novel DKD susceptibility locus with consistent directions of effect across diverse ancestral groups and provide insight into the genetic architecture of DKD.
| Type 2 diabetes is the most common cause of severe kidney disease worldwide and diabetic kidney disease (DKD) associates with premature death. Individuals of non-European ancestry have the highest burden of type 2 DKD; hence understanding the causes of DKD remains critical to reducing health disparities. Family studies demonstrate that genes regulate the onset and progression of DKD; however, identifying these genes has proven to be challenging. The Family Investigation of Diabetes and Nephropathy consortium (FIND) recruited a large multi-ethnic collection of individuals with type 2 diabetes with and without kidney disease in order to detect genes associated with DKD. FIND discovered and replicated a DKD-associated genetic locus on human chromosome 6q25.2 (rs955333) between the SCAF8 and CNKSR genes. Findings were supported by significantly different expression of genes in this region from kidney tissue of subjects with, versus without DKD. The present findings identify a novel kidney disease susceptibility locus in individuals with type 2 diabetes which is consistent across subjects of differing ancestries. In addition, FIND results provide a rich catalogue of genetic variation in DKD patients for future research on the genetic architecture regulating this common and devastating disease.
| Diabetic kidney disease (DKD) is a devastating complication in patients with diabetes mellitus (DM) and is associated with high risk for cardiovascular disease and death.[1,2] DKD is the leading cause of end-stage kidney disease (ESKD) requiring renal replacement therapy in developed nations; these procedures incur high healthcare costs with great personal, family and societal burden.[3] The prevalence of DKD continues to rise in the United States in proportion to the growing prevalence of DM. Unfortunately, intensification of glycemic, lipid and blood pressure control have not dramatically impacted the prevalence of DKD.[3,4] Hyperglycemia alone is insufficient to cause DKD. Genetic factors appear critical in its pathogenesis based upon variable incidence rates of DKD between population groups, aggregation of DKD-associated ESKD in families, and the highly heritable nature of diabetic renal histologic changes, estimated glomerular filtration rate (eGFR) and proteinuria.[5]
Genome-wide association studies (GWAS) have identified multiple loci for kidney function and chronic kidney disease (CKD) in population- and community-based cohorts, primarily of European ancestry.[6–10] However, CKD phenotypes in many studies included minimally to moderately reduced eGFR, not fully reflective of the progressive forms of CKD seen in kidney disease clinics. In early reports, published GWAS signals for DKD were equivocal, confounded by small sample sizes and failure to consistently replicate. Recently, the GEnetics of Nephropathy: an International Effort (GENIE) consortium identified genome-wide significant, replicated signals in a meta-analysis of over 12,000 type 1 (T1) DM patients with DKD of European ancestry.[9] Type 2 (T2) DM is far more prevalent than T1DM, accounting for 90% of cases worldwide and for the majority of prevalent cases of DKD. Relative to European Americans (EAs) with T2DM, African American (AA), American Indian (AI), and Mexican American (MA) patients with T2DM are disproportionately affected by severe DKD,[3] yet under-represented in genetic analyses. Defining the underlying genetic architecture responsible for advanced T2DM-associated kidney disease in multiple populations could provide critical insights into pathogenesis and identify new molecular targets for therapy. We report the results of a GWAS in AA, EA, MA, and AI patients with DKD enrolled in the National Institute of Diabetes and Digestive and Kidney Diseases (NIDDK)-sponsored “Family Investigation of Nephropathy and Diabetes” (FIND) [11] and the corresponding large replication study and trans-ethnic meta-analysis.
Demographic characteristics of the Discovery and Find Large Replication study (FILR) samples that met FIND phenotype qualifications and genotype quality control (QC) are summarized in Table 1 and S1 Table. The proportion of females, age at ESKD or enrollment, hemoglobin A1c and proportion with diabetic retinopathy (DR) varied by ancestry but was generally comparable between the Discovery and FILR samples within a specific ethnic population. Phenotypic differences among these populations and their genetic and DKD prevalence differences motivated the meta-analysis approach.
The principal component (PC) analysis identified PCs that genetically partitioned the Discovery sample into ancestry groups consistent with self-report. S1 Fig displays the two-dimensional partitioning via PC analysis with the boundaries for inclusion into the GWAS analysis. The logistic regression model that included the PCs as covariates reduced the inflation factor to nominal levels and combined with the P-P plot show no evidence of a systematic inflation (S2 Fig). In the replication, the inflation factor was λ = 1.05 using 278 AIMs. If we scale to 1,000 cases and 1,000 controls this would be λ1000 = 1.017, an appropriate inflation factor for the replication study. S3 Fig provides a summary of the statistical power analyses for the race-specific discovery, replication analysis, and meta-analyses. These calculations show that for risk-predisposing variants shared across ancestries the study has power >0.50 and 0.80 to detect odds ratios (OR) on the order of 1.06 to 1.15 for a minor allele frequency (MAF) = 0.45 and 1.30 to 1.36 for a MAF = 0.05, respectively. Further, leveraging the differences in linkage disequilibrium (LD) among the four ancestries, the study is powered to potentially reduce the size of the associated region via trans-ethnic mapping.
The only locus that reached genome wide significance for DKD in the trans-ethnic meta-analysis encompassing all FIND Discovery and Replication samples was rs955333 on chromosome 6 (minimum p-value 1.31x10-8 [additive]; minimum p-value 9.02x10-11 [dominant]) (Table 2). Fig 1 contains the Manhattan plot for the meta-analysis across all ancestries included in the Discovery and Replication samples. Consistent directions of association were present in three ethnic groups (only AA samples did not pass QC) and several supporting single nucleotide polymorphisms (SNPs) were detected in the region (regional plot in Fig 2). This SNP lies between the SR-like carboxyl-terminal domain associated factor 8 gene (SCAF8) and the connector enhancer of KSR family of scaffold proteins gene (CNKSR3), suggesting a possible role in transcription regulation. CNKSR3 is a direct mineralocorticoid receptor target gene involved in regulation of the epithelial sodium channel (ENaC) on the apical membrane of cells in the distal nephron.[12] CNKSR3 is highly expressed in the renal cortical collecting duct and upregulated in response to physiologic aldosterone concentrations. ENaC precisely regulates renal sodium absorption and plays important roles in maintenance of plasma volume and blood pressure. Ziera et al. [12] suggested that CNKSR3, a PSD-95/DLG-1/ZO-1 (PDZ) domain containing protein, inhibits the RAS/ERK signaling pathway, stimulating ENaC activity with enhanced renal sodium absorption. More recently, CNKSR3 was shown to function as an aldosterone-induced scaffolding platform that orchestrated assembly of ENaC and its regulators Nedd4-2, Raf-1 and SGK-1 and was essential for stimulation of ENaC function by aldosterone.[13] Clinically, renin-angiotensin-aldosterone system (RAAS) blockade serves as a mainstay of therapy for patients with DKD and other proteinuric kidney diseases.[14,15] Inhibition of aldosterone may further limit renal fibrosis, independent of natriuretic effects.[16,17] Hence, significant association between DKD and markers near CNKSR3 is consistent with clinical trial data demonstrating that blockade of the renin angiotensin system or the aldosterone receptor slows DKD progression. However, further experiments are needed to demonstrate that the associated SNP regulates the pathogenesis of progressive DKD. Further studies will be necessary to assess if the CNKSR3 regulates DKD pathogenesis indirectly by its effects on ENaC activity or directly by promoting aldosterone-dependent fibrosis.
Less is known about the function of SCAF8, also known as RBM16. SCAF8 is a RNA maturation factor recruited to the carboxy-terminal domain of RNA polymerase II in a phosphorylation-dependent manner.[18] It also is a target for ataxia telangiectasia mutated (ATM) kinase, a crucial component of the DNA damage response required for DNA repair and cell cycle control.[19] ATM kinase is associated with responsiveness of patients with DM to the insulin sensitizer metformin in some but not all studies.[20,21] Thus, genes in the region of rs955333 are suggestive of DKD-related pathogenesis.
GWAS loci identify elements that may regulate gene expression, and recent data indicate GWAS associations are located in regions bounded by recombination hot spots near non-coding causal variants, which regulate transcription.[22,23] We next contrasted transcript abundance of the genes within the megabase region centered on rs955333, TIAM2, SCAF8, CNKSR3, IPCEF1 and OPRM1, in DKD and living donor kidney biopsies. DKD biopsies were obtained from European and AI cohorts and were analyzed separately. All five genes show statistically significant differential expression in at least one kidney tissue compartment of one population. SCAF8 steady state mRNA levels show increased expression in DKD compared to living donor biopsies in glomerular and tubulo-interstitial compartments of both populations (S2B Table); TIAM2 and OPRM1 show glomerular-specific differential expression; IPCEF1 is repressed in both tissue compartments of AI subjects; and CNKSR3 is increased in the tubulo-interstitial compartment of AIs (S2B Table). Normalized tubulo-interstitial expression of CNKSR3 correlated with urine albumin (r = 0.78, q = 0.0056) and urine albumin:creatinine ratio (UACR) (r = 0.74, q = 0.0107). In addition, IPCEF1, located downstream of CNKSR3, has been reported to be translated with CNKSR3 as one protein,[24] and has a tubulo-interstitial expression quantitative trait locus (eQTL) (NM_001130699, rs249964, P = 2.34E-04) (S2A Table). LD between this SNP and the sentinel variants in the region significantly associated with DKD in the trans-ethnic (rs955333) and AI association analysis (rs12523822; see below) is negligible (D’ = 0.43, r2 = 0.01 in AI). However, tubulo-interstitial expression of IPCEF1 in kidney tissue from AIs was significantly correlated with the DKD phenotype UACR (r = -0.54, q = 0.031). These studies were limited by the small number of available biopsies the narrow criteria used to define the region of interest (see Methods). As proxies, disease-dependent differential gene expression and the rs249964 eQTL demonstrate DKD regulatory activity in the locus. Significant results of eQTL and differential gene expression analyses for other loci in Table 2 are also sown in S2A Table and S2B Table, respectively.
No SNP reached genome-wide significance (P<5x10-8) in the AA GWAS; however, a number provided suggestive evidence for association with DKD (Table 3; S5A Table and S6A Table summarize the top 200 SNP associations in the discovery GWAS and replication study, respectively). The strongest associations were found within the apolipoprotein L1 (APOL1) and non-muscle heavy chain 9 gene (MYH9) region on 22q (Table 2, Discovery + FILR meta-analysis: rs5750250, P = 7.7x10-8; rs136161, P = 5.23x10-7). Since G1 and G2 variants of APOL1 are strongly associated with non-diabetic nephropathy in AA patients,[25–27] the G1/G2 compound risk was modeled under a recessive genetic model and these variants accounted for the associations on 22q in Table 2 (rs5750250 P = 7.70x10-8, OR = 1.27; rs136161 P = 5.23x10-7, OR = 1.36). Association with G1/G2 within APOL1 likely exists due to inclusion of non-FIND AA cases with coincident DM and unrecognized non-diabetic kidney disease.[28] APOL1 was not associated with T2D-ESKD in a logistic regression analysis adjusting for age, gender and global ancestry restricted to FIND MALD and CHOICE (Choices for Healthy Outcomes In Caring for End-stage renal disease) study cases meeting the original FIND DKD case definition (rs73885319 P = 0.1098; rs71785313 P = 0.1182).[29]
Regions beyond 22q provided suggestive evidence of association in the AA Discovery + FILR meta-analysis including rs1298908 on 10q22 (OR = 1.36, P = 8.83x10-7) between MAT1A and ANXA11, in a region dense with regulatory elements and transcription factors. There was also an association on 3p26 (rs304029, OR = 1.26 P = 1.10x10-6) within inositol 1,4,5-trisphosphate receptor, type 1 (ITPR1), a gene involved in cerebellar and autoimmune disorders but not renal involvement.[30] The genes in these other candidate regions (ANXA11, MAT1A and ITPR1) also show statistically significant differential expression in at least one population and compartment; as do IGSF22 near candidate rs11766496 on chromosome 11, and TNFRSF19 near rs95107795 on chromosome 13. Other top AA associated regions in Table 2 do not have clear connections to kidney disease. Since APOL1 association likely reflected inclusion of non-FIND cases with non-diabetic nephropathy, a GWAS was re-computed within AAs in the discovery sample, which only included subjects lacking two APOL1 risk variants. The top 200 associations from this GWAS are summarized in S7 Table. The correlation between the–log10 (p-value) for GWAS with and with AA subjects with and without two APOL1 risk variants is r = 0.82 (S4 Fig). The top association in this subset GWAS was rs2780902 on 1p31 (OR = 0.52, P = 2.98x10-7) within Janus kinase 1 (JAK1), a member of the protein-tyrosine kinases.[31] The ENCODE data shows that this SNP resides within a region with numerous transcription factors and DNase I hypersensitivity sites. JAK1 is a widely expressed membrane associated phosphoprotein and is involved in interferon transduction pathway. This kinase links cytokine ligand binding to tyrosine phosphorylation of various known signaling proteins and the signal transducers and activators of transcription (STATs). Another interesting association among the top 10 associations is rs2596230 on 15q14 (OR = 1.56, P = 9.36x10-6) within ryanodine receptor 3 (RYR3).[32] The protein encoded by RYR3 functions to release calcium from intercellular storage in many cellular processes and the gene is expressed in the kidney. The closely related gene, RYR2, is associated with albuminuria.[33] Our prior analyses of transcript expression in DKD biopsies provide additional support for the associations. Both JAK1 and RYR3 (and RYR2) show differential expression that is restricted to the European subjects with Stage III and Stage IV CKD. JAK1 expression is increased in DKD in both compartments, while RYR3 and RYR2 are depressed in the glomerulus.[34] We also recomputed the genome wide discovery and trans-ethnic meta-analysis removing AA subjects with APOL1. The top 200 associations are summarized in S8 Table.
Several regions provided evidence of association with DKD in AIs (Table 3; S5B Table and S6B Table summarize the top 200 SNP associations in the discovery GWAS and replication study, respectively). The strongest association was with rs12523822 on 6q.25 in the SCAF8-CNKSR3 gene region (OR = 0.57, P = 5.74x10-9). This SNP is in strong LD with rs955333, the top hit in the trans-ethnic meta-analysis (r2 = 0.96 in AI unrelated controls); S5 Fig graphically illustrates the extended linkage disequilibrium in this region in all but the AA samples. The A allele at rs955333 is the ancestral allele and confers susceptibility to DKD (as the G allele has OR<1 in Table 2); the A allele has a frequency of 0.76 in the American Indian samples and 0.85 in European American samples, but is nearly monomorphic in African American samples. The allele frequencies are very similar in population-based samples: 0.85 in HapMap CEU, 1.00 in YRI, 0.77 in MEX and 0.76 in full-heritage American Indians from the southwestern United States (R Hanson, personal communication). Thus, the high risk allele at this locus does not appear to be Amerindian specific. The p-value for association in European Americans is 0.0013 and 1.3x10-6 in American Indian, suggesting that the signal does not come entirely from American Indians samples. Further fine-mapping or sequencing will be necessary to fully characterize the association signal within and across ethnic groups. Another association that approached genome-wide significance was rs13254600 (OR = 0.58, P = 5.54x10-8) on 8q24 within WD repeat domain 67 (WDR67). This gene is expressed in a wide variety of tissues, including kidney, and may affect cellular membrane functions by regulating Rab GTPase activity.[35] TBC1D31 (WDR67) mRNA is increased in both compartments of kidney tissue from AIs, but only in the glomerulus for European subjects with more advanced DKD.
Another SNP of interest is rs10019835 (OR = 0.70, 5.47x10-7) on 4q32 within guanylate cyclase 1, soluble, alpha 3 (GUCY1A3); the protein encoded by GUCY1A3 serves as a receptor for nitric oxide,[36] which through its role in endothelial function may be a mediator of DKD.[37] GUCY1A3 is differentially expressed in both tissue compartments and both DKD biopsy cohorts, and shows one of the strongest differences of all genes in candidate regions (especially among the European subjects who have more advanced DKD) (S3A Table and S3B Table; S6 Fig). In addition, the candidate SNP rs10019835 has a tubulo-interstitial specific eQTL with the full-length isoform of GUCY1A3 (NM_000856, P = 4.97x10-4). The shortest isoform of the gene (NM_001130687) has a glomerular eQTL with rs12504357 (P = 2.63x10-5), an intronic SNP that is 5kb upstream of the associated variant. These two eQTL SNPs have D’ = 1 in some populations, likely reflecting low allele frequencies in the reference populations. Integrin alpha 6 (ITGA6, rs13421350, 2q31, OR = 0.58, P = 5.54x10-8) is involved in cell adhesion and is expressed in the kidney. The gene shows negative differential expression in Europeans with DKD, and it has both glomerular and tubulo-interstitial eQTL. The glomerular eQTL is with the SNP rs6758468 (P = 5.41x10-4), which is 143kb from the candidate; while the tubulo-interstitial eQTL is with rs12469788 (P = 3.26x10-4), which is 5kb from the candidate with D’ = 1, but negligible r2. Finally, rs10952362 on 7q36 near XRCC2 (rs10952362, OR = 1.91, P = 7.99x10-8), a gene involved in DNA repair was strongly associated with DKD.[38] We find that XRCC2 is repressed in the tubulo-interstitial kidney tissue from AIs.
EA subjects comprised the smallest group within FIND and power to detect variants associated with DKD was limited (S3 Fig). None of the associations in the EA Discovery + FILR meta-analysis had a p-value <10−5 (Table 3; S5C Table and S6C Table summarize the top 200 SNP associations in the discovery GWAS and replication study, respectively).
Several suggestive associations were identified in the MA Discovery GWAS (Table 3; S5D Table summarizes the top 200 SNP associations in the GWAS). No replication cohort was available to be genotyped in FILR, so only the Discovery GWAS and trans-ethnic meta-analysis are reported (Tables 2 and 3). The strongest association was on 12q24 for rs7975752, located ~242 kb downstream of the mediator complex subunit 13-like (MED13L) gene (OR = 1.76, P = 1.67 x 10−6). MED13L functions as a transcriptional coactivator for RNA polymerase II-transcribed genes. While its functional significance in DKD is unclear, gene variants 4 Mb downstream (rs614226) and upstream (rs653178) on 12q24 show genome-wide significant association with ESKD [9] and CKD [39] in Europeans. We see that MED13L is repressed in both compartments in kidney tissue from AIs but only in the glomerular transcriptome in the European subjects. Association was observed between DKD and rs731565 (P = 4.06 x 10−6) residing within an intronic region of the contactin-associated protein-like 2 (CNTNAP2) gene on 7q36. SNP rs7805747, approximately 4 Mb downstream from rs731565 has been associated with CKD in European populations [39] Finally, rs4849965, 1.2 Mb upstream of the SRY-related HMG-box 11 (SOX11) gene on 2p25.2 trended toward association with DKD (OR 1.50, 95% CI 1.26–1.79; P = 6.18x10-6) and has previously been associated with CKD in Europeans.[39] We find that absolute tubulo-interstitial expression of SOX11 in AIs is correlated with ACR (r = 0.66, q = 0.029).
The current FIND GWAS comprises the largest genetic analysis for severe DKD based upon risk for progression to ESKD in EA and high-risk non-European ethnic groups including AAs, AIs, and MAs. As in other GWAS, results support a role for multiple DKD susceptibility genes, each with weak effects. A number of the SNPs most strongly associated with DKD had additional support from compartment-specific gene expression measures and eQTL analysis obtained in European and American Indian populations. A novel chromosome 6q25.2 DKD locus was identified in AI samples; SNPs in this region had genome-wide significant association and consistent directions of effect in the meta-analysis across all ethnic groups. Independent support for this region comes from an association with serum creatinine/eGFR in a GWAS in East Asian populations (P = 2.6 x 10−5 at rs4870304) [40]. Strengths of the FIND GWAS were the severe phenotype in cases, focus on DKD in T2D, and inclusion of non-European populations. The 6q25.2 locus requires fine mapping and additional replication in independent sample sets of diabetic subjects with and without DKD that has sufficient power to detect associated, common variants with moderate effect size. Once localized and replicated, functional studies in animal and cell culture models will be necessary to discover the biological mechanisms responsible for the association of DKD with the underlying genetic architecture.
As in other GWAS for complex disease, many previously identified DKD loci were not replicated in the FIND analyses. The inconsistency between our data and published DKD GWAS could reflect that FIND limited the DKD case group to subjects with ESKD and DKD with heavy proteinuria felt to be at high risk for progression to ESKD. FIND did not include microalbuminuric participants as “cases” in the Discovery cohort, choosing instead to focus on advanced nephropathy. However, some microalbuminuric participants with ACR<100 mg/g were included in the replication analysis. Prior GWAS focused on European and Asian DKD populations, often enriched for T1D-associated DKD. Genetic associations may not replicate across other populations; for example, association of APOL1 variants with non-diabetic kidney disease is limited to populations with recent African ancestry. Another possible interpretation is the variants, which regulate DKD pathogenesis, are distinct for T1D and T2D, although a meta-analysis including both T1D and T2D subjects may identify shared loci. Finally, the DKD phenotype in the FIND GWAS relied on standard, stringent clinical criteria for advanced DKD. This approach limited phenotypic heterogeneity but potentially minimized the utility of cross-study comparisons. Although heavy proteinuria is a hallmark of DKD, recent analyses suggest approximately one third of patients with diabetes and an eGFR <60 ml/min per 1.73 m2 had normal urinary protein excretion.[4] This would justify the focus of FIND on advanced DKD. Although not the only DKD phenotype with a genetic component, several investigators recently proposed using ESKD as the optimal DKD phenotype in genetic association studies.[41,42] The availability of bio-samples from patients with advanced DKD is limited. Therefore, entry criteria in the present replication cohorts were loosened to increase sample size; this likely included a small number of participants with non-diabetic CKD (or DKD less likely to progress to ESKD). The AA non-FIND cases used in our replication cohort appear to have included individuals with DM and coincident focal segmental glomerulosclerosis (FSGS), an effect addressed via partitioning based on APOL1 G1 and G2.[28] As in all GWAS, some non-nephropathy controls may develop DKD. This effect would bias results toward the null making it less likely to detect significant association.
FIND was well-powered to detect common risk variants with moderate effect sizes shared across ethnic groups. It was also well powered to use differences in effect sizes to help localize the region of association via transracial mapping. However, it was not powered to detect modest ethnic-specific effects that are not shared with another ethnicity or gene-gene interactions. Thus, these ethnic-specific scans provide important hypothesis generating results for subsequent meta-analyses, pathway enrichment analyses and hypothesis generation.
The FIND was completed in accordance with the principles of the Declaration of Helsinki. Written informed consent was obtained from all participants. The Institutional Review Board at each participating center (Case Western Reserve University, Cleveland, OH, Harbor-University of California Los Angeles Medical Center, Johns Hopkins University, Baltimore, National Institute of Diabetes and Digestive and Kidney Diseases, Phoenix, AZ, University of California, Los Angeles, CA, University of New Mexico, Albuquerque, NM, University of Texas Health Science Center at San Antonio, San Antonio, TX, Wake Forest School of Medicine, Winston-Salem, NC) approved all procedures, and all study subjects provided written informed consent. A certificate of confidentiality was filed at the National Institutes of Health.
See Supplementary Methods (S1 Text).
The DNA samples that comprise the Discovery cohorts, plus an additional 244 blind duplicates were genotyped on the Affymetrix Genome-Wide Human 6.0 SNP array (see S1 Text Supplemental Methods for details). The FILR replication samples were genotyped for 3,937 SNPs selected based on the strength of the statistical association from the Discovery GWAS. Additional SNPs were included based on the FIND eQTL association and candidate gene SNPs previously reported to be associated with DKD (see S1 Text Supplemental Methods for details). Specifically, within each ancestry group, the SNPs with the strongest statistical evidence of association were identified; a few additional SNPs from each region with supportive but weaker evidence of association were also identified (i.e., associations due to LD but r2<0.95 with the primary associated SNP). This redundancy was designed to limit the number of regions not represented in the replication study due to genotyping failure. In total, 3,019 SNPs (821 AA, 790 AI, 608 EA, and 800 MA) were genotyped for FILR based solely on statistical association with DKD within an ethnicity. The trans-ethnic meta-analysis of the discovery cohort identified another 436 SNPs nominally associated with DKD (p<0.0003). In addition, 482 SNPs (121 AA, 133 AI, 122 EA, 14 MA, meta-analysis 92) were chosen with the smallest L2-norm (i.e., Euclidean distance) of the–log10 (p-values) from GWAS and eQTL association analyses, provided that p <0.01 from GWAS. Here, the L2-norm was defined relative to the maximum of the–log10 (p-values) from the GWAS and eQTL and provides an ordering of the combined evidence for eQTL and association with DKD. SNP associations in FILR were considered “replicated” if both the association reached statistical significance and direction of the association was consistent with the Discovery analysis. Finally, 278 AIMs were genotyped to allow for adjustment of potential population substructure. Thus, FILR was designed as a replication study and not a large-scale trans-ethnic fine-mapping study. Subsequent studies will complete fine-mapping to localize associations.
|
10.1371/journal.pntd.0002576 | Cholera Vaccination Campaign Contributes to Improved Knowledge Regarding Cholera and Improved Practice Relevant to Waterborne Disease in Rural Haiti | Haiti's cholera epidemic has been devastating partly due to underlying weak infrastructure and limited clean water and sanitation. A comprehensive approach to cholera control is crucial, yet some have argued that oral cholera vaccination (OCV) might result in reduced hygiene practice among recipients. We evaluated the impact of an OCV campaign on knowledge and health practice in rural Haiti.
We administered baseline surveys on knowledge and practice relevant to cholera and waterborne disease to every 10th household during a census in rural Haiti in February 2012 (N = 811). An OCV campaign occurred from May–June 2012 after which we administered identical surveys to 518 households randomly chosen from the same region in September 2012. We compared responses pre- and post-OCV campaign.
Post-vaccination, there was improved knowledge with significant increase in percentage of respondents with ≥3 correct responses on cholera transmission mechanisms (odds ratio[OR] 1.91; 95% confidence interval[CI] 1.52–2.40), preventive methods (OR 1.83; 95% CI 1.46–2.30), and water treatment modalities (OR 2.75; 95% CI 2.16–3.50). Relative to pre-vaccination, participants were more likely post-OCV to report always treating water (OR 1.62; 95% CI 1.28–2.05). Respondents were also more likely to report hand washing with soap and water >4 times daily post-vaccine (OR 1.30; 95% CI 1.03–1.64). Knowledge of treating water as a cholera prevention measure was associated with practice of always treating water (OR 1.47; 95% CI 1.14–1.89). Post-vaccination, knowledge was associated with frequent hand washing (OR 2.47; 95% CI 1.35–4.51).
An OCV campaign in rural Haiti was associated with significant improvement in cholera knowledge and practices related to waterborne disease. OCV can be part of comprehensive cholera control and reinforce, not detract from, other control efforts in Haiti.
| In October 2010, Haiti experienced a cholera outbreak that is now considered one of the largest cholera epidemics in recent history. A comprehensive approach is necessary to successfully fight the epidemic and proven methods for controlling cholera include improving access to clean water and sanitation as well as widespread hygiene education. In addition, there are two safe cholera vaccines approved for use. The authors conducted surveys before and after a cholera vaccination campaign, that included a public health educational component, in rural Haiti; surveys addressed knowledge of cholera and hygiene practices such as hand washing and water treatment, which are crucial for preventing waterborne diseases such as cholera. The authors found that after the vaccination campaign, knowledge of cholera improved significantly. There was also significant increase in reported hand washing and water treatment post vaccination. Furthermore, there was an association between knowledge and hygiene practices. Therefore, this study demonstrates that cholera vaccination can be a complementary tool in the fight against cholera in Haiti and will not detract from other control efforts.
| In October 2010, a cholera outbreak began in the Artibonite and Centre Departments of Haiti [1]. By December, cholera had been identified in all 10 departments of Haiti and has since reached neighboring countries [2], [3]. Cholera is an acute, watery diarrheal infection caused by the bacterium Vibrio cholerae of the O1 or O139 serogroup; and it can rapidly lead to severe dehydration and death if untreated. However, effective therapy can decrease mortality rate from more than 50% to less than 0.2% [4].
Efforts to control the cholera outbreak have been hampered by weak health systems and lack of clean water and adequate sanitation in Haiti. In 2008, only 17% of Haiti's population used improved sanitation facilities while 12% had access to piped, treated water [5]. In addition, conditions in Haiti further deteriorated on January 12, 2010 when the country suffered a devastating 7.0-magnitude earthquake that killed thousands and rendered approximately 2 million individuals homeless [6]. Pockets of densely populated areas resulting from internal migration after the earthquake likely contributed to an explosive outbreak in Haiti. Rural areas and urban slums were particularly vulnerable to the rapid spread of a waterborne disease such as cholera. Furthermore, Haiti's population had no prior exposure or immunity to V. cholerae [7]. Moreover, analysis of the V. cholerae strain in Haiti revealed a variant strain (serotype Ogawa, biotype El Tor) known to be associated with more severe illness [8], [9]. Between October 2010 and May 2013, there were over 600,000 cases of infection and more than 8,000 cholera deaths reported [10]. In 2011, the cholera epidemic in Haiti accounted for 58% of all cholera cases and 37% of all cholera deaths reported to the World Health Organization (WHO) [11].
A comprehensive approach is necessary to fight the cholera epidemic in Haiti and proven cholera control measures include: active case finding, improving water and sanitation, and widespread hygiene education [12]–[14]. In addition, there are two safe oral cholera vaccines (OCV), approved by the WHO for use in cholera endemic areas [15]. Some have argued that cholera vaccination might detract from other prevention efforts and result in diminished hygiene practices among vaccine recipients [16]–[18]. Yet, there is no evidence indicating that cholera vaccination reduces hygiene practice.
Knowledge, Attitude, and Practice (KAP) surveys have been used in various settings to assess existing knowledge and hygiene practices relevant to prevention and transmission of diarrheal diseases, including cholera [19]–[22]. KAP surveys have also been employed in areas of cholera outbreak to measure uptake of knowledge and behavioral changes in response to educational activities aimed at cholera control [23], [24]. In December 2010, a KAP survey was conducted in resource-limited communities of Port-au-Prince, Haiti to assess the effectiveness of public health campaigns on cholera education [24]. The study showed high knowledge of cholera signs and transmission mechanisms as well as improvement in water treatment practices after the outbreak. However, there have been no studies evaluating the effect of vaccination campaigns for waterborne, diarrheal diseases on knowledge and practices related to these diseases.
We aimed to assess the impact of an OCV campaign on knowledge of cholera and health practice related to waterborne illness in rural Haiti. We hypothesized that the campaign, which had been implemented with an educational component, would lead to improved knowledge and behavior critical for cholera control and therefore had served to bolster efforts in the fight against cholera in Haiti.
Ethics Statement: Data were collected as part of a public health campaign; therefore informed consent was not required from survey respondents. Institutional Review Board approval was received from Partners Healthcare for post-hoc analysis of the de-identified dataset.
We analyzed data from the rural 5th section of St. Marc, also known as Bocozel (Figure S1), in the Artibonite Department of Haiti, where between May and June 2012, the non-profit organization, Partners In Health, carried out a pilot OCV campaign in support of the Haitian Ministry of Health [25]. In February 2012, prior to vaccine implementation, a census was undertaken in Bocozel, resulting in enumeration of 9,517 households. Empty households were visited twice, and if neighboring households could not provide information to confirm that a third visit was warranted, the household was not counted in the census. During the census, every 10th household was invited to participate in a baseline survey on knowledge and practices regarding cholera and waterborne disease.
The survey gathered information on sociodemographic characteristics; knowledge about means of cholera transmission, preventive measures, and water treatment modalities; practices related to frequency of water treatment and hand washing; type of toilet access; and source of drinking water. Knowledge questions prompted respondents to provide as many answers as they could to the following questions: “How can a person get cholera?” “What can you do to avoid getting cholera?” and “What are the methods of treating water that you drink?” Examples of appropriate responses for cholera transmission mechanisms included: “drinking untreated water,” “eating uncooked food,” and “dirty hands.” For cholera prevention methods, suitable answers included: “treat water,” “eat cooked food, and “wash hands.” For hygiene practices, respondents were asked to choose the option that described their frequency of water treatment among: “always,” “almost always,” “often,” “sometimes,” and “almost never.” Respondents were also asked to report the number of times they washed their hands with soap and water daily. Knowledge questions were directed to the individual responding, and practice questions were related to the household. Trained enumerators (locally recruited Haitians who had completed high school) administered surveys to one adult individual (male or female, ≥18 years) identified by members of the household as the head or, in the absence of head of household, a representative of the household. Enumerators received a 2-day training on the use of hardware and software used for data collection as well as the survey modules. Refresher trainings were conducted prior to the administration of each vaccine dose.
The OCV campaign was executed in 2 phases with individuals aged 10 years and above targeted in the first phase, and children between the ages of 1 and 10 years targeted in the second phase. The campaign is described in detail elsewhere [25]. Prior to the campaign, meetings with key stakeholders, community focus groups, and Ministry of Health representatives led to the generation of key messages about cholera prevention and cholera vaccine that were used as part of the vaccination campaign (Table S1). Before and throughout the period of vaccination, educational information was disseminated verbally via radio shows, sound trucks, town criers, local television and was printed on T-shirts and posters. Members of the vaccination team were encouraged to share education messages at every contact with the public. These messages were also communicated by enumerators to household members in the census, after all data collection was complete. Education information was thus provided directly to at least one representative of all enumerated households. All vaccine recipients received the same information during vaccination days, and the entire community received information during the period of the campaign. Printed educational information was not a major focus of the campaign because of low literacy rates in the region.
In September 2012, after the vaccine campaign, a follow-up survey was conducted to estimate vaccination coverage, and as a secondary objective, to evaluate knowledge and practice about cholera. De-identification of pre-vaccine survey data precluded resurveying the same participants; therefore, a list of 600 households was randomly generated from the 9,517 households enumerated during the census using a random number generator in Microsoft Excel. The same survey tool used in the pre-vaccination phase was administered to these households in addition to questions about receipt of cholera vaccine. The same enumerators collected census data and conducted both surveys with the exception of a few staff who were not available at the second time point.
We analyzed results from both surveys using Statistical Analysis System (SAS 9.3). Chi-square and Wilcoxon rank-sum tests were used to compare knowledge and practice variables from the pre- and post-vaccination surveys. We used multivariable logistic regression analysis to (1) evaluate changes in knowledge of cholera prevention and transmission and hygiene practices after the vaccine campaign; (2) examine whether proxies for socioeconomic status (i.e. ever having attended school and access to electricity at home) were associated with these outcomes; and (3) assess whether cholera knowledge was associated with hygiene practices. Multivariable models included a variable for survey (1 versus 2), ever having attended school, and electricity access in the home. To assess for confounding, we first identified baseline variables that were differentially distributed between the two surveys and were associated with any outcome at a p-value≤0.05. These variables (farming occupation, latrine, open defecation) were then included in the multivariable models and those that altered the effect estimate for the survey variable by >10% were retained in the final model.
A total of 811 households from 53 different localities were surveyed pre-vaccination (Survey1), and 518 households from 47 localities were interviewed post-vaccination (Survey2). Eighty-two of the 600 households randomly selected to complete Survey2 (13.7%) were not interviewed: 43 households had been destroyed or no longer existed, 12 households were empty despite two visit attempts, and 1 household resident was deceased. The remaining 26 households were either not accessible because of challenges presented by the rainy season or they could not be physically located based on the information in the census. Because there were few official addresses in this area, drawn markings had been made during the census to label and number houses; and in some cases, they were no longer legible.
Vaccine coverage is described in detail elsewhere and was estimated between 76.7–92.7% of the population of the region, with the lower limit of the range estimated by census and registration data and the upper limit estimated from Survey2 [25]. A total of 41,242 individuals received 2-dose series of the OCV. Of the 518 Survey2 respondents, 480 (92.7%) [95% CI 90.1%–94.6%] reported receipt of at least one dose of the cholera vaccine, and 419 (80.8%) [95% CI 77.3%–84.0%] provided their vaccination cards for verification.
Baseline demographic characteristics for pre-and post-vaccine survey respondents were generally similar (Table 1); however statistically significant differences between the two time points were observed for household size, number of people sharing a toilet, toilet type, and having a farming occupation. 65.2% of Survey1 respondents reported use of latrine compared to 46.9% in Survey2. Farming was the most common occupation representing 69.5% of Survey1 respondents and 76.1% in Survey2.
Nearly all respondents pre-vaccine (99.1%) and post-vaccine (99.6%) had heard of cholera. A high level of knowledge was defined as greater than the median number of correct answers in Survey1 (Table 2). A significantly higher proportion of Survey2 respondents (63.8%) knew ≥3 correct modes of cholera transmission compared to 48.1% in Survey1 (p<0.0001). A similar pattern was observed with cholera prevention questions. Pre-vaccination, 50.0% of respondents provided ≥3 correct answers on how to avoid cholera compared to 64.5% post-vaccine (p<0.0001). Finally, a higher percentage of individuals in Survey2 (44.1%) knew ≥3 means of water treatment compared to Survey1 (22.6%) with p<0.0001 (Figure 1).
None of the differentially distributed baseline variables significantly changed the effect estimates for any outcome; therefore, only the socioeconomic proxy variables (ever having attended school and access to electricity at home), and no additional variables, were included as covariates in the final multivariable models. For cholera knowledge, post-vaccination surveys were associated with a statistically significant increase in the odds of providing at least 3 correct responses on means of cholera transmission (odds ratio [OR] 1.91; 95% CI 1.52–2.40; p<0.0001). For cholera prevention measures, the odds ratio of knowing 3 or more correct answers in Survey2 compared to Survey1 was 1.83 (95% CI, 1.46–2.30; p<0.0001). Similarly, there was also greater odds of knowing ≥3 ways to treat water in Survey2 relative to Survey1 (OR 2.75; 95% CI, 2.16–3.50; p<0.0001). Ever having attended school and electricity access in the home, were not generally associated with increased knowledge (Table 3); however, we did observe a positive relationship between access to electricity in the home and knowing 3 or more means of avoiding cholera of borderline statistical significance (OR: 1.37; 95% CI: 1.00–1.89).
The percentage of respondents who reported “always” treating their water increased from 49.4% in Survey1 to 62.0% in Survey2 (p<0.0001). The most common reasons provided for not always treating water were related to access to products. 35.9% had “no products” in Survey1 and 49.2% reported the same reason in Survey2. Products were “hard to get” for 28.2% and 35.0% of respondents in Survey1 and Survey2 respectively. Regarding hand washing practices, 46.7% of Survey2 respondents reported hand washing with soap and water >4 times a day compared to 41.1% in Survey1 (p 0.05). We observed decreased use of river water in Survey2 (42.7%) versus Survey1 (48.0%), although this was not statistically significant (p 0.06).
Multivariable regression analysis of hygiene practice revealed that relative to the pre-vaccination period, post-vaccination participants were more likely to report always treating water (OR1.62; 95% CI, 1.28–2.05; p<0.0001). Similarly, odds of washing hands with soap and water >4 times a day was increased in Survey2 relative to Survey1 (OR1.30; 95% CI, 1.03–1.64; p 0.03). Higher socioeconomic status, as measured by ever having attended school and access to electricity, was associated with increased odds of always treating water and hand washing with soap and water >4 times a day (Table 3). There were no confounding variables associated with practice questions.
Knowledge of water treatment as a means of preventing cholera was associated with the practice of always treating water (OR 1.47; 95% CI, 1.14–1.89; p 0.003). Overall, there was no statistically significant association between knowledge of hand washing as a cholera preventive measure and practice of frequent hand washing (OR 1.10; 95% CI, 0.82–1.46; p 0.53). However, in stratified analyses, knowledge of hand washing as a preventive measure was significantly associated with the practice of washing hands >4 times a day post-vaccine (OR 2.47; 95% CI 1.35–4.51; p 0.003) but not pre-vaccine (OR 0.85; CI 0.61–1.19; p 0.35).
This study aimed to evaluate the impact of a cholera vaccination campaign in rural Haiti on knowledge of cholera and health practice related to transmission and prevention of waterborne illnesses. It revealed that post-vaccination campaign, there was a significant increase in baseline knowledge and improvement in practice essential to cholera control.
Our pre-vaccination surveys revealed that at baseline, 48.1%, 50%, and 22.6% of respondents knew at least 3 means of cholera transmission, prevention methods, and treating water respectively. Nationwide health education campaigns on cholera prevention and transmission seem therefore to have reached this rural community, although, these proportions appear low. This may partly be related to the timing of our study that occurred almost two years after the outbreak when the intensity of public health messaging may have waned. Furthermore, the remote location of our rural study population combined with limited electricity may have hampered access to national mass media campaigns. A KAP survey conducted two months after the onset of cholera in the capital city, Port-au-Prince, showed 71.9% of respondents indicated consumption of contaminated water as a cholera transmission mode while 86.0% identified hand washing as a preventive measure [24]. In cholera endemic regions, rates of high knowledge on cholera from survey data range from 46.0% in Bangladesh to 84.8% in Tanzania [20], [21].
This study demonstrates that an OCV campaign with a strong public health education component was associated with increase in knowledge of cholera transmission, preventive measures, and methods of treating water. We also observed significant improvement in health practices essential for prevention of waterborne diseases after the vaccine campaign. Beau de Rochars et al. similarly reported significant improvement in water treatment practices in Haiti from 30.3% before cholera to 73.9% after community wide education campaigns in response to the outbreak [24]. Currently, there are no available data on the impact of a cholera vaccine program on knowledge and behavior related to cholera. Our cholera vaccination campaign provided an opportunity to raise awareness and directly reinforce public health messages about cholera control in the target population. Our findings demonstrate that an OCV campaign can be complementary to and even strengthen other cholera control efforts during an epidemic. Similarly, other vaccination programs may potentially function as health system strengthening tools in resource limited settings.
Our study also showed an association between knowledge and practice. Although a KAP study in Bangladesh demonstrated that good knowledge of cholera was associated with better prevention practices [20], other studies have shown hygiene practice rates were not commensurate with knowledge [21], [23]. It is important to note that KAP surveys do not explore the nuances of the social and economic context that influence or even deter the translation of knowledge into practice. For example, our surveys identified access to products as an important barrier to the practice of frequent water treatment. We also found that surrogates of higher socioeconomic status were associated with increased frequency of hand washing and water treatment. This may be attributed to the fact that individuals of higher socioeconomic status are likely able to afford soap and products for treating water. Although these products are distributed periodically free of charge by government and non-government organizations, they ordinarily must be purchased. They were not distributed to households at the time of the survey, but distributions did take place to some extent between the two surveys. Despite the apparent association between knowledge and practice, it is crucial to consider the various factors beyond information that influence health practices, particularly in resource limited settings. Moreover, it is not yet evident how levels of knowledge and hygiene practices as measured by KAP surveys actually impact cholera epidemics. To our knowledge, no data exists to confirm that higher knowledge and improved hygiene as measured by KAP surveys result in improved outcomes (e.g. decreased incident cases and mortality rates) in areas experiencing an epidemic.
This study has some limitations. First, we cannot exclude that other factors or interventions, external to the OCV campaign, were responsible for the findings. Nevertheless, despite our organization's presence in the area, work with the Ministry of Health, and consultation with the local water authorities at the time of writing, we are unaware of any other blanket community hygiene and education programs that occurred between the two surveys, other than our OCV campaign activities and routine public health messaging about the epidemic. There are technical water improvement initiatives that began in April 2012, but they do not have significant community educational components related to cholera or waterborne disease. A pre and post survey outside the catchment area of the OCV campaign would have provided comparison data, but this was not feasible as part of the vaccination campaign. Second, our study relied on self-report to assess water treatment and other hygiene practices so we cannot verify that reported practice was actual practice. Third, while we aimed for random, systematic sampling, the programmatic nature of the survey and the environment presented challenges in its execution. In Survey1, we interviewed 8.5% of households that completed the census, which was lower than 10% that would be expected when surveying every 10th household. If some enumerators restarted their count of every 10th household daily, instead of continuing the count across the days over the two-week census, this would explain the lower than expected survey rate. For post-vaccine surveys, we were unable to survey 13.7% of the 600 randomly generated households, partly due to lack of visible address markings on homes, families who moved away, lack of directions for homes in the census data, and challenges related to the rainy season. We lack information to assess whether households surveyed and not surveyed were comparable and whether respondents were similar across the two surveys. Nonetheless, we believe that it is unlikely there was a systematic bias in the inclusion households, and therefore it is unlikely that excluded households had significantly better or worse knowledge and practice about cholera than the included households. If surveyed households had different knowledge levels and practices than those not surveyed, this would bias our absolute estimates of knowledge and practice, but would unlikely influence our overall findings of improved knowledge and practice unless the extent or pattern of excluded households differed across the two Surveys. Finally the unequal distribution of some sociodemographic characteristics between the two survey populations raises the possibility of unmeasured differences in populations sampled. However, we believe that the observed differences reflect population-level changes over time such as seasonal variations in occupation and latrine access. For instance, post-vaccine surveys were administered later in the agricultural season when more participants may have identified as being farmers. Latrines are also at increased risk of overflowing in the rainy season, thus potentially forcing more individuals to resort to open defecation.
After an integrated cholera vaccination campaign in rural Haiti, surveys demonstrate a significant increase in knowledge of cholera transmission and prevention mechanism as well as improvement in practices of water treatment and frequent hand washing, which are critical for curbing the spread of diarrheal diseases such as cholera. This provides evidence that oral cholera vaccination can be part of comprehensive cholera control and can reinforce, rather than detract from, other prevention activities in Haiti.
|
10.1371/journal.pbio.2003268 | A map of protein dynamics during cell-cycle progression and cell-cycle exit | The cell-cycle field has identified the core regulators that drive the cell cycle, but we do not have a clear map of the dynamics of these regulators during cell-cycle progression versus cell-cycle exit. Here we use single-cell time-lapse microscopy of Cyclin-Dependent Kinase 2 (CDK2) activity followed by endpoint immunofluorescence and computational cell synchronization to determine the temporal dynamics of key cell-cycle proteins in asynchronously cycling human cells. We identify several unexpected patterns for core cell-cycle proteins in actively proliferating (CDK2-increasing) versus spontaneously quiescent (CDK2-low) cells, including Cyclin D1, the levels of which we find to be higher in spontaneously quiescent versus proliferating cells. We also identify proteins with concentrations that steadily increase or decrease the longer cells are in quiescence, suggesting the existence of a continuum of quiescence depths. Our single-cell measurements thus provide a rich resource for the field by characterizing protein dynamics during proliferation versus quiescence.
| The cell cycle is by nature highly dynamic, but we lack a standardized map of how core cell-cycle regulators change over time. In this study, we used time-lapse microscopy to track the dynamics of key cell-cycle proteins in individual human cells and found several unexpected patterns, even for well-studied proteins such as Cyclin D1. Our data provide a rich resource for those focused on the cell cycle, or on any biological process that is impacted by the cell cycle, by providing a series of maps of protein dynamics during cell-cycle progression and cell-cycle exit.
| Cellular proliferation is driven by the mitotic cell cycle, a highly regulated process consisting of DNA synthesis (S phase) and mitosis (M phase), separated by gap phases (G1 and G2). Decades of cell-cycle research have led to in-depth understanding of the biochemical processes involved in cell-cycle progression, but the temporal dynamics of these processes, and how they differ in non-cycling cells, are less well characterized. Simplified diagrams of cell-cycle dynamics are sometimes depicted in textbooks [1,2,3], but these diagrams are not always in agreement, typically only Cyclin dynamics are represented, and information on protein behavior during quiescence is absent. Thus, although the cell cycle is one of the most dynamic processes in biology, we lack quantitative information about the chronology of key events during cell-cycle progression versus cell-cycle exit.
An abbreviated explanation of the events surrounding cell-cycle entry and cell-cycle progression follows, with Fig 1A serving as a simplified network diagram. In quiescent or resting cells, Cyclin-Dependent Kinase (CDK) activities are low or off, and the master regulator of cell-cycle entry, the retinoblastoma protein (Rb), is in a non-phosphorylated state in which it binds and inhibits the E2F transcription factor. Cell-cycle entry can be triggered when resting cells receive extracellular mitogenic signals. Mitogenic signaling leads to Erk-dependent activation of transcription factors, such as c-Myc [4] and Ets-1 [5], which in turn up-regulate Cyclin D. Cyclin D binds its cognate Cyclin-Dependent Kinases, CDK4 and CDK6, which initiate hypo-phosphorylation of Rb. In the textbook model, this initial hypo-phosphorylation of Rb liberates the E2F transcription factor, a key driver of genes involved in the G1/S transition, including Cyclin E [6,7]. Transcriptional up-regulation of Cyclin E drives Cyclin-Dependent Kinase 2 (CDK2)/Cyclin E activity, leading to “hyper” phosphorylation of all 14 sites on Rb, and liberating additional E2F in a positive feedback loop. However, this model was recently called into question by the observation that E2F target genes were only up-regulated at the time of Rb hyper-phosphorylation and not with the initial hypo-phosphorylation [8]. Nevertheless, it is generally accepted that Rb hyper-phosphorylation marks passage through the Restriction Point (R-point) [9], defined as the time at which cells no longer require mitogens to complete the rest of the cell cycle [10]. Concordantly, activation of CDK2 was shown via single-cell time-lapse microscopy to mark cells that had passed the R-point [11].
At the beginning of S phase, Cyclin A protein levels begin to rise, and Cyclin A/CDK2 becomes the dominant source of CDK activity driving cells through S phase. DNA replication is initiated when origins of replication, previously prepared for replication by licensing factors such as Cdt1, fire due to phosphorylation by Dbf4-dependent kinase and CDK activities [12]. To prevent relicensing and re-replication of DNA, Cdt1 is degraded at the start of S phase by the E3 ubiquitin ligases SCFSkp2 and CRL4Cdt2 [13]. Any residual Cdt1 is bound and inhibited by Geminin, the levels of which rise during S and G2 [14,15]. Toward the end of S phase, Cyclin B levels rise rapidly, giving rise to Cyclin B/CDK1 activity that propels cells into mitosis [16]. The anaphase-promoting complex/cyclosome (APC/C) triggers exit from mitosis and is responsible for resetting the cell cycle at the end of mitosis via the degradation of Cyclin A, Cyclin B, Geminin, and many other substrates [17]. Cell-cycle progression is also controlled by protein inhibitors of CDKs, including p21 and p27, the ubiquitination and degradation of which promote S phase entry [18,19,20,21,22].
Cells can also temporarily exit the cell cycle by transitioning to a resting state, termed quiescence or G0. Relative to our knowledge of G1, S, G2, and M, the G0 phase remains poorly understood, both in terms of when and how cells transition into and out of G0 and in terms of a molecular definition of G0. Although there are multiple forms of quiescence, a universal feature of quiescence is lack of progression through the cell cycle [23]. Previous efforts to characterize quiescence in human cells have used serum starvation, contact inhibition, or loss of adhesion to induce quiescence, identifying a set of genes expressed across all three modes of quiescence induction, as well as sets of genes specific to the initiating quiescence signal [24,25]. Indeed, synchronization procedures have been shown to induce stress responses specific to the synchronization procedure used [26,27,28]. Characterization of quiescent cells from unperturbed populations has been hindered by the lack of a molecular marker to identify living quiescent cells.
The recent development of a sensor for CDK2 activity enables the identification of live cells that are in quiescence [11]. This sensor consists of an mVenus-tagged section of DNA Helicase B (DHB-mVenus) containing CDK2 phosphorylation sites close to a nuclear localization sequence (NLS) and a nuclear export sequence (NES) (S1A Fig). Phosphorylation of the sensor by CDK2 masks the basic residues of the NLS and unmasks the NES, causing translocation of the sensor to the cytoplasm in a manner that is correlated with CDK2 activity. The cytoplasmic:nuclear ratio of this sensor thus serves as a readout for CDK2 activity. Cells early in the cell cycle show nuclear localization of the sensor and low CDK2 activity, whereas cells toward the end of the cell cycle show cytoplasmic localization of the sensor and high CDK2 activity. S phase begins when the cytoplasmic:nuclear ratio of the sensor is approximately 1. Addition of a CDK2 inhibitor at any time during the cell cycle causes an immediate drop in CDK2 activity, visualized by rapid nuclear translocation of the sensor [11].
When single-cell traces of CDK2 activity from asynchronously cycling cells are aligned to the time of mitosis, a bifurcation in CDK2 activity becomes apparent, which corresponds to the proliferation-quiescence cell fate decision [11]. One subset of cells completes mitosis with residual CDK2 activity (cytoplasmic:nuclear ratio of the sensor ≥ 0.5), which then steadily increases over the course of the cell cycle (CDK2inc cells). Another subset of cells completes mitosis with low or no CDK2 activity and enters quiescence (CDK2low cells; cytoplasmic:nuclear ratio of the sensor < 0.5). This quiescence is transient in nature. Indeed, CDK2low cells experience a second cell fate decision in which they can continue to remain quiescent or emerge from quiescence and re-enter the cell cycle. Cells that emerge from quiescence can be identified by a renewed increase in CDK2 activity (CDK2emerge cells). Thus, upon completion of mitosis, cells can become proliferating CDK2inc cells or quiescent CDK2low cells. CDK2low cells can remain CDK2low for variable amounts of time or re-enter the cell cycle by becoming CDK2emerge cells.
Entry into the CDK2low state occurs in all cell lines examined thus far, even under optimal culture conditions (full-growth media at subconfluent densities). While it is known that the bifurcation in CDK2 activity is regulated by p21 [11], our understanding of why cells enter the CDK2low state is incomplete. We recently showed that 50% of the transits through the CDK2low state can be explained by replication errors carried over from the previous (mother) cell cycle [29]. The trigger for entry into the CDK2low state in the other 50% of CDK2low cells remains unknown, but it is possible that these cells are also experiencing an unidentified stress. Because cells enter the CDK2low state without any exogenous trigger, we refer to CDK2low cells that exist under optimal culture conditions as “spontaneously” quiescent, to contrast with other well-established types of quiescence in which cells are “forced” into quiescence (e.g., serum starvation or contact inhibition).
Despite substantial knowledge about the mechanism of cell-cycle transitions, we do not have a clear picture of overall cell-cycle dynamics detailing the rise and fall of protein levels and appearance and disappearance of protein post-translational modifications. In large part, this is because biochemical approaches in synchronized cells typically monitor only a few protein species at low time resolution. Proteomic surveys of the cell cycle have provided a more global view of cell-cycle events in mammalian cells but also suffer from low temporal resolution [26,30]. Furthermore, any method that relies on cell synchronization to enrich for cells at a specific cell-cycle stage is likely to exert stress on cells, which pollutes actual cell-cycle regulation with regulatory mechanisms operative as cells emerge from an arrested state. In addition, bulk analysis approaches blur heterogeneity in cell-cycle behavior, potentially resulting in incorrect interpretations of biological data. In contrast, time-lapse microscopy can offer single-cell measurements at millisecond temporal resolution in asynchronous cells but is limited by the difficulty of designing live-cell fluorescent readouts of multiple cell-cycle regulators and by the challenges of automated image processing and cell tracking. Most recently, immunofluorescence (IF) staining of fixed-cell snapshots has been used to infer cell-cycle kinetics of a handful of proteins [31,32,33], but without distinguishing proliferating from quiescent cells. Given that spontaneously quiescent cells appear in varying proportions in all cycling populations examined thus far, failure to distinguish proliferating cells from spontaneously quiescent cells leads to increased apparent cell-to-cell variability and decreased accuracy in quantifying protein behavior.
Here we combine the best of live-cell microscopy and antibody-based measurement to map key molecular events during cell-cycle progression versus spontaneous cell-cycle exit. By categorizing cells by their CDK2 activity trajectory (CDK2inc, CDK2low, CDK2emerge) and computationally aligning their IF signal as a function of time-since-anaphase, we reduce cell-to-cell variability in protein measurements and eliminate potential artifacts from synchronization procedures. In this way, we identify several unexpected differences in protein levels and modification states between cells that are progressing through the cell cycle and have increasing CDK2 activity (CDK2inc cells) and cells that are quiescent (CDK2low). One noteworthy example is Cyclin D, which is well known (and confirmed here in MCF10A cells) to be expressed at low levels in cells forced into quiescence by serum starvation or contact inhibition, but which we show is more abundant in spontaneously quiescent CDK2low cells compared with proliferating CDK2inc cells. We also identified 4 proteins with concentrations that steadily increase or decrease the longer the cells are in spontaneous quiescence. This result suggests that there exists a continuum of quiescence depths. Together, our single-cell data provide a chronology of key events during the active cell cycle and reveal key molecular differences between forced quiescence, spontaneous quiescence, and proliferation.
We used 2 complementary single-cell methods to chronicle the dynamics of key cell-cycle regulators. The first method uses 4-color IF snapshot images to categorize individual cells as G1, S, G2, M, or G0/quiescent. This approach has the advantage of being readily applicable to any cell line without the need to insert fluorescent sensors or perform time-lapse microscopy but does not explicitly carry time-dependent information. By co-staining cells with Hoechst (to measure DNA content) and EdU (a marker for DNA synthesis) [34], we could subdivide the cell cycle into 5 categories (Fig 1B–1D and S1B Fig): cells with 2N DNA content and no EdU incorporation were classified as G0 or G1; cells with near 2N DNA content and intermediate EdU signal were classified as early S phase; cells with high EdU signal were classified as S phase; cells with near 4N DNA content and intermediate EdU signal were classified as late S phase; and cells with 4N DNA content and no EdU incorporation were classified as G2 or M (Fig 1B).
To further distinguish cells in G0 from cells in G1, we co-stained cells with an antibody against phospho-Rb at either Serine 780 or Serine 807/811. These sites are phosphorylated by CDK2 and thus can serve as a fixed-cell readout of CDK2 activity. The phospho-Rb signal is bimodally distributed, representing hypo- and hyper-phosphorylated Rb (Fig 1C). Newly born cells with hypo-phosphorylated Rb were previously shown to be in the CDK2low state, whereas newly born cells with hyper-phosphorylated Rb are in the CDK2inc state [11]. Therefore, EdU-negative cells with 2N DNA content and hypo-phosphorylated Rb are classified here as G0/quiescent, and EdU-negative cells with 2N DNA content and hyper-phosphorylated Rb are classified here as G1 (Fig 1C). To distinguish cells in G2 from cells in M, we used an antibody against phospho-Histone H3 (pHH3), a well-established marker for mitosis. EdU-negative cells with 4N DNA content that were pHH3-negative were classified as G2, and cells that were pHH3-positive were classified as mitotic (Fig 1D). We used 3 fluorescent channels to stain cells with Hoechst, EdU, and either phospho-Rb or pHH3 (S1C Fig), and used the fourth channel to measure 1 of 14 proteins of interest in MCF10A human mammary epithelial cells. We also validated our results in Hs68 human foreskin fibroblasts. We avoided use of cancer cell lines, which often have mutations in the core cell-cycle regulatory network.
The second method involves time-lapse microscopy over 24 hours of MCF10A cells expressing Histone 2B (H2B) fused to mTurquoise and the CDK2 sensor fused to mVenus. Immediately after the last frame was taken, cells were fixed with para-formaldehyde, processed for IF, and reimaged. Custom MATLAB-based cell-tracking scripts were used to extract single-cell traces of CDK2 activity, with a custom “jitter correction” to re-register the images before and after IF (see Materials and methods). In this way, we can match each cell’s IF staining to its history. The H2B signal is used to automatically identify the frame of anaphase for each cell, which enables automated alignment of all CDK2 activity traces (and consequently each cell’s IF signal) to each cell’s final anaphase of the movie.
The resulting plot demonstrates the bifurcation in CDK2 activity that is evident as cells complete mitosis and assume either a CDK2inc, CDK2low, or CDK2emerge state (Fig 1E) [11]. Another subset of cells has no mitoses during the course of the 24-hour movie, of which a further subset has low CDK2 activity for the entire 24-hour imaging period. Although these cells are not cycling, they are also not senescent (S2A–S2C Fig) and thus appear to be in a prolonged quiescence (Fig 1F). Indeed, we confirmed that these cells can emerge from this prolonged quiescence (S2D Fig). In our unperturbed MCF10A cells, 95.6% ± 5.4% of the total population divided at least once during the 24-hour imaging. Of the total population,79.0% ± 6.2% entered the CDK2inc state after mitosis, 8.8% ± 2.6% remained CDK2low after mitosis, and 7.8% ± 1.7% entered the CDK2low state after mitosis but built up their CDK2 activity before the end of the imaging period (CDK2emerge) (Fig 1E). Among the 4.4% that did not divide during the course of the movie, 52.3% ± 17.7% stayed in a prolonged quiescence (representing 2.3% ± 0.8% of the total population, Fig 1F) and 47.7% ± 17.7% (or 2.1% ± 0.8% of the total population) were observed to build up CDK2 activity before the end of the imaging period (S1 Movie and S2D Fig). For CDK2emerge cells, automated identification of the time point when cells begin building up CDK2 activity after being in a CDK2low state allows automated alignment of CDK2emerge traces to this event (Fig 1G), which we have previously argued represents the R-point [11,35]. Alignment of the CDK2 activity traces in these various ways allows for the staging of IF-based protein levels or modification states as a function of time-since-anaphase, or time-since-R-point.
Cells treated with EdU for 15 minutes at the end of a 24-hour time-lapse sequence illustrate the power of this approach—CDK2inc cells display the classic “rainbow” pattern of EdU as a function of time-since-anaphase, allowing us to identify and label G1, S, and G2 phases of the cell cycle in our time-lapse + IF experiments (Fig 1H, blue). CDK2low cells (Fig 1H, red) and prolonged quiescent cells (Fig 1H, purple) display no EdU signal. A moving average through the CDK2inc and CDK2low subpopulations further illustrates the effect (Fig 1I). CDK2emerge cells aligned to the time at which CDK2 activity begins to increase show a pattern similar to the CDK2inc cells but with less clarity due to the difficulty of automating the identification of the first frame of CDK2 activity rise (relative to the easy automatic identification of the first frame of anaphase; Fig 1J). This plot shows that CDK2emerge cells begin S phase at a similar time after the initial CDK2 activity buildup as CDK2inc cells.
These 2 methods were used to chronicle the dynamics of 14 proteins during cell-cycle progression and spontaneous quiescence. The proteins were chosen because of the availability of selective antibodies, their role as core cell-cycle regulators (Cyclin A2, Cyclin B1, Cyclin E, Cyclin D1, p21, p27, Cdt1, Geminin, total Rb, and phospho-Rb), or as important signaling inputs to the cell cycle (cMyc, Fra1, phospho-cJun, and p53). Nine proteins were highly dynamic over the course of the cell cycle (Cyclin A, Cyclin B, Cyclin E, Cyclin D, p21, Cdt1, Geminin, cMyc, and phospho-Rb; Figs 2–5 and S3 Fig), whereas others tested were relatively invariant over the cell cycle in the cell types examined here (p27, total Rb, p53, Fra1, and phospho-cJun; S4 and S5 Figs).
Using multi-color IF in MCF10A cells, we began by inferring the dynamics of various cell-cycle proteins using (1) a density scatter plot of signal intensity versus DNA content (Fig 2, Column 1); (2) a contour plot of signal intensity versus DNA content, in which cells are grouped into 7 cell-cycle phases, as described in Fig 1B–1D and S1B Fig (Fig 2, Column 2); and (3) a histogram of signal intensity in G0/quiescent cells (2N, EdU-negative, hypo-phosphorylated Rb) versus G1 cells (2N, EdU-negative, hyper-phosphorylated Rb), as defined in Fig 1C (Fig 2, Column 3). We also repeated these experiments in a second cell type, non-immortalized Hs68 human foreskin fibroblasts (S3 and S4 Figs).
Cyclin A2 and Cyclin B1, two of the best-understood cell-cycle proteins, behaved in textbook fashion and serve as a proof-of-principle. In cycling cells, Cyclin A2 rose linearly during S and G2, consistent with previous reports (Fig 2A, scatter and contour plots; and Fig 3A) [31,32,33,36,37]. Cyclin B1 levels did not begin to rise until late S but then rose supra-linearly, as previously reported (Fig 2B, scatter and contour plots; and Fig 3B) [16,31,32]. Cyclins A2 and B1 were both degraded in mitosis [38,39], and both were undetectable in G0 and G1 cells (Fig 2A and 2B, histograms; and Fig 3A and 3B).
When Cyclin E levels were plotted against DNA content, we detected a subtle “N”-shaped pattern in which Cyclin E rose in G1 and fell in S phase, as expected (Fig 2C, scatter plot; [31,36,40,41]). The rise in G1, and fall in early S phase, of Cyclin E is also detected in the time-lapse + IF data for CDK2inc cells (Fig 3C). In contrast with Cyclins A2 and B1, which remained “off” in CDK2low cells, Cyclin E levels rose steadily in CDK2low cells (Fig 3C). This is surprising because Cyclin E is overexpressed in several cancers and Cyclin E/CDK2 activity is a major driver of cell-cycle progression [42]. Therefore, we expected Cyclin E levels to be lower in spontaneously quiescent cells compared with proliferating cells. We note, however, that these high levels of Cyclin E in G0/quiescent cells are not accompanied by high CDK2 activity and thus are not able to stimulate cell-cycle progression; by definition, we identify these quiescent cells because of their lack of CDK2 activity (CDK2low). This lack of CDK2 activity despite high levels of Cyclin E is likely due to the accompanying high levels of p21 in these cells (see below). Thus, a likely explanation for the high levels of Cyclin E in G0/quiescent cells may be that Cyclin E in these cells has not been subjected to S phase–mediated degradation, which depends on CDK2 activity [40,41]. We also observed that the Cyclin E antibody utilized here, the widely used clone HE12, detects a strong nonspecific signal in MCF10A cells, in addition to detecting Cyclin E (S6A and S6B Fig). Thus the difference in Cyclin E signal between CDK2low and CDK2inc cells may be partly obscured by the nonspecific signal.
The patterns displayed by Cyclin D1 were also unexpected. MCF10A cells express Cyclin D1, D2, and D3, with Cyclin D1 at the highest level of the three [43]. Thus Cyclin D1 is the prevalent D-type cyclin in our cells, and the antibody used in this study is selective for Cyclin D1 (S6A and S6B Fig). When Cyclin D1 levels were plotted against DNA content, we detected a “U”-shaped pattern in which Cyclin D1 is high in cells with 2N DNA content, low in S phase, and elevated again in cells with 4N DNA content (Fig 2D, scatter and contour plots). This pattern has been reported previously [44,45] but is not widely appreciated. Upon closer inspection, the EdU-negative cells with 2N DNA content reveal highly heterogeneous expression of Cyclin D1—cells with hypo-phosphorylated Rb (G0 cells) have much higher levels of Cyclin D1 than cells with hyper-phosphorylated Rb (G1 cells) (Fig 2D, histogram). Like Cyclin E, this is surprising because Cyclin D is considered a driver of the cell cycle and is overexpressed in several cancers [46]; therefore, its levels are expected to be higher in proliferating cells than in quiescent cells. When we examined our time-lapse + IF data, we observed the same phenomenon—cells born into the quiescent CDK2low state had high Cyclin D1 levels, whereas CDK2inc cells that were actively progressing through the cell cycle again displayed a “U”-shaped pattern, with Cyclin D1 levels being moderate in G1, low in S phase, and moderate again in G2 (Fig 3D). In addition, prolonged quiescent cells also have high Cyclin D1 levels (Fig 3D, purple).
By way of explanation, we considered the possibility that Cyclin D1 levels appear higher in G0 cells simply because Cyclin D1 in these cells has not been subjected to S phase-mediated degradation [47,48]. However, the moving average of Cyclin D1 levels indicated that CDK2low cells have higher levels of Cyclin D1 than CDK2inc cells, even in cells 2 hours after birth, before S phase-mediated degradation could play a role. We also note that CDK2low cells have more Cyclin D1 than CDK2inc cells ever have, at least on average. However, high levels of Cyclin D do not necessarily correspond to high CDK4/6 activity [49,50], and there is as yet no single-cell assay to measure CDK4/6 activity in these cells. An alternative explanation for the high Cyclin D1 levels in CDK2low cells is that Cyclin D1 protein levels are stabilized by high levels of p21 in these cells [51,52,53].
Indeed, p21 displays the same “U”-shaped pattern as Cyclin D1 does when plotted against DNA content (Fig 2E, scatter and contour plots). As with Cyclin D1, cells with hypo-phosphorylated Rb (G0/quiescent cells) have high levels of p21, whereas EdU-negative, 2N DNA content with hyper-phosphorylated Rb (G1/proliferating cells) have very low levels of p21 (Fig 2E, histogram). Moreover, time-lapse + IF data revealed that p21 levels are high in newly born G0/CDK2low cells and very low in newly born G1/CDK2inc cells, as reported previously (Fig 3E) [11]. CDK2emerge cells show initially high levels of p21 that then decay around the time that CDK2 activity turns back on (Fig 3E, green), consistent with the notion that a decay in p21 enables a rise in CDK2 activity.
CDK2inc cells maintain very low levels of p21 throughout all of G1 and S phase (Fig 2E, contour plot; and Fig 3E, blue). While these data are consistent with our previous studies [11,29], these results differ from the common notion that p21 levels are generally high in G1 cells [54]. A likely explanation for this discrepancy is that many previous studies used various treatments (e.g., nocodazole or serum starvation) for cell synchronization, which exert stress on cells and can increase p21 levels [55,56]. Furthermore, immunoblotting does not allow fine-grained analysis of p21 heterogeneity or temporal behavior. More recent single-cell experiments tracking exogenous YFP-p21 in U2OS osteosarcoma cells detected newly born cells with and without YFP-p21 [57]. However, without a live-cell marker to distinguish G0 from G1, it is not possible to know if the newly born cells with elevated p21 are actually passing through a G0/CDK2low state rather than going straight to G1. Similarly, the cells born without detectable p21 could represent a G1/CDK2inc subpopulation.
The dynamics of Cdt1 are expected to have some similarities to p21 because both proteins are substrates of the E3 ubiquitin ligase CRL4Cdt2 [13], a feature reflected in our IF data (Fig 2F, scatter and contour plots). However, in direct contrast to p21, Cdt1 levels are high in G1 cells and lower in G0/quiescent cells (Fig 2F, histogram). Time-lapse + IF data show a similar trend, revealing that any residual Cdt1 present in CDK2low cells is quickly degraded to the basal level seen in S phase cells (Fig 3F). The levels of Geminin, an inhibitor of Cdt1, are out of phase with Cdt1, as expected [13,58,59]. Geminin levels are undetectable in quiescent CDK2low cells and begin to rise in early S phase, consistent with Geminin’s role as a substrate of the APC/C (Figs 2G and 3G) [35,58]. However, unlike Cyclin A2, which rises steadily and linearly, Geminin levels plateau by mid-to-late S phase, a feature seen in both IF and time-lapse + IF data, suggesting an additional layer of transcriptional regulation.
We also examined the cell-cycle dynamics of c-Myc, a key protein that links MAPK signaling to cell-cycle entry [60]. More recently, c-Myc has been shown to act as a “transcription amplifier” as opposed to a classic transcription factor [61,62]. Here we show that c-Myc is strongly cell-cycle regulated. Immunofluorescence reveals that c-Myc levels are higher in G1 cells than in G0/quiescent cells (Fig 2H, histogram), consistent with a pro-proliferation role for c-Myc. CDK2low cells maintain low c-Myc levels as long as they remain in the CDK2low state but then up-regulate c-Myc upon emerging from the CDK2low state (Fig 3H). c-Myc levels rise steadily in S and G2 phases (Fig 2H, scatter and contour plots; and Fig 3H).
Phospho-Rb is bimodally distributed among EdU-negative cells with 2N DNA content (Fig 1C) [11]. The switch from hypo- to hyper-phosphorylated Rb marks passage through the R-point [9,63,64], and while this event is often cited as occurring in mid- to late G1 [9,63], we have shown previously that MCF10A cells are born into a state of either hypo- or hyper-phosphorylated Rb immediately upon completion of mitosis [11]. Here we extend this result by confirming that the same is true using an antibody against Rb phosphorylation at another site, Serine 780 (Fig 3I)—cells born into the quiescent CDK2low state have hypo-phosphorylated Rb, whereas cells born into the cell cycle-committed CDK2inc state have hyper-phosphorylated Rb. This phospho-Rb-S780 signal continues to rise as CDK2inc cells progress through the cell cycle. Examination of the CDK2emerge cells provides additional information by revealing that cells present with hyper-phosphorylated Rb as soon as the rise in CDK2 activity can be detected, indicating that hyper-phosphorylation of Rb occurs prior to or concurrently with activation of CDK2 (Fig 3I, green).
Given the surprising behavior of several proteins in spontaneous quiescence (e.g., rising Cyclin D1, Cyclin E, and p21 levels), we compared our results in spontaneously quiescent cells with quiescence induced by well-established methods, namely serum starvation (Fig 4A and 4B) and contact inhibition (S6C Fig). By both quantitative western blotting and IF, we were able to reproduce the canonical protein dynamics upon serum starvation or contact inhibition in which the levels of Cyclin D1, Cyclin E, p21, and all other proteins examined, fell as a function of time in quiescence. We also validated the selectivity of the antibodies used for IF via siRNA knockdown (S6A and S6B Fig) and provide sample images for each IF stain (S7 Fig).
We next sought to further validate the unexpected dynamics of Cyclin D1 using an antibody-independent method. We used CRISPR-mediated genome editing of MCF10A cells to tag Cyclin D1 at its endogenous locus with mCitrine, a yellow fluorescent protein, and subsequently transduced the cells with H2B-mTurquoise and mCherry-tagged CDK2 sensor. Western blotting and PCR revealed that both alleles of Cyclin D1 were tagged with mCitrine (S8A and S8B Fig) and IF revealed a linear correlation at the single-cell level between the mCitrine-Cyclin D1 signal and an antibody stain against Cyclin D1 (S8C Fig). In agreement with our time-lapse + IF results for Cyclin D1, single-cell tracking of the mCitrine-Cyclin D1 cell line showed that CDK2low cells have elevated Cyclin D1 levels compared with CDK2inc cells (Fig 4C red traces, and Fig 4D bottom panel), and that the levels of Cyclin D1 for CDK2inc cells are moderate in G1, low in S phase, and moderate again in G2 (Fig 4C blue traces, and Fig 4D top panel). These results explain why Cyclin D1 expression was recently reported to be a poor predictor of the time spent between mitosis and S phase [45].
Given that c-Myc levels are low and Rb is hypo-phosphorylated (and thus that E2F transcription is inhibited) in the spontaneously quiescent CDK2low cells, what factors could be driving the high levels of Cyclin D1? Since these 2 major cell-cycle transcription factors are likely off in CDK2low cells, we hypothesized that the high Cyclin D1 levels in these cells could be due to a lack of degradation. Indeed, Cyclin D1 levels are strongly regulated not only by transcription but also by protein degradation via cullin-RING ligases (SCF with various F-box proteins) [65]. Because the majority of cullin-RING ligases require covalent modification by NEDD8 for holoenzyme ubiquitin ligase activity, their activity can be inhibited by blocking their neddylation with the small molecule MLN4924 [66]. We therefore filmed mCitrine-Cyclin D1 cells before and after an acute treatment with 1.4 μM of MLN4924 and selected for analysis only those cells that received drug 1–2 hours after mitosis (during G0/G1). Consistent with our hypothesis, inhibition of cullin-RING ligases caused an increase in Cyclin D1 in CDK2inc cells to a level that was comparable with that in CDK2low cells. Thus, lack of Cyclin D1 degradation in CDK2low cells is a major contributor to the high levels of Cyclin D1 seen in these cells.
Cyclin D1 is well known for its short half-life. These results suggest that the stability of Cyclin D1 varies with cell-cycle phase—the half-life of Cyclin D1 is short in CDK2inc cells but much longer in CDK2low cells. Together, these validation experiments lend confidence in our overall approach and in the unexpected findings in this work.
To compare the relative protein dynamics in proliferating versus quiescence cells, we normalized and overlaid the moving average data from Fig 3 for CDK2inc and CDK2low cells (Fig 5A; note that normalizing the signals masks differences in dynamic range among proteins). Proteins were grouped into 2 plots according to their behavior in quiescent CDK2low cells—Group 1 contains signals that are “off” in quiescent cells (Fig 5A, top), and Group 2 contains signals that change dynamically over time in quiescent cells (Fig 5A, bottom). The identification of 4 proteins in Group 2 that either steadily increase (Cyclin D1, Cyclin E, p21) or steadily decrease (Cdt1) the longer a cell has been quiescent suggests that quiescence is not just a single static state but rather that certain aspects of a cell’s proteome evolve as a function of time spent in quiescence (at least over the 24-hour period that we measured).
We then schematized these results to create diagrams that depict the chronology and dynamics of cell-cycle events (Fig 5B). The 5 proteins in Group 1 all increase their levels as cells progress through the proliferation cycle, albeit with different dynamics. Cyclin A2 starts to accumulate in S phase and continues to increase until M phase. Geminin begins to accumulate at the same time but plateaus in G2. Cyclin B1 and c-Myc remain low until late S phase. Rb phosphorylation on Serine 780 steadily increases throughout the whole proliferative cycle. All of these proteins reset at mitosis and maintain low levels in quiescent cells. The dynamics of Group 2 proteins are more variable. Cyclin D1 and Cdt1 turn on in G2 after being low or off in S phase. Cyclin D1 increases further if cells go into G0, and degrades when CDK2inc cells re-enter the cell cycle. Cdt1 decreases slowly when cells enter the CDK2low state and decreases rapidly when CDK2inc cells enter S phase. Cyclin E starts to increase at the completion of mitosis and continues to increase throughout G0 and G1; Cyclin E levels drop because of degradation at the G1/S transition but remain elevated in G0/quiescent cells. p21 levels are low in proliferating cells but increase steadily once cells enter quiescence. Such diagrams provide a quantitative resource for understanding the dynamics of cell-cycle proteins relative to one another.
Using single-cell time-lapse microscopy and IF, combined with automated image processing and cell tracking, we have characterized the dynamics of key cell-cycle proteins in unperturbed proliferating and spontaneously quiescent cells and compared these with cells forced into quiescence by serum starvation or contact inhibition. Our measurements provide a rich resource for those focused on the cell cycle, or on any biological process that is impacted by the cell cycle, by providing a map of standard cell-cycle behavior in non-tumorigenic cells. Unlike most characterizations of cell-cycle behavior, which use chemical synchronization such as nocodazole or double thymidine block, our data come from asynchronous, unperturbed single cells. We are therefore able to chart, at high time resolution, both mean population behavior as well as cell-to-cell variability in protein levels and modification states. All cultured populations of somatic human cells that we have examined thus far actually contain mixtures of proliferating and spontaneously quiescent cells. This generates extensive cell-to-cell variability, which would obscure even single-cell IF data aligned by time-since-anaphase, if one were unable to distinguish the proliferating, quiescent, and emerging populations using the CDK2 sensor.
When we classified proteins based on their behavior in quiescent CDK2low cells, we identified a set of 4 proteins (Cyclin D1, Cyclin E, p21, and Cdt1), whose concentrations increase or decrease the longer cells are in quiescence. This suggests that quiescence is not a homogenous “off” state, but rather that the quiescent cell state changes continually, at least over our 24-hour observation period. These data support the existence of a continuum of quiescence depths.
It is well documented that the levels of Cyclins D and E are dramatically reduced in cells forced into quiescence via serum starvation [42,67]. Here we compared the dynamics of multiple key cell-cycle proteins, including Cyclin D1 and Cyclin E, in forced versus spontaneous quiescence. In contrast to the declining levels of Cyclin D1 and Cyclin E in serum-starved or contact-inhibited cells, the levels of Cyclin D1 and Cyclin E rise while cells are in the quiescent CDK2low state. We confirmed this result using endogenously tagged mCitrine-Cyclin D1 and further showed that the high levels of Cyclin D1 in CDK2low cells arise because of reduced cullin-RING ligase-mediated protein degradation in the CDK2low state. Similarly, high levels of Cyclin E in CDK2low cells likely arise because this Cyclin E has not been subjected to S phase–mediated degradation, which depends on CDK2 activity [40,41].
Examination of cells emerging from a transient quiescence (CDK2emerge cells) reveals that CDK2emerge cells recapitulate the protein dynamics of cells that immediately enter the CDK2inc state after mitosis. Put another way, the protein dynamics of CDK2inc cells that are born committed to the cell cycle with elevated CDK2 activity are similar to the protein dynamics of CDK2emerge cells that commit to the cell cycle at variable times after dividing. This result argues that the beginning of the active cell cycle is marked by the increase in CDK2 activity and that any time cells spend prior to the activation of CDK2 represents a period of cell-cycle exit that we have referred to as G0/quiescence.
Which signaling events are causes of, and which are simply consequences of, entry into quiescence? The full answer will require extensive analysis using acute perturbations of the proteins in question, but based on the data presented here, we can already speculate that the behavior of proteins with levels that are similar in newly born CDK2inc and CDK2low cells is likely to simply be a consequence of entry into quiescence (e.g., Geminin, Cyclin A2, and Cyclin B1). In contrast, proteins with levels that are already distinct in newly born CDK2inc and CDK2low cells have already been shown to be causative (e.g., p21 [11]) or have the potential to be causative in the proliferation-quiescence decision (e.g., Cyclin D1 and phospho-Rb).
In summary, the experimental and computational approaches employed here enable the creation of chronological maps of protein dynamics during cell-cycle progression and cell-cycle exit in asynchronous single cells, revealing several differences compared with previous results generated from synchronized cells. These maps will be informative for mathematical modeling of the cell cycle and can also serve as a benchmark for comparing the cell cycle of non-transformed cells with the cell cycle of various cancer cells. Our work also highlights the fact that there are multiple molecularly distinct states of quiescence, depending on the initiating trigger. Together, our data provide new information for answering fundamental questions about normal cellular control over proliferation and add new molecular knowledge to the poorly documented state(s) of G0/quiescence.
MCF10A human mammary epithelial cells were maintained in DMEM/F12 (ThermoFisher) supplemented with 5% horse serum (Invitrogen), 20 ng/ml epidermal growth factor (EGF, Sigma-Aldrich), 0.5 mg/ml hydrocortisone (Sigma-Aldrich, St. Louis, MO), 100 ng/ml cholera toxin (Sigma-Aldrich), 10 μg/ml insulin (Invitrogen), and penicillin/streptomycin. For serum starvation media, the horse serum, EGF, and insulin were removed, and 0.3% BSA was added. For live-cell time-lapse imaging, phenol-red free DMEM/F12 was used. Hs68 primary human foreskin fibroblasts were cultured in DMEM with 10% FBS and penicillin/streptomycin. Both cell lines were purchased from ATCC. Hs68 cells can be propagated for 42 passages according to ATCC and are not immortalized; cells were received at passage 12 and were used within 13 passages of receipt. MCF10A cells expressing the CDK2 sensor (DHB-mVenus) and tagged histone H2B (H2B-mTurquoise) are as described [11].
Integration of the mCitrine-encoding gene into the CCND1 locus was carried out using CRISPR technology [68]. A CRISPR-Cas9 ribonucleoprotein (RNP) complex was generated using the CRISPR-Cas9 System from IDT. The RNP contains crRNA (GGAGCUGGUGUUCCAUGGCUGUUUUAGAGCUAUGCU) annealed to tracrRNA and Cas9 nuclease. The RNP was electroporated into MCF10A cells using the Neon system from Life Technologies following the manufacturer’s protocol with 2 pulses, 30 ms at 1150 V. Single cells were sorted by flow cytometry into 96-well plates and grown into clones. Western blot and IF against Cyclin D1 protein, as well as PCR of the CCND1 gene, were carried out as validation. Data from clone 2A7 is shown in this work (S8 Fig). For functional validation, cells were treated with Mek inhibitor (PD0325901, S1036 from Selleckchem) at 100 nM for 32 hours, or treated for 32 hours followed by a Mek inhibitor washout for 6 hours. To confirm a similar response to inhibition of degradation for both mCitrine-Cyclin D1 and endogenous Cyclin D1, the mCitrine-Cyclin D1 line and parental wild type MCF10A line were treated with MLN4924 (Active Biochem, A-1139) at 1.4 μM or Bortezomib (Cayman Chemical,10008822) at 1 μM for 2 hours (S8A Fig).
siRNA oligos were synthesized by Dharmacon: CCNA2 (MU-003205-02-002), CCNE1 (MU-003213-02-0002), CCNE2 (MU-003214-02-0002), CDKN1A (MU-003471-00-0002), CCND1 (MU-003210-05-0002), RB1 (MU-003296-03-0002) or IDT: CCNB1 (hs.Ri.CCNB1.13.1), GMNN (hs.Ri.GMNN.13.1), CDT1 (hs.Ri.CDT1.13.2), MYC (hs.Ri.MYC.13.2), and Negative Control DsiRNA (51-01-14-04). The oligos were electroporated into MCF10A cells following manufacturer’s instruction (Neon system, Life Technologies). Cells were fixed for IF or lysed for western blotting 20 hours (for short-live proteins: Cyclin A2, Cyclin B1, Cyclin E, Cyclin D1, c-Myc, p21, Geminin, and Cdt1) or 48 hours (for longer-live proteins: Rb) after the electroporation.
Antibodies used in this study are p21 Waf1/Cip1 (CST #2947) at 1:250, phospho-Rb (Ser807/811) (CST #8516) at 1:250, phospho-Rb 780 (BD Biosciences #668385) at 1:250, p21 (BD Biosciences #556430) at 1:250, total Rb (a gift from Julien Sage) at 1:200, p53 (DO-1) (Santa Cruz sc-126) at 1:100 and p53 (Ab-1) (Calbiochem OP03) at 1:100, Fra-1 (Santa Cruz #28310) at 1:200, Cyclin E clone HE12 (Zymed #32–1600) at 1:400, Cyclin D1 clone SP4 (Thermo Scientific RM-9140-S0) at 1:250, Cyclin A2 (Santa Cruz #751) at 1:500, phospho-Histone H3 (Ser10) (CST #9706 and #9701) at 1:200, p27 (BD Bioscience #610241) at 1:100, Geminin (CST #5165) at 1:250, Cyclin B1 (CST #4138) at 1:100, c-Myc (CST #5605) at 1:250, CDT1 (CST #8064) at 1:200, phospho-c-Jun (Ser73) (CST #3270) at 1:800, and Alexa Fluor-488, -546, -647 secondary antibodies (ThermoFisher) at 1:500.
For Cyclin E IF, cells were fixed in −20°C methanol for 5 minutes and then washed twice with PBS. For all other antibodies, cells were fixed with 4% paraformaldehyde and then washed twice with PBS. Cells were then incubated with a blocking/permeabilization buffer (10% FBS, 1% BSA, 0.1% TX-100 and 0.01% NaN3 for antibodies against Cyclin E, p21, Cdt1, Geminin, Fra1, p53, and phospho-c-Jun) for an hour at room temperature, or sequentially permeabilized with 0.2% TX-100 for 15 minutes at 4°C and blocked with 3% BSA for an hour at room temperature (for antibodies against Cyclin A2, Cyclin B1, Cyclin D1, c-Myc, p27, and total Rb). Primary antibody staining was carried out overnight at 4°C in the corresponding blocking buffer and visualized using secondary antibodies conjugated to Alexa Fluor-488, -546, or -647. Where phospho-Rb and phospho-Histone H3 antibodies were used in conjunction with an antibody for a protein of interest in Fig 2, cells were processed using the method appropriate for the protein of interest. Where indicated, cells were incubated in media containing 10 μM EdU for 15 minutes, and then fixed and processed according to manufacturer’s instructions (ThermoFisher #C10340).
Images were acquired on an ImageXpress Micro XLS widefield microscope (Molecular Devices) with a 10X 0.45NA objective and processed using custom scripts in MATLAB.
Cells were plated at least 24 hours prior to imaging in phenol red-free full-growth media in a 96-well plate (Greiner bio-one #655090) such that the density would remain subconfluent until the end of the imaging period. Images were acquired every 12 minutes on an ImageXpress Micro XLS widefield microscope (Molecular Devices) with a 10X 0.45NA objective; CFP exposure = 75 ms; YFP exposure = 200 ms. Cells were imaged in a humidified, 37°C chamber at 5% CO2.
Images were processed as described in Cappell et al., 2016 ([35]), with a general description reproduced here: Mean nuclear intensities were measured by averaging the background-subtracted pixel intensities in each nucleus as defined by a nuclear mask. The nuclear mask was established by performing segmentation on H2B-mTurquoise- or Hoechst-stained images as follows. Log-transformed images were convolved with a rotationally symmetric Laplacian of Gaussian filter and objects were defined as contiguous pixels exceeding a threshold filter score. In order to segment cells in contact with their nearest neighbor, a custom segmentation algorithm was implemented to detect and bridge concave inflections in the perimeter of each object (hereafter referred to as the “deflection bridging algorithm”). The deflection bridging algorithm was implemented on every identified object in the first imaging frame and then only adaptively in subsequent frames. This was accomplished by iteratively tracking cells in each frame, detecting probable merge events (as discussed below) and selectively implementing the deflection bridging algorithm on putative merged objects. Local background subtraction was performed on images of sensors or antibodies that were nuclear in subcellular distribution. For local background subtraction, the nuclear mask was expanded by 25 μm and the background for each cell was calculated as the median pixel intensity of local nonmasked pixels. For cytoplasmically localized sensors or antibodies, the nuclear mask was dilated by 50 μm, and the global background was calculated as the mode intensity of all nonmasked pixels. As before, CDK2 activity was calculated as the ratio of cytoplasmic to nuclear mean DHB fluorescence, with the cytoplasmic component calculated as the mean of the top 50th percentile of a ring of pixels outside of the nuclear mask. Tracking of cells between frames was implemented by screening the nearest future neighbor for consistency in total H2B-mTurquoise fluorescence (“conservation of mass”).
Because the stage jittered slightly after fixation and IF in the time-lapse + IF dataset, we implemented the following jitter correction procedure to ensure precise matching of the CDK2 activity trace of each cell to its IF intensity: We first subtracted the image at a specific time from the image in the next frame to get a “difference score” between 2 images. We then repeated the process, with 1 image moving in a 2-dimensional manner, to get multiple “difference scores” when the stage jittered. The position with the lowest score indicated the amount of jittering and the images were aligned accordingly.
“Conservation of mass” was further exploited to detect merges or splits, which allowed recovery of overlapping traces. Mitosis events (called at anaphase) were called when the total H2B fluorescence of the 2 nearest future neighbors of a given cell were both between 45% and 55% of the total H2B fluorescence of the past cell. The R-point was defined as the time CDK2 activity first began to rise. Computationally, this involves calculating slopes of CDK2 activity using windows of 6–10 time points and then maximizing a linear function for time-since-mitosis, CDK2 activity, and CDK2 slope (long times-since-mitosis, low CDK2 activity, and high CDK2 slope).
The tracking code is available for download here: https://github.com/scappell/Cell_tracking.
Traces were computationally classified, and manually verified, as CDK2inc (blue), CDK2low (red), or CDK2emerge (green) based on CDK2 activity at 2 hours after mitosis: CDK2inc traces must remain ≥ 0.5 for all frames post-anaphase; CDK2low traces must remain < 0.5 for all frames post-anaphase; CDK2emerge traces initially enter the CDK2low state and then emerge—these traces must remain < 0.5 for at least 3 hours post-anaphase before rising.
|
10.1371/journal.pgen.1000553 | The Caenorhabditis elegans HNF4α Homolog, NHR-31, Mediates Excretory Tube Growth and Function through Coordinate Regulation of the Vacuolar ATPase | Nuclear receptors of the Hepatocyte Nuclear Factor-4 (HNF4) subtype have been linked to a host of developmental and metabolic functions in animals ranging from worms to humans; however, the full spectrum of physiological activities carried out by this nuclear receptor subfamily is far from established. We have found that the Caenorhabditis elegans nuclear receptor NHR-31, a homolog of mammalian HNF4 receptors, is required for controlling the growth and function of the nematode excretory cell, a multi-branched tubular cell that acts as the C. elegans renal system. Larval specific RNAi knockdown of nhr-31 led to significant structural abnormalities along the length of the excretory cell canal, including numerous regions of uncontrolled growth at sites near to and distant from the cell nucleus. nhr-31 RNAi animals were sensitive to acute challenge with ionic stress, implying that the osmoregulatory function of the excretory cell was also compromised. Gene expression profiling revealed a surprisingly specific role for nhr-31 in the control of multiple genes that encode subunits of the vacuolar ATPase (vATPase). RNAi of these vATPase genes resulted in excretory cell defects similar to those observed in nhr-31 RNAi animals, demonstrating that the influence of nhr-31 on excretory cell growth is mediated, at least in part, through coordinate regulation of the vATPase. Sequence analysis revealed a stunning enrichment of HNF4α type binding sites in the promoters of both C. elegans and mouse vATPase genes, arguing that coordinate regulation of the vATPase by HNF4 receptors is likely to be conserved in mammals. Our study establishes a new pathway for regulation of excretory cell growth and reveals a novel role for HNF4-type nuclear receptors in the development and function of a renal system.
| The function of many important biological structures requires the construction of very complex cellular shapes. For example, mammalian kidneys or related renal systems in other animals rely on the formation of elongated tubes that maximize surface area to facilitate the exchange of ions between the body and excreted fluid. Defects in kidney development or function may lead to kidney failure or polycystic kidney disease. Mechanisms involved in orchestrating the formation and function of the elaborate tube structures in renal systems are still poorly characterized. Here, we show a novel transcription factor involved in the growth and elongation of an excretory tube in C. elegans. This factor helps manage tube development by regulating genes involved in ion transport and membrane fusion, likely helping to balance the growth of the inner and outer portions of the excretory tube as this structure elongates. This transcription factor shares significant homology with a mammalian protein that participates in hormone signaling and is present in the kidney tubules, suggesting that elongation and growth of tube structures may rely on a new kind of hormonal communication that occurs between distant parts of the cell; this signaling mechanism may be important for appropriate kidney development in humans.
| Nuclear receptors (NRs) comprise a large family of transcription factors distinguished by a highly conserved DNA binding domain and a structurally conserved ligand-binding domain. NRs are notable for their ability to interact with small molecule ligands, enabling these factors to respond to autocrine, paracrine, and endocrine signals in order to mediate transcriptional effects at a distance [1],[2]. The canonical NR family is exclusively found in metazoans and the number of nuclear receptor members varies dramatically depending on species; from 21 NR genes in Drosophila melanogaster, to ∼50 in rodents and humans, to over 250 NRs in Caenorhabditis elegans and related nematodes [3]. The extraordinarily large NR family of C. elegans is particularly intriguing. Of the 283 predicted NR genes, only 15 are directly orthologous to NRs found in other metazoans, including Drosophila and mammals [4]. The remaining 268 NRs are thought to be derived from extensive duplication and diversification of an ancestral gene most closely related to the mammalian and Drosophila HNF4 receptors [5]. The presence of both highly similar and divergent HNF4-type receptors in nematodes implies that many of these proteins will carry out conserved structural and physiological functions, whereas others will have evolved to adopt responsibilities more specific to the nematode lineage. This idea is supported by the fact the C. elegans NHR-49 nuclear receptor shares many of the metabolic functions of the mammalian HNF4α, but not the developmental activities [6],[7]. Thus, study of C. elegans NRs should not only be helpful for understanding mammalian NR function and physiology, but should also reveal novel regulatory activities for the nuclear receptor family.
The prospect that the responsibilities of mammalian receptors may be divided among a larger number of NRs in C. elegans may be advantageous for understanding the physiological function these complex proteins. For example, the mammalian HNF4α plays numerous roles in development, metabolism, and disease [8]; because of this widespread physiological impact, the functional and mechanistic diversity of this receptor is far from understood. Indeed, mutations in the human HNF4α are associated with maturity onset diabetes of the young (MODY) and late onset type II diabetes; yet, how these HNF4α lesions lead to diabetes has not been established [9]–[11]. Furthermore, there is considerable controversy over the quantity and identity of HNF4 target genes [12]–[14]. These complications may be due, at least in part, to the fact that HNF4α carries out essential functions in several different tissues, and that HNF4α likely regulates different target sets depending on metabolic, developmental, and nutritional context.
HNF4α is also expressed in many cell types for which its function has not yet been established; for example, the epithelial cells of the intestine and the proximal and convoluted tubules of the kidney, and while HNF4α has been shown to regulate proliferation of transformed kidney cell lines, its role in kidney development remains to be defined [15],[16]. The C. elegans renal system is comprised of only three cells, yet these cells carry out many of the same functions as mammalian kidneys [17],[18]. Therefore, C. elegans might be an advantageous system in which to study the role of HNF4 receptors in renal development. The largest portion of the C. elegans excretory system consists of the excretory cell (EC). The development of the EC is extraordinary, as it involves the formation and growth of four branches that project outward from a single nucleus located near the anterior bulb of the pharynx [17]. These branches grow along the length of the animal to near the tip of the head and tail in early development, and then continue to grow along with the animal until adulthood. Each branch of the EC contains an inner membrane that coalesces to form a lumen; thus, the excretory cell becomes a large, single cell tube. Consequently, the EC has been effectively used to understand the development of tubes and to investigate mechanisms involved in excretory function [17],[19],[20]. At this point, factors known to participate in the development and function of the C. elegans excretory cell include vATPases, WNK kinases, CLIC-like proteins, Patched related proteins, and mucins [17], [21]–[24]. Additionally, the CEH-6 homeobox protein has also been implicated as the only transcriptional regulatory factor, thus far, involved in excretory cell development [25]. How the complex structure of the EC is developed and maintained so precisely, even at points very distant from the primary sites of gene regulation, remains a mystery.
We have found a highly conserved C. elegans HNF4 paralog, NHR-31, that is specifically expressed in the excretory cell of the nematode, suggesting that investigation of this receptor may provide unique insight into the role of nuclear receptors in renal development and tube formation. In this study, we show that NHR-31 specifically regulates the expression of genes that coordinate the synchronous growth and elongation of excretory canals, demonstrating a novel NR mediated pathway for renal system development and function.
nhr-31 is predicted to encode an HNF4α related nuclear receptor (NR) protein with a highly conserved DNA binding domain (DBD) and ligand binding domain (LBD) (Figure 1A). To help establish the physiological function of this NR, we determined the tissues in which the nhr-31 gene is expressed. A GFP reporter construct was generated by fusing 3.0 kb of nhr-31 upstream regulatory sequence to the gfp gene (Pnhr-31::gfp). Injection of Pnhr-31::gfp into WT worms revealed that the nhr-31 promoter drives strong expression in the excretory cell (EC). In transgenic animals, GFP protein was first observed in the EC cell shortly after EC birth and persisted in the EC for the remainder of worm embryogenesis, larval development, and adulthood (Figure 1B and data not shown). GFP was observed throughout the cytoplasm of the H-shaped excretory cell. Because our reporter construct was designed by fusing only the nhr-31 promoter to the gfp gene, the GFP localization pattern does not represent NHR-31 protein sub-cellular localization. Pnhr-31::gfp expression was also observed, at lower levels, in the intestine and in several unidentified cells located near the tail (Figure 1B and data not shown).
In C. elegans, the EC functions cooperatively with duct and pore cells, and together these cells are important for maintaining osmolarity homeostasis [26],[27]. To determine if nhr-31 RNAi animals displayed compromised excretory function, we treated animals with nhr-31 RNAi or control RNAi from the L1 to L4 stage of development and then stressed L4 animals with acute exposure to a standard growth plate supplemented with 500 mM NaCl, and determined their ability to respond to these unfavorable conditions. 250 animals were assayed at each time point. After just two hours, less than 5% of nhr-31 RNAi animals could be rescued from 500 mM NaCl exposure. In contrast, L4 animals fed control RNAi were able to thrive for much longer under these same conditions, with over 50% of animals maintaining the ability to recover even after 8 hours of high salt exposure (Figure 1C). These data indicate that reducing nhr-31 gene expression strongly impairs the ability to survive acute osmotic stress.
Three different nhr-31 deletion strains have been isolated, and all of these strains are inviable (www.wormbase.org). Using one of these strains (nhr-31(tm1547)), we found that nhr-31 deletion leads to early embryonic lethality (data not shown). Additionally, application of nhr-31 RNAi throughout growth and development results in significant embryonic lethality in the F1 generation (data not shown). Thus, NHR-31, like its mammalian homolog HNF4α, plays an essential role in early embryonic development. Because we found that the nhr-31 gene is primarily expressed in the excretory cell during larval and adult stages, however, we investigated the participation of nhr-31 in EC development and morphology using an RNAi feeding strategy that specifically reduced nhr-31 expression during larval development and adulthood. In postembryonic animals, the EC is an H-shaped cell, with four canals emanating from a main cell body located near the terminal bulb of the pharynx [17]. Two canals project along each side of the animal towards the posterior end, and two canals project forward towards the anterior end (Figure 2A). To monitor EC morphology, WT animals were injected with the Pnhr-31::gfp reporter. In WT adult animals, GFP localization revealed that the outer diameter of the excretory cell was relatively uniform through the entire length of the canal, measuring ∼3.5 µm in proximal sections of the posterior canal, and tapering to ∼2.4 µm in distal sections of the posterior canal (Figure 2B and 2C).
When WT animals carrying the Pnhr-31::gfp construct were treated with nhr-31 RNAi from the L1 stage of larval development through adulthood, the morphology of the adult EC was dramatically altered (Figure 2B and 2C). In particular, the excretory canals were not uniform in diameter; instead, they contained multiple enlarged varicosities, with diameters up to 10 µm (Figure 2B and 2C). These varicosities showed considerable variability in size and shape and were located along the entire length of the EC, including the proximal, middle, and distal portions of the posterior arms, as well as in the anterior branches of the EC canal (Figure 2B and 2C and data not shown). DIC images of nhr-31(+/−) heterozygotes also revealed similar excretory cell abnormalities, providing support for the specificity of our nhr-31 RNAi construct (Figure S1).
High magnification of the GFP images obtained in nhr-31 RNAi animals suggested that the varicosities consisted of dense cellular material with an abundance of vacuoles (Figure 3A). This phenotype was different from previously reported EC abnormalities, which showed enlargement of the EC cell due to fluid accumulation or cyst formation [19],[27]. To more closely examine the morphological defects in the EC of nhr-31 RNAi animals, we employed high pressure freezing transmission electron microscopy (HP-TEM). Table 1 shows quantitative analysis of sections obtained from the middle region of the EC in 5 different control RNAi animals and 5 different nhr-31 RNAi animals. Cross sections of the EC of a WT animal showed a single circular lumen with an average diameter of 1.6 µm (Table 1). Additionally, an abundance of well-formed canaliculi were clearly visible in WT animals (Figure 3C and Table 1). Canaliculi are smaller “mini-canals” surrounding the canal lumen; these canals are thought to greatly increase the apical surface area of the EC lumen (Figure 3B) [17]. Canaliculi were visible in the wild type excretory canal cross section as small, round, circular shapes and were regular in size and consistent (∼70/section) in number from section to section (Figure 3D and Table 1). According to our EM measurements, the average diameter of the EC was ∼2.8 µm, which agreed nicely with our GFP measurements (Figure 2C and Table 1).
HP-TEM imaging revealed multiple morphological defects in the excretory canals of nhr-31 RNAi animals, particularly in the varicosities (Figure 3D and Table 1). First, the average canal diameter increased to 5.8 µm, with larger varicosities displaying diameters of up to 8 µm, and the narrow regions showing diameters from 2–3 µm. Second, the average diameter of the lumen in nhr-31 RNAi animals was increased by 26% to 1.95 µm, and the lumen often appeared multi-lobed. The diameter of the lumen correlated strongly with the outer cell diameter, as the largest lumen diameter measurements were found within large varicosities (Table 1). Third, we found that the canaliculi were uncharacteristically irregular in size and present at much higher numbers (∼126/cell) in nhr-31 RNAi animals (Figure 3D and Table 1). Finally, the varicosities of nhr-31 RNAi animals possessed an unusually high number of large vesicles, elevated endoplasmic reticulum abundance, and a considerable increase in mitochondria (Figure 3E and Table 1). Importantly, the TEM cross sections showed that the varicosities were not a result of an EC canal lumen that was folded back on itself or bent away from the normal lateral alignment, or due to osmotic “swelling”, both of which have been previously reported for mutants that affected EC structure [19],[27]. Consequently, the EC phenotypes resulting from loss of nhr-31 function are different from previous observations and suggest that nhr-31 defects are distinctive in their mechanism of origin. In summary, both fluorescence confocal microscopy and TEM showed that loss of nhr-31 function leads to significant defects in EC canal size, shape, and microstructure. The abundance of cellular material and organelles, along with significant structural abnormalities, implies that the abnormal varicosities observed in adult nhr-31 RNAi animals are likely to result from regions of uncontrolled cellular growth.
We next applied gene expression profiling to establish downstream regulatory targets of nhr-31. Gene expression was measured using C. elegans oligomer based microarrays. We carried out this study in L4 larvae, as this is the larval stage at which the EC morphology differences between WT animals and nhr-31 RNAi animals first begin to show. Overall, we found that, in nhr-31 RNAi worms, the expression of 20 genes was suppressed by greater than 2-fold and the expression of 63 genes were enhanced by greater than 2-fold (Table S1).
The most striking outcome of our microarray experiments was the discovery that RNAi of nhr-31 dramatically affected the expression of 15 genes that encode subunits of the vacuolar ATPase (gene names are referred to as vha), and one gene predicted to code for a vATPase cofactor (gene name, R03E1.2). In fact, of the 30 genes most strongly reduced by inhibition of nhr-31, 15 of these were vha genes (Table S2). The vacuolar ATPase (vATPase) is an ATP-dependent proton pump, which transports protons across cellular membranes (Figure 4A). Each C. elegans vha gene encodes for one subunit of the holoenzyme, and there are 15 separate subunits that make up the holoenzyme. For several of the vATPase subunits, C. elegans possesses multiple gene isoforms; consequently there are 18 vha genes in total. As a secondary confirmation of the microarray data, we employed quantitative RT-PCR to specifically measure the mRNA levels of all 18 vATPase genes found in C. elegans. We found that the expression of 16 of these genes was reduced when nhr-31 was inhibited (Figure 4A). Importantly, previously published data show that nearly all vha subunits are expressed in the excretory cell, indicating NHR-31 is likely to be mediating expression of these vha genes directly in the EC (Table 2) [19], [20], [29]–[33]. Additionally, most vha genes are also expressed in the intestine, where NHR-31 also resides. Accordingly, the only two vha genes not regulated by NHR-31, vha-7 and unc-32, are not expressed in the excretory cell. In sum, our microarray and QRT-PCR convincingly demonstrate that a primary function of NHR-31 is to coordinately promote the expression of almost the entire complement of vacuolar ATPase genes. NHR-31 localization to the excretory cell, where nearly all vha genes are expressed, also argues that NHR-31 is regulating vha genes in this cell type.
Because the vacuolar ATPase subunits are highly expressed in the EC, we suspected that the impact of nhr-31 on EC development might be a consequence of vacuolar ATPase regulation. To test this hypothesis, we used RNAi feeding to specifically reduce the expression of three different vacuolar ATPase subunits: vha-5 (small a subunit), vha-8 (catalytic E subunit) and vha-12 (B subunit). Because previous studies have shown that RNAi of the vacuolar ATPase subunits leads to larval lethality, we did not apply vha or nhr-31 RNAi until the L3 stage of development. Using this approach, we found that RNAi of each of these subunits was sufficient to cause excretory canal formation defects similar to those of nhr-31 RNAi animals (Figure 4B). These results imply that the control of EC development by NHR-31 is mediated, at least in part, by its stimulation of vATPase expression. We also note that this experiment shows that knockdown of nhr-31 or vATPase expression specifically in late larval development is sufficient to cause irregular EC growth and adult varicosities.
Although the large, irregular, varicosities observed in nhr-31 RNAi animals were never observed in WT adults, we did notice varicosity-like structures early in WT larval development, residing at regular intervals along the EC canal in L1 and early L2 animals (Figure 5A and Table 3). These varicosities differed from those present in nhr-31 RNAi adults in that they displayed a consistently symmetrical oval shape (Figure 5A). In L1 larvae, ∼10 of these varicosities were observed in each EC canal branch, but as worms developed the regions of the excretory cell between varicosities grew wider and the varicosities consequently decreased in prominence such that, by the late L3 stage of development, the entire length of the excretory cell possessed a diameter similar to the varicosities observed in L1 animals (Figure 5B and 5E). The presence of these growth varicosities in WT L1 larvae was confirmed by hp-TEM (Figure 5C and 5D). According to these TEM measurements, L1 varicosities displayed a diameter that was 2.8 times that of narrow regions, and a lumen diameter that was about 2-fold larger than the narrow regions (Table 3). Additionally, the varicosity regions harbored many more canaliculi (Figure 5C and 5D and Table 3). This data implies that the varicosities may form in L1 animals and spread horizontally along the excretory cell to help increase cellular diameter, and perhaps also length. Thus, we suspected that nhr-31 RNAi animals might improperly maintain these structures such that they continue to enlarged and become irregularly shaped as animals developed into adults. However, examination of nhr-31 RNAi animals revealed no obvious signs of varicosities in the L3 stage of development, implying that knockdown of nhr-31 did not interfere with the normal dissipation of these structures during mid-larval development (Figure 5E). The varicosities that arise in nhr-31 RNAi animals first appear in the late L4 stage of development and continue to grow larger as animals grow older (Figure 5E). Consequently, the varicosities observed in adult nhr-31 RNAi animals must either occur from growth of new structures, or the reactivation and renewed growth of these original varicosities. We also note that the varicosities caused by nhr-31 loss of function continue to grow larger during adulthood, such that by day 2 of adulthood they are nearly twice as large as varicosities in early adults (Figure 5E).
Nuclear receptors typically associate with complex binding motifs comprised of two hexameric half-sites [2]. These half sites may be paired in multiple orientations with various amounts of spacing, and this architecture helps determine the type of NHR that binds. To identify NREs in the promoters of the C. elegans vacuolar ATPase genes, we used the NHR-computational analysis program “NHR-scan” [34]. This program identified strong NRE candidates in nearly all of the vha promoters; 15 out of 18 vha genes harbored candidate NREs in close proximity to their transcription start site. If a vha gene was expressed as part of an operon, NREs were found near the transcription start site of the first gene in the operon. Analysis of predicted NREs showed a strong presence of an AGTTCA consensus half site (Figure 6A and Table 4). The most common repeats were an ER6 (40% of all binding sites), which is an everted repeat separated by 6 base pairs and an ER8 (27% of all binding sites), an everted repeat separated by 8 base pairs. In fact, 13 of 19 vATPase genes had at least one highly conserved ER6 or ER8 site in their promoters, while several other types of AGTTCA repeats were also found once or twice in vATPase promoters. Interestingly, we also found a consensus ER6 site in the nhr-31 promoter, implying that nhr-31 may regulate its own expression through a feedback or feed-forward mechanism. This putative regulation did not manifest in our GFP reporter studies, however, implying that self-regulation in the excretory cell is not very significant during development.
The most common spacing for mammalian HNF4α receptors is a DR1 or DR2, however it would not be surprising if NHR-31 adopted a different NRE specificity, as nematode NR binding sites have likely evolved to generate NREs to help distinguish between all of the different HNF4 paralogs in C. elegans. Consistent with this notion, the NHR-31 LBD does not retain two conserved amino acids that help direct HNF4α homodimerization on DR1 and DR2 sites [35]. It is also possible, however, that everted repeats have not yet been widely characterized as HNF4α sites in other organisms. The presence of so many binding sites that closely match a consensus site is quit remarkable, especially since NHR response elements are notoriously degenerate [36]. Furthermore, nuclear receptor regulated genes often contain several conserved and cryptic NREs that are necessary for modulating expression level, consequently, there are likely to be important cryptic NREs in these promoters as well [37].
Analysis of the vATPase gene promoters from mice (Mus musculus) showed an astonishing enrichment of HNF4α binding sites (Table 4). In fact, we found highly conserved HNF4α binding sites in 10 vacuolar ATPase genes, and most of these genes harbored at least two independent HNF4α binding sites. The repeats were almost always in DR1 or DR2 configuration and the consensus half-site sequence for these sites was AG(G/T)TCA (Figure 6B), which matches the consensus site previously reported for HNF4α binding sites [38]. As with the C. elegans NREs, the enrichment of these binding sites is highly significant.
Taken together, these data strongly argue that coordinate regulation of vacuolar ATPase genes by the HNF4 nuclear receptor is conserved in mammals. We should note, however, that the DR1 and DR2 elements can also bind other mammalian nuclear receptors; therefore, even though NHR-31 is most closely related to HNF4α, and expressed along with vATPases in the excretory system, the participation of other mammalian nuclear receptors in coordinate regulation of vATPase genes cannot be ruled out. Similarly, we cannot rule out the involvement of additional C. elegans NRs in regulation of nematode vATPase genes.
We have identified a new pathway involved in the development of the C. elegans renal system. In summary, we have shown that the NHR-31 nuclear receptor, through promotion of vacuolar ATPase gene expression, is essential for the appropriate growth, morphology, and function of the C. elegans excretory cell. This study not only identifies a new transcriptional regulator necessary for EC development, but also establishes the specific regulatory targets that mediate its effects, and highlights potential nuclear receptor response elements. The regulatory or developmental activities carried out by NHR-31 have not yet been observed for a nuclear receptor; consequently our findings expand the physiological repertoire of the NR superfamily.
A primary function of NHR-31 is to maintain the structure of the EC canal during the transition from larval development into adulthood. When exposed to nhr-31 RNAi throughout larval development, or specifically in late larval development, we observed numerous large and irregular varicosities all along the length of the posterior and anterior EC canals, these varicosities first manifested in L4 development and continued to amplify and grow several days into adulthood. As the excretory cell is involved in the regulation of ion transport and osmolarity, we considered that these varicosities might have been due to accumulation of fluid within the EC cytoplasm to create “cyst-like” structures. However, HP-TEM revealed numerous sub-cellular abnormalities within the varicosities that could not be explained by an abnormal accumulation of fluid. For example, nhr-31 RNAi dependent varicosities generally contained abnormally shaped lumens, significant increases in the number of canaliculi, ER and mitochondria, and abnormally large numbers of ectopic vesicles. These data imply that the EC varicosities are not fluid filled, but rather overdeveloped. In contrast, in the narrow regions of the nhr-31 RNAi EC, we found normal numbers of mitochondria, ER, and canaliculi, implying the majority of EC irregularities that occur in nhr-31 RNAi animals are localized to the enlarged varicosities. This excessive growth phenotype significantly differs from previously characterized excretory cell phenotypes [19],[27].
Another intriguing finding of our study is that NHR-31 has a surprisingly specific and strong impact on the expression of v-ATPase encoding genes (vha genes). The vacuolar ATPase (v-ATPase) is an ATP-dependent proton pump that is organized into a peripheral domain (V1), which is responsible for ATP hydrolysis, and an integral domain (V0), responsible for proton transport. Although it is referred to as the vacuolar ATPase, this enzyme is found in multiple intracellular membranes, including endosomes, lysosomes, Golgi-derived vesicles, clathrin coated vesicles, secretory vesicles, as well as the plasma membrane [39],[40]. vATPases are important for numerous cellular functions, including ion transport, substrate transport, acidification of vesicles and other organelles. Additionally, recent studies have shown that vATPases also play a predominant role in vesicular trafficking of the endocytic and exocytic pathways, participating directly in membrane fusion by not only providing the proper acidic environment, but also by directly forming protein complexes during the fusion process [40]. Given the diversity of vATPase functions, it seems likely that the transcription of vATPase would be precisely regulated both spatially and temporally in order to facilitate the development and function of different cell types. Although numerous factors have been shown to regulate the vATPase at the enzymatic level, our study has identified a transcription factor with a specific role in regulating vATPase expression in a tissue specific manner.
In C. elegans it has been shown that vATPase subunits of either the V0 sector or the V1 sector, are important in excretory cell development and morphology [19]. In this previous study, several distinct vha subunits were knocked down early in development resulting in several defects in the hypodermis, cuticle, and excretory cell. Specifically, abnormal structures were observed in the ECs that were described as “whorls”. Because RNAi of nhr-31 leads to the reduced expression of 17 out of the 19 genes that encode vha subunits, we suspected that the role of nhr-31 in EC development may be due, at least in part, to regulation of vATPase gene expression. In support of this hypothesis, we found that larval specific knockdown of NHR-31 target genes encoding either an a subunit, an E subunit, or a B subunit of the vATPase, led to excretory cell phenotypes nearly identical to those observed in nhr-31 RNAi animals. Although varicosities found in our experiments may be related, in some fashion, to the “whorls” observed in the previous study ([19], it should be noted that the previous study focused on reduction of vATPase expression much earlier in development. In contrast, in our study, vha expression was knocked down specifically in late L3 development through early adulthood. Thus, our findings show that regulation of vATPase expression is a prominent factor in NHR-31 function.
The phenotypic abnormalities observed in nhr-31 RNAi animals, combined with the predicted function of nhr-31 regulatory targets, provide several clues into how this nuclear receptor may impact the generation of a healthy EC (Figure 7). A critical component of EC development is the outgrowth of the excretory canals. During larval development, four excretory canals must grow out of the main cell body and migrate towards the posterior and anterior ends of the animal and then continue to grow as the animal increases in length. We observed that, during early larval development, the EC migrates along the length of the animal and is periodically punctuated with small oval shaped varicosities. By the time a worm reaches later larval stages, these varicosities are no longer present and adult EC canals are exquisitely uniform in diameter. We suggest that the growth varicosities that form during early larval development may be regions of high cellular growth activity, where robust protein, organelle, and membrane synthesis occur, these areas of growth then serve to supply material to the cytosol, as well as the basal and apical membranes of the EC, thus enabling the EC canal to elongate in a bidirectional manner. TEM images of the EC in L1 larvae, which show periodic varicosities with a more dense supply of membrane and organelles, support this hypothesis (Figure 5C and 5D). As the EC reaches its full-length, precise regulation of new cellular synthesis and cellular elongation reaches equilibrium such that regions of high EC cellular mass become evenly distributed and the EC adopts a fully mature and uniform shape.
Many of defects observed in nhr-31 RNAi animals are consistent with an inability to properly regulate the coordination between EC cell outgrowth and new synthesis of cellular material. Thus, the NHR-31 nuclear receptor may play an important role in regulating the growth and elongation of the EC cell, and, in nhr-31 RNAi animals, excess lipid synthesis and other factors involved in cellular growth proceed unchecked leading to the production of new EC cellular material, even as this cell is no longer growing lengthwise. In this scenario, actively growing regions of the excretory cell could not expand laterally in either direction; consequently, excess cellular material would accumulate in varicosities that continue to grow larger even after animals reach adulthood.
It is astonishing that NHR-31 controls such a small and specific set of target genes, and that nearly all of its targets comprise subunits or cofactors of the vATPase. While the fundamental conclusions of this study are not dependent upon the mechanism by which NHR-31 regulates gene transcription, NHR-31 is a transcription factor of the nuclear receptor type, and therefore it is tempting to propose that NHR-31 regulates the vATPases in response to a ligand signal by directly binding to the vATPase promoters. Consistent with this hypothesis, our binding site analyses of the vATPase promoters revealed a significant enrichment of nuclear receptor response elements in the form of ER6 or ER8 everted repeats with an AGTTCA consensus half site (Figure 6A and Table 4). The fact that this response element does not perfectly match the preferred response element architecture of the mammalian HNF4α is not surprising, as C. elegans contains dozens of HNF4-like receptors, and it is likely that NREs have evolved in nematodes in order to distinguish between NHR paralogs. We did, however, find strong enrichment of classical HNF4α binding sites (DR1 and DR2) in the promoters of the mouse vATPase genes, suggesting that coordinate regulation of the vATPase by HNF4 type receptors may be well conserved in mammals, even though the exact response element architecture may have changed.
The physiological functions and target genes of nhr-31 have not been previously linked to an HNF4-type receptor, or any other nuclear receptor. NHR-31 shares a high degree of homology with mammalian HNF4 receptors, including nearly perfect conservation of key DNA binding elements and a strongly conserved ligand-binding domain (LBD). Interestingly, it has been proposed that the mammalian HNF4 receptors interact with free fatty acids and fatty acyl-CoA molecules [41],[42]. An ability of NHR-31 to bind to the acyl chain of a fatty acid or lipid molecule would provide a provocative explanation for how NHR-31 may be coordinating membrane synthesis and cellular elongation in the EC, which is likely to be occurring at sites distant from the nucleus. Because intensive membrane synthesis, transport, and fusion must take place in order to meet the needs of a growing excretory cell, such processes may release lipid based signals that activate or repress NHR-31 control of vacuolar ATPases and other genes associated with membrane biogenesis. Whether or not the functions of NHR-31 are conserved in mammals remains to be determined; however, the fact that both HNF4α and vacuolar ATPases are expressed at high levels in the proximal tubules of the mammalian kidney, combined with our demonstration that the mammalian vATPase genes contain a high density of HNF4α binding sites, implies that a functional role for HNF4 receptors in coordinate regulation of the vATPase in the renal system may indeed be a conserved process [16],[39].
The N2 Bristol strain of C. elegans was used for all experiments. Worms were maintained by standard techniques at 20–22°C. nhr-31 RNAi constructs were created by introducing the full-length NHR-31 cDNA into the L4440 RNAi feeding vector (Andy Fire, Stanford University). RNAi constructs for vha-5, vha-8, and vha-12 were obtained from the Ahringer RNAi library (University of Cambridge, Cambridge, UK). All RNAi constructs were transformed into the HT115 strain of E. coli and RNAi was introduced to N2 worms by RNAi feeding. RNAi expression was induced in the feeding bacteria by growing bacteria on NGM plates containing 3 mM IPTG and 100 µg/ml carbenicillin. Bacteria containing an empty L4440 RNAi vector were used for the RNAi control. Although NHR-31 is part of a large family of related nuclear receptors, these receptors have extensively diverged from one another during evolution, such that the closest paralog of NHR-31 shares only 55% homology in cDNA sequence; therefore, it is highly unlikely that there will be cross reactivity of the RNAi. Furthermore, C. elegans RNAi prediction programs do not indicate any cross reactivity (www.wormbase.org) [28]. Finally, the fact that nhr-31(+/−) heterozygotes displayed similar EC defects further supports the specificity of this RNAi construct (Figure S1).
An nhr-31 promoter/gfp reporter construct (Pnhr-31::gfp) was generated by fusing ∼3 kb of upstream regulatory sequence and 17 base pairs of the first nhr-31 exon to the gfp gene, primers were created using the nhr-31a.1 predicted isoform. Promoter DNA was amplified from genomic DNA using the following primers: NR-31UPGF (5′-TAA CTC GAG GAC GCA GGA AAG TCG GCA GTA GG-3′), as the 5′ upstream primer and NR-31-ExonI (5′-TCA CCC GGG TAC TCC CAA TCT TCG A-3′) as the 3′ downstream primer. Amplified DNA was inserted into the L3691 GFP reporter vector (from Andy Fire, Stanford University).
The Pnhr-31::gfp reporter vector was introduced into N2 worms by microinjection at a concentration of 50 ng/µl, worms were selected by EC fluorescence and no co-injection marker was used. Worms harboring the Pnhr-31::gfp transgene were examined by both standard fluorescence microscopy and confocal microscopy. Images were taken using AxioVision 4.6 software in multi-channel acquisition mode with an AxioCam MRU camera (Carl Zeiss Microimaging). For observation, larval or adult worms were mounted on glass slides with 2% agarose pads containing azide. Stack images of animals treated with nhr-31, vha-5, vha-8 and vha-12 RNAi were taken in both the FITC channel (488 nm) and DIC channels.
To measure excretory cell diameter, control and nhr-31 RNAi animals expressing Pnhr-31::gfp (5–6 animals) were analyzed by taking images which captured 50–100 µm of the proximal, middle, or distal regions of the posterior excretory cell tube. Diameter measurements were taken every 4–5 µm within the imaged regions using Zeiss measurement software. Data were plotted using Origen 5.0 (OrigenLab, Northampton, MA) software, and data displayed in dot plots reflected values from each independent measurement, along with the mean, and standard deviation from the mean.
Day 2 adults or L1 larvae were placed into a 20% BSA/PBS buffer solution and prepared in a Leica-Impact-2 high-pressure freezer according to the following protocol: 1) 60 hours in 100% acetone and uranyl acetate at −90°C. 2) Temperature was ramped from −90°C to −25°C over the course of 32.5 hours. 3) Next, sample was incubated at −25°C for 13 hours. 4) Next, the temperature change was brought from −25°C to 27°C in a 13 hour temperature ramp. Serial sections were post-stained in uranyl acetate followed by lead citrate. Thin cross sections were taken from resin-embedded clusters of young adults or L1 larvae. Sections for nhr-31 RNAi and control RNAi adult animals were obtained from 5 different animals, and sections for L1 larvae were also taken from 5 independent animals.
Synchronized L1 populations were prepared by hypochlorite bleaching of gravid N2 adults according to established protocols [6]. Synchronized L1 larvae were grown on control RNAi bacteria or nhr-31 RNAi until animals reached the early L4 stage of development. Worms were then harvested in M9, washed five times and immediately frozen in liquid nitrogen. RNA was extracted using a TRIZOL based method as described [6]. RNA was then labeled with Cy3 or Cy5 and hybridized to Washington University manufactured C. elegans microarrays (http://genome.wustl.edu). Data were obtained from three independent biological replicates and analyzed using GenePix Pro 6.0 software (Molecular Devices, Sunnyvale, CA). Ratios were calculated using background corrected, and normalized data (global mean).
For QRT-PCR, RNA was extracted and cDNA was prepared using our previously published protocol [6] with the following exception: RNA was separated from genomic DNA with a Turbo DNA free prep kit from Ambion (Austin, TX). qPCR was performed using a BioRad iCycler (MyiQ Single Color, Bio-Rad Laboratories, Hercules, CA). The data were analyzed as previously described [6]. QRT-PCR primers amplified ∼100 base pair regions of NHR-31 target genes. Primers were designed using Primer3 software and calibrated by serial dilution of cDNA and genomic DNA. Primer sequences are available upon request.
Worms treated with control RNAi or nhr-31 RNAi from the L1 to L4 stage of development were plated on high salt (500 mM NaCl) NGM-Lite plates seeded with E. coli. After various periods of high salt exposure, worms were scored for the ability to survive when rescued to a standard salt plate. Data for each time point was obtained from 250 animals. For rescue, worms were collected from the salt plates using M9 buffer+300 mM NaCl and transferred to standard NGM plates containing 50 mM NaCl. Worms were scored for survival after 12 hours of recovery [28].
To identify putative nuclear receptor response elements (NREs), we use the online computer program NHR-scan (http://nhrscan.genereg.net), which was first presented in a study by Sandelin and Wasserman [34]. The promoters of C. elegans vATPase genes were defined as the sequence between the ATG translational start site of the vha gene of interest and the beginning or end of the next upstream gene in the C. elegans genome. For vATPase genes expressed in operons, the promoter was chosen using the ATG translational start site of the first gene in the operon. For mouse promoters, 2000 nucleotides of upstream sequence were extracted from each vATPase gene. This sequence included 1950 nucleotides upstream of the translational start site +50 nucleotides of coding sequence. In all cases, the isoform with the most 5′ translational start site was selected for promoter sequence extraction. To calculate and display the consensus half sites shown in Figure 6, all half site sequences were analyzed using the WebLogo online program (http://weblogo.berkeley.edu/logo.cgi) [43].
|
10.1371/journal.pntd.0002307 | Cathelicidin-like Helminth Defence Molecules (HDMs): Absence of Cytotoxic, Anti-microbial and Anti-protozoan Activities Imply a Specific Adaptation to Immune Modulation | Host defence peptides (HDPs) are expressed throughout the animal and plant kingdoms. They have multifunctional roles in the defence against infectious agents of mammals, possessing both bactericidal and immune-modulatory activities. We have identified a novel family of molecules secreted by helminth parasites (helminth defence molecules; HDMs) that exhibit similar structural and biochemical characteristics to the HDPs. Here, we have analyzed the functional activities of four HDMs derived from Schistosoma mansoni and Fasciola hepatica and compared them to human, mouse, bovine and sheep HDPs. Unlike the mammalian HDPs the helminth-derived HDMs show no antimicrobial activity and are non-cytotoxic to mammalian cells (macrophages and red blood cells). However, both the mammalian- and helminth-derived peptides suppress the activation of macrophages by microbial stimuli and alter the response of B cells to cytokine stimulation. Therefore, we hypothesise that HDMs represent a novel family of HDPs that evolved to regulate the immune responses of their mammalian hosts by retaining potent immune modulatory properties without causing deleterious cytotoxic effects.
| In mammals, secreted host defence peptides (HDPs) protect against a wide range of infectious pathogens. They also perform a range of immune modulatory functions which regulate the immune response to pathogens, ensuring that the protective inflammatory response is not exacerbated and that post-infection repair mechanisms are initiated. We identified a novel family of molecules secreted by medically-important helminth pathogens (termed helminth defence molecules; HDMs) that exhibit striking structural and biochemical similarities to the HDPs. To further investigate the extent of this similarity, we have performed a comparative functional study between several well characterized, anti-microbial, mammalian HDPs and a series of parasite-derived peptides. The parasite HDMs displayed immune modulatory properties that were similar to their HDP homologs in mammals, but possessed no antimicrobial or cytotoxic activity. We propose that HDMs of these helminth pathogens underwent specific adaptation, losing their anti-microbial activity but retaining their ability to regulate the immune responses of their mammalian hosts. This absence of cytotoxicity and retention of immune-modulatory activity offers an opportunity to design novel immunotherapeutics derived from the HDMs which could be used to combat destructive inflammatory responses associated with microbial infection and immune-related disorders.
| Host defence peptides (HDPs) are found in all living organisms and play a pivotal role as effector components of the innate immune system [1], [2]. They act as the first line of defence against pathogenic assaults from bacteria, fungi, eukaryotic parasites and viruses [3]–[5]. A range of HDPs with varied sequence lengths, structures and activities have been characterized [6] and since sequence identity between them is often very poor, their classification is based largely on homologous secondary structures. The two predominant HDP groups found in nature are the cathelicidins, characterized by α-helical secondary structure, and the defensins, which contain β-sheets stabilized by intra-molecular disulfide bridges [7]–[9]. Despite the diversity in their sequences and structures, HDPs are typically small amphipathic peptides (12–50 amino acids) with a net positive charge (+2 to +9) and consist of at least 50% hydrophobic amino acids [10]. These biochemical properties are central to the HDPs antimicrobial function by allowing their interaction with, and disruption of, negatively charged bacterial membranes [10].
The contribution of mammalian HDPs to the innate immune response extends beyond direct bacterial killing. The elevated expression of HDPs in response to damage (injury or infection) has led to the suggestion that mammals utilize these peptides as ‘alarmins’ to activate the mobilization of a comprehensive immune response [11]. Besides their antimicrobial activity, HDPs function as potent immune regulators, selectively altering host gene expression, inducing chemokine production, inhibiting bacterial- or hyaluronan-induced pro-inflammatory cytokine production, promoting wound healing and modulating T and B cell function [reviewed in [12]–[14]. The net result of these activities is a balance between pro- and anti-inflammatory immune responses which prevents an exacerbated inflammatory response while concurrently stimulating the resolution of infection and repair of damaged epithelia.
The immune response elicited by helminth (worm) parasites is akin to the innate immune response to tissue injury and wound healing [15], [16]. Typically, this consists of a suppression of classical pro-inflammatory responses and the induction of anti-inflammatory regulatory Th2 type immune responses. While classical Th1-associated inflammatory mediators can provide protection from helminths [17], there is a substantial cost in collateral damage to host tissue [15], [18]. In addition, due to their migration and feeding activities, helminth parasites cause considerable local tissue damage. Therefore, it has been proposed that on exposure to helminths, the most beneficial outcome is to shut down a destructive Th1-type response in favour of a Th2 response that rapidly and effectively heals tissue [15], [17], [18]. Ultimately this means that the parasite is tolerated by the host, remaining in situ for many years and thus successfully completes its lifecycle.
Some advances have been made in identifying the signalling molecules that initiate helminth-associated Th2 responses. Many of these (such as IL-33 and Thymic stromal lymphopoietin (TSLP)) are thought to be released by epithelial cells damaged by migrating parasites [15], [19]. However, a number of helminth-derived products have also been shown to modulate the function of innate immune cells and thus are potentially instrumental in the initiation of Th2 immune responses [19], [20]. We have previously shown that a cysteine protease secreted by the trematode helminth Fasciola hepatica prevented the induction of pro-inflammatory macrophages and dendritic cells [21], [22]. In addition, peroxiredoxin, also secreted by F. hepatica, promoted the development of Th2 host immune responses [23], [24]. Importantly, homologues of these proteins are found in other medically-important trematode parasites which we have suggested reveals a common mechanism of immune-modulation employed by this class of pathogen.
As part of our on-going analysis of the secretome of F. hepatica we recently discovered a novel 8 kDa protein. On analysis, this protein shared structural and biochemical similarities to mammalian cathelicidins and was therefore termed F. hepatica Helminth Defence Molecule 1 (FhHDM-1) [25]. Like the human cathelicidin LL-37 precursor CAP-18, FhHDM-1 is proteolytically processed (by a parasitic endopeptidase, cathepsin L1) to release a 34-residue C-terminal peptide previously named FhHDM-1 p2 [25]. This peptide adopts an amphipathic helix structure and, like LL-37, can bind to Escherichia coli lipopolysaccharide (LPS) to prevent its interaction with the toll-like receptor (TLR) 4/MD2/CD14 complex on macrophages. Hence we proposed that F. hepatica utilized FhHDM-1 as a molecular mimic of mammalian cathelicidin-like HDPs as a means of controlling host innate immune responses [25].
Phylogenetic analysis showed that FhHDM-1 is a member of a family of HDMs conserved throughout several major animal and human trematodes such as Schistosoma, Fasciola, Opisthorchis, Clonorchis and Paragonimus species [25]. Importantly, all HDM molecules in this family have preserved the C-terminal amphipathic helix. Here, we have performed a comparative functional study between several anti-microbial HDPs derived from well-characterized mammalian cathelicidins and parasite-derived peptides. For the helminth-derived peptides we selected FhHDM-1p2 and two homologs derived from Schistosoma mansoni which we term S. mansoni HDM-1 (SmHDM-1p146) and HDM-2 (SmHDM-2p58). In addition, we included a peptide derived from a previously characterized secretory molecule of S. mansoni, termed Sm16-p73 [26], which our phylogenetic studies suggest is a divergent member of the HDM superfamily [25]. We show that, in contrast to the mammalian HDPs (LL-37, CRAMP, BMAP-28 and SMAP-29), the trematode-derived HDMs are not cytotoxic or bactericidal. However, like the mammalian HDPs, the trematode HDMs suppress the activation of macrophages by microbial stimuli and alter the isotype of immunoglobulin secreted by B cells. We propose that HDMs represent a novel family of HDPs that have undergone specific adaptation to retain potent immune modulatory properties in the absence of deleterious cytotoxic effects and are exploited by helminth pathogens to regulate the immune responses of their mammalian hosts.
Four synthetic cathelicidin-derived peptides from diverse mammalian species were used: SMAP-29 from sheep [27], [28], CRAMP from mice [29], LL-37 from human [30], and BMAP-28 from cattle [31]. The 34-residue C-terminal FhHDM-1 peptide, termed FhHDM-1p2, has been previously described [25]. Sm16-p73 peptide is 35 residues in length, corresponding to residues 73–107 of the full-length protein S. mansoni Sm16 (GenBank accession number: AAD26122.1). SmHDM-1p146 is 35 residues in length and corresponds to residues 146–180 of the full length protein (GenBank accession number XP_002580563.1). Finally, SmHDM-2p58 is a 32 residue peptide that corresponds to residues 58–98 from the full length protein (GenBank accession number: XP_002576627.1). All mammalian and trematode peptides were synthesized by GenScript (NJ, USA) and supplied endotoxin-free. The single letter code sequence of each peptide is shown in Table 1.
The biochemical characteristics for each of the HDMs and HDPs were calculated using tools available from the Antimicrobial Peptide Database (http://aps.unmc.edu/AP/main.php) [32] and are presented in Table 1. Predicted properties were total net charge, Boman index, hydrophobic ratio and total hydrophobic residues on the same hydrophobic surface of the alpha helix.
We have previously shown, using circular dichroism (CD) spectroscopy, that FhHDM-1 has the propensity to adopt alpha-helical structure in solution [25]. To assess whether HDMs from the related trematode parasite S. mansoni also form alpha helices, secondary structure prediction was performed using JPred 3 ([33]; http://www.compbio.dundee.ac.uk/www-jpred/). Specific regions predicted to form alpha helices were then subjected to helical wheel analysis using Heliquest ([34]; http://heliquest.ipmc.cnrs.fr/cgi-bin/ComputParams.py) to identify those with distinct hydrophobic faces; i.e. are amphipathic. The atomic structures of the vertebrate HDPs LL-37 (PDB ID: 2K6O) and BMAP-28 (PDB ID: 2KET) were visualised for comparison using the PyMOL Molecular Graphics System, Version 1.5.0.4 Schrödinger, LLC. (http://pymol.org/).
Lipopolysaccharide (LPS) binding was performed using a quantitative chromogenic Limulus amoebocyte assay (Chromo-LAL assay; Associates of Cape Cod Incorporated) following manufacturer's recommendations. Assays were performed in flat-bottom endotoxin- and glucan-free 96-well plates (Associates of Cape Cod Incorporated). Stock solutions of each peptide were prepared in endotoxin-free water (80 µg/ml) and diluted to a final concentration of 250 pmol/ml. In the first step, 25 µl of peptide solution was mixed with 25 µl of a solution containing 1 endotoxin U/ml of Escherichia coli O113:H10 LPS and incubated for 30 min at 37°C to allow peptide and LPS binding to occur. The second step involved the addition of 50 µl of the chromo-LAL reagent. The liberation of ρ-nitroaniline was monitored every 60 sec at 405 nm with a Synergy H1 hybrid reader (Biotek) while the temperature was maintained at 37°C. Each peptide concentration was also incubated with 25 µl of LPS-free water as a control to determine if the peptide itself could activate the Chromo-LAL assay. The experiment was conducted twice in triplicate. Standard deviation was calculated from these six replicates.
The minimal inhibitory concentration (MIC) of each peptide against various bacteria was determined using a standardized dilution method according to NCSLA guidelines [35]. Overnight colonies of E. coli, Pseudomonas aeruginosa, Salmonella typhimurium, Staphylococcus epidermis and Staphylococcus aureus, were suspended to a turbidity of 0.5 OD units and further diluted in Mueller-Hinton broth (MHB). For determination of MIC, peptides were prepared in an acetic acid/BSA solution and used in graded concentrations (0, 1, 2, 4, 8, 16, 32, 64, and 128 µg/ml) from a stock solution. Ten microliters of each concentration was added to each corresponding well of a 96-well polypropylene microtiter plate and 1×105 bacteria in the volume of 90 µL. The plate was incubated at 37°C for 16 h and then read at 600 nm with a Synergy H1 hybrid reader (Biotek).
The peptide's haemolytic activities were determined using human red blood cells (RBCs; Research Blood Components, LLC) in 96-well polypropylene microtiter plates. One hundred µl of 0.5% RBC suspension was added to an equal volume of a peptide (8–256 µg/ml). After 1 h at 37°C, plates were centrifuged at 1420×g for 5 min and the optical density of the supernatant was measured at 414 nm with a Synergy H1 hybrid reader (Biotek). Values for 0% and 100% lysis were obtained by adding PBS or Triton X-100 (1%; final concentration) to RBCs, respectively. All assays were performed in triplicate and the values of percent lysis were within a 1% standard deviation range.
RAW 264.7 murine macrophages (5×105 cells) were incubated in the presence of the fluorescent dye TO-PRO (Life Technologies) for 60 sec. After the peptides (50 µM) were added to the culture media the uptake of dye was measured for 360 sec by flow cytometry.
RAW 264.7 murine macrophages (1×106cells) were incubated with a range of concentrations (2.5–50 µM) of peptides for 1 h at 37°C. The culture supernatants were collected and assayed for LDH activity with the CytoTox LDH release kit (Promega) according to the manufacturer's instructions. The amount of LDH released is expressed as a percentage of the total amount of LDH released from cells treated with lysis buffer (regarded as 100% cytotoxicity).
Oocysts of the Iowa C. parvum isolate [36] were propagated in experimentally infected newborn Cryptosporidium–free Holstein bull calves to obtain parasite material for study as previously described [37], [38]. Oocysts were isolated by sucrose density gradient centrifugation, stored in 2.5% (W/V) potassium dichromate (4°C) and used within 6 weeks of isolation [39]. Oocysts of the TU502 C. hominis isolate [40], [41] were propagated in gnotobiotic piglets and isolated from feces at Tufts University [42] and used within 4 weeks of isolation. Prior to excystation, oocysts were treated with hypochlorite [37]. In vitro excystation (37°C, 0.15% [W/V] taurocholate, 2 h) of oocysts used for all experiments was ≥90%. Sporozoites were isolated from excysted oocyst preparations by passage through a polycarbonate filter (2.0 µm pore size; Poretics, Livermore, California) and used immediately.
Sporozoite viability after incubation with peptides was assessed using fluorescein diacetate (FDA) and propidium iodide (PI) with modification [43]. In brief, freshly excysted sporozoites were incubated (15 min, 37°C) in minimal essential medium (MEM) containing individual peptides (2.5, 0.25, 0.025 µM) or in MEM alone (n = 3). Peptide concentrations were selected based largely on studies by our group and others evaluating the effects of various antimicrobial peptides on C. parvum viability [44]–[47]. Heat-killed (20 sec, 100°C) sporozoites were used as a control. FDA (8 mg/ml final concentration) and PI (3 mg/ml final concentration) were then added to the sporozoite preparations and incubated further (5 min, 21°C). A minimum of 100 sporozoites were then examined by epifluorescence microscopy for each preparation, and the percent viability was determined. Percent reduction of viability was calculated as ([MEM-treated sporozoite viability−peptide-treated sporozoite viability]÷MEM treated sporozoite viability)×100. The mean values for test and control preparations were examined for significant differences using Student's t-test.
CD11b+F4/80+ macrophages were derived (99% purity) from the bone marrow of BALB/c mice by culturing with M-CSF (ebioscience) over 6 days and then seeded in RPMI (with 10% FBS v/v) at a concentration of 1×105 cells/ml. These cells were incubated with peptides (0.5 µM–5 µM) for 1 h at 37°C, in the absence of mCSF. Following two washes with ice-cold PBS, cells were incubated with a combination of E. coli LPS (10 ng/ml; Sigma) and IFNγ (10 ng/ml; BD Pharmingen) overnight. Supernatants were then collected and the amount of TNF measured by ELISA according to the manufacturer's instructions (BD Pharmingen).
B cells were isolated from the spleens of BALB/c mice by negative selection using a B cell isolation kit containing biotin-conjugated mAbs to CD43, CD4, and Ter-119 (Miltenyi Biotec) and then seeded at a concentration of 1×106 cells/ml in RPMI (with 10% FBS v/v). The cells were treated with a range of concentrations of peptides (0.5 µM–5 µM) for 1 h at 37°C. After washing, the B cells were incubated with a combination of either E. coli LPS (10 µg/ml; Sigma) and IL-4 (10 ng/ml; BD Pharmingen) or E. coli LPS (10 µg/ml) and IFNγ (200 ng/ml; BD Pharmingen) for six days. Supernatants were then collected and the amount of IgG1 or IgG2a measured by ELISA (Sigma).
Statistical comparisons were performed with Prism 4.0 Software (Graph- Pad), using two-tailed Student's t test for comparisons of two data sets, and ANOVA for multiple comparisons. Statistically significant differences were determined by a p value of *<0.05, **<0.01, ***<0.001.
The biochemical properties for each of the HDMs and HDPs are presented in Table 1. The cathelicidins are known to be highly cationic peptides. Except for the FhHDM-1p2 peptide, which has a net charge of 0, all the peptides in the present study have a net positive charge (+3 to +9) with a percentage of hydrophobic residues ranging from 34 to 44. The Boman index is an estimated potential of peptides to bind to other proteins. For this index, a low value (≤1) suggests that a peptide has more antibacterial activity, whereas values ranging from 2.5–3.0 indicate that a peptide is multifunctional with hormone-like activities [48]. While BMAP-28 has a low Boman index (0.81) which correlates to its high antimicrobial activity, the Boman indices for the other peptides range from 1.34–3.11. Overall, there is no striking difference between the biochemical properties of HDMs and HDPs.
The atomic structures of the vertebrate HDPs LL-37, CRAMP, BMAP-28 and SMAP-29 have been previously solved and determined that each form an amphipathic helix [31], [49]–[51] (figure 1). The structures of LL-37 and BMAP-28 are shown in figure 1A as representative for this group of peptides. Secondary structure prediction of the parasite-derived peptides FhHDM-1p2, SmHDM-1p146, SmHDM-2p58 and Sm16-p73 revealed that all possessed regions likely to form alpha helices. Furthermore, helical wheel analysis showed that each molecule contained an alpha helix (32–35 amino acids) toward the C-terminal that was distinctly amphipathic. The number of residues forming the hydrophobic face of the parasite molecules ranged from 6–9 (figure 1B). Thus, like their vertebrate HDP counterparts, the secreted helminth parasite molecules also form distinct amphipathic helices.
Mammalian HDPs have the ability to bind to and thus neutralize the bacterial endotoxin LPS [52]–[57]. The capacity of different mammalian and helminth peptides to bind LPS from E. coli O113:H10 was compared using the chromogenic Limulus amoebocyte assay (Chromo-LAL). The Limulus amoebocyte lysate contains enzymes that are activated in a cascade of reactions in the presence of LPS [58]–[60]. The final enzyme in the series splits the chromophore, ρ-nitroaniline (ρNA), from the chromogenic substrate, generating a yellow color. The amount of ρNA released is proportional to the amount of free LPS present in the system.
Consistent with that published in the literature, mammalian HDP LL-37, BMAP-28, SMAP-29 and CRAMP inhibited the activation of the Chromo-LAL assay at a concentration of 250 pmol/ml, indicating that they interact with LPS and thus prevent the activation of the enzymatic cascade [52], [53], [55], [61] (figure 2). BMAP-28 was the most potent, preventing the activation of the enzyme cascade with a Vmax of 3.1 and SMAP-29 was less effective with a Vmax of 26.5. Of the helminth-derived HDPs, both Sm16-p73 and SmHDM-1p146 displayed no LPS-binding capacity with a Vmax of 40.4 and 31.1, respectively, which were above the control reaction with no peptide (Vmax of 28.5). By contrast, the helminth peptide SmHDM-2p58 was the second best inhibitor of the series with a Vmax of 12.5, just after BMAP-28.
We found that FhHDM-1p2 was capable of directly activating the Chromo-LAL assay itself (data not shown) and, therefore, its binding capacity could not be evaluated using this test. However, using a plate-binding assay we have previously demonstrated that FhHDM-1p2 does indeed bind to LPS [25].
The bactericidal properties of mammalian HDPs are mediated by direct antimicrobial activities, and are therefore easily evaluated as the minimal concentration capable of inhibiting visible microbial growth (MIC) against a panel of bacterial species. Consistent with previous reports [28], [31], [62], [63], we found that SMAP-29 and BMAP-28 were effective against a broad group of gram-negative bacteria, including E. coli, P. aeruginosa, and S. typhimurium and two gram-positive bacteria, S. aureus and S. epidermis, with MIC values of <0.25–8 µg/ml (Table 2). LL-37 and its mouse counterpart, CRAMP, showed bactericidal activity against gram-negative bacteria with MIC ranging from 2 to 8 µg/ml, but were ineffective against the gram-positive bacteria tested (MIC<128 µg/ml). The inactivity of LL-37 against S. aureus and S. epidermis is consistent with other studies [64]. However, the anti-microbial effect of LL-37 on Staphylococcus could be strain dependent as several studies have reported an effect of this peptide on both S epidermis [65] and S. aureus [66], [67]. Despite structural similarities with the mammalian peptides tested, none of the HDMs demonstrated bactericidal activity against any species of bacteria at the concentrations tested (<0.25 to 128 µg/ml).
We have recently shown that cationic peptides, including the cathelicidin LL-37, are highly parasiticidal against the apicomplexan parasite C. parvum in vitro [47]. Using the same methodology, we compared the anti-parasite activity of the mammalian HDPs to that of the four helminth-derived peptides. In keeping with our previous data [35], LL-37 exhibited parasiticidal activity at a concentration of 2.5 µM against C. parvum and the related species C. hominis (figure 3). The other mammalian cathelicidins showed greater parasiticidal activity, reducing the viability of protozoans at lower concentrations of 0.025 and 0.25 µM. Of particular note, BMAP-28 demonstrated the highest potency against both species of protozoan at 2.5 µM (P<0.01). In stark contrast to these results, was the relative absence of parasiticidal activity of the helminth-derived peptides, even at the highest concentration of 2.5 µM.
The predominant mechanism of HDP bactericidal activity is the formation of pores in the membrane lipid bilayer, destroying its integrity and causing cell death [68]. However, this effect is not specific to bacterial cells and HDPs have also been shown to be cytolytic to eukaryotic cells, particularly at high concentrations [69]. TO-PRO is a membrane impermeant dye and therefore its detection within cells is indicative of pore formation. Using this dye, we demonstrated that, as expected, all mammalian peptides at a concentration of 50 µM (equivalent to the highest concentration tested in the bactericidal assays) induced the formation of pores in a murine macrophage cell line (figure 4). However, at the same concentration none of the helminth peptides exhibited this effect.
To more completely assess the cytotoxicity of the peptides, we first examined their haemolytic activity against human RBCs at various concentrations (8–256 µg/ml). After one hour of co-incubation, all of the mammalian peptides induced concentration dependent hemolysis (Table 3). BMAP-28 was the most potent of all the peptides, with 50% of RBCs lysed at the lowest concentration of peptide tested (8 µg/ml) and 70% at the highest concentration of 256 µg/ml. In comparison, at this highest concentration, the other mammalian peptides were much less cytotoxic, lysing only 14.5–29.3% of RBCs. Notably, under the same experimental conditions, the S. mansoni-derived peptides did not lyse the cells at any concentration tested. The F. hepatica-derived FhHDM-1p2 showed some low-level cytolytic activity, with 11.4% of cells lysed at the highest concentration of 256 µg/ml.
Lactate Dehydrogenase (LDH) is a soluble cytosolic enzyme that is released into culture media following loss of membrane integrity resulting from either apoptosis or necrosis. Therefore LDH release is an indicator of cell membrane integrity and acts as a measure to assess cytotoxicity. Consistent with the demonstration of hemolysis, higher concentrations (>10 µM) of mammalian peptides also resulted in the death of murine macrophages, with the highest concentration tested (50 µM) resulting in 100% cytotoxicity, compared to the effect of a lysis buffer (figure 5). By contrast, none of the helminth peptides induced cell death at any concentration tested. This lack of LDH detection was not due to enzyme inhibition by the helminth peptides; when the peptides were added directly to culture supernatant from lysed cells the LDH activity was unchanged. Consistent with these data, the cells treated with helminth peptides (10–50 µM) looked morphologically normal by light microscopy (data not shown).
Activation of macrophages by microbial stimuli is central to the induction of innate immune responses. IFNγ is one of the key cytokines in the innate immune response to intracellular pathogens, and augments cellular responses to TLR ligands such as bacterial LPS [70], [71]. To prevent excessive inflammation potentially leading to sepsis, HDPs have been shown to inhibit the response of macrophages to these inflammatory mediators using mechanisms that are independent of direct binding to LPS [72]. Significantly, macrophages isolated from animals and humans infected with helminth parasites are also hyporesponsive to stimulation with LPS and IFNγ [73], [74]. Therefore, here we investigated whether helminth-derived peptides, like mammalian HDPs, could suppress the activation of inflammatory macrophages. At concentrations below cytotoxic levels (<5 µM), all mammalian peptides significantly inhibited TNF production in response to the combined stimulation with LPS and IFNγ (figure 6). Titration of the peptide concentration showed that even at concentrations as low as 0.5 µM, the inhibitory activity was preserved. In contrast, at the lowest concentration tested, the helminth-derived peptides had no effect on the activation of macrophages. However, as the concentration was increased, the helminth peptides significantly suppressed the inflammatory response of macrophages and in most cases more effectively than the mammalian peptides (figure 6). It is worth noting that the helminth-derived peptides could be tested at concentrations up to 50 µM as they are non-toxic to cells, whereas due to their cytotoxicity the HDPs were not tested at concentrations above 10 µM (figures 4,5). At these higher concentrations (10, 25 and 50 µM) the helminth-derived peptides significantly reduced the production of TNF from activated macrophages in a concentration dependent manner (data not shown).
In addition to directly inhibiting inflammatory innate immune responses, there is evidence that mammalian HDPs have an additional role in regulating the magnitude of the adaptive antibody responses. For example, it has been shown that CRAMP functions to positively regulate the level of IgG1 produced by B cells [75], and LL-37 reportedly decreased the production of IgG2a from mouse splenic B cells activated with LPS and IFNγ [76]. Consistent with these reports, our analyses showed that with the exception of BMAP-28, all the mammalian HDPs significantly increased the production of IgG1 in response to a Th2 biased environment (LPS and IL-4) (figure 7A). The apparent reduction in IgG1 production recorded for the higher concentration of BMAP-28, likely reflects some level of cell death rather than a reduction in antibody production. Due to this cytotoxicity the HDPs were not tested at concentrations above 5 µM. While there was greater variability between the HDMs, peptides from F. hepatica and S. mansoni significantly increased the production of IgG1 in response to LPS and IL-4 even at low concentrations.
Conversely, mammalian peptides reduced the production of IgG2a in a Th1 (LPS and IFNγ) biased environment (figure 7B). For the helminth peptides, only concentrations above 5 µM had the same effect on B cells, significantly (p<0.001) inhibiting IgG2a secretion (data not shown), suggesting a lower potency than mammalian-derived peptides.
Parasitic helminths secrete molecules that modulate host immune responses to establish an environment that facilitates their survival and a prolonged reproductive phase [20], [77], [78]. Co-evolution of helminths with their hosts means that these parasites are well adapted to the host's immune system, making use of endogenous regulation mechanisms to manipulate the immune response to their benefit. In this study, we compared the biological activities of a series of helminth-derived cathelicidin-like peptides to that of their mammalian homologues and suggest how their production by helminths can facilitate a successful parasitic life cycle.
In vitro, most mammalian HDPs are effective antimicrobial agents against a range of organisms including gram-negative and gram-positive bacteria, protozoa, viruses and fungi [10], [79], [80]. In general, the expression of HDPs is increased at the onset of an infection and therefore the anti-pathogenic activity was thought to be one of the most important immediate responses that the mammalian host evolved to deal with invading pathogens. It has been proposed that the specificity of HDPs for particular microbes is subjected to significant variation and is particularly influenced by the types of microbial biotas to which each HDP species is exposed [81]. Consistent with this theory, we showed some variation in the anti-microbial capabilities of the mammalian HDPs examined. While sheep and bovine derived peptides were effective against a broad range of both gram-positive and gram-negative bacteria, the mouse and human HDPs were largely ineffective against the gram-positive species tested. However, we found that none of the helminth-derived peptides displayed gram positive or negative bactericidal activity, even at the highest concentration tested, implying that their specialised function is not anti-microbial. However, we cannot exclude the possibility of the peptides having anti-microbial activity on other, untested pathogenic bacteria.
Despite lacking bactericidal activity, we showed previously that FhHDM1-p2, like the mammalian peptides, interacts with LPS, thus effectively neutralising the ability of infecting bacteria to induce an inflammatory response [25]. We suggested that this may be a mechanism used by the parasite to prevent excessive activation of innate cells in response to the translocation of microbes into circulation occurring as a result of damage to the skin and/or gut epithelium during migration of the parasite [25]. As the two major factors mediating interaction between LPS and HDPs are hydrophobicity and cationicity [82], inspection of the sequence of the other HDMs would predict a universal ability to bind to LPS. However, while SmHDM-1p146 appeared to bind LPS as efficiently as the mammalian peptides, neither SmHDM-2p58 or Sm16-p73 were particularly potent, indicating that the neutralization of LPS may not be a common function of the helminth-derived peptides.
A number of studies have demonstrated the ability of amphibian and mammalian HDPs to kill protozoan parasites in vitro [83], [84]. For example, BMAP-28 possesses potent activity against the agent of human leishmaniasis, Leishmania major [83], and we have shown that LL-37 can reduce the viability and infectivity of sporozoites of C. parvum, an intestinal infection of humans and agricultural animals [47]. In the present study we confirm the activity of LL-37 against C. parvum and the related parasite C. hominis and show that the other mammalian cathelicidins tested also have anti-protozoan activity; BMAP-28 exhibited the most potent in vitro activity. By contrast, and consistent with their lack of antibacterial activity, the helminth-derived HDMs did not kill either parasite in vitro. Clearly, the particular physico-chemical properties of HDMs do not confer an ability to penetrate and disrupt the surface membrane of these parasitic organisms.
Although widely defined as antimicrobial, in fact, at the concentrations normally found at human mucosal surfaces and in physiological salt conditions, the mammalian cathelicidin-like peptides do not display bactericidal activity. However, at these same concentrations and under the same conditions, the peptides exhibit a variety of immune modulatory functions [12], [85]. This has led to the suggestion that the cathelicidins are principally immune modulators rather than antimicrobials, and like the mammalian defensins, have traded their bactericidal capacities to acquire the ability to broadly regulate the immune response [12], [86]. The mammalian cathelicidins have a diverse effect on the cellular immune response, but in particular, the peptides have a crucial role in regulating TLR-dependent innate inflammatory responses. This means that they function to maintain homeostasis in response to natural shedding of microflora-TLR agonists as well as controlling the systemic inflammatory response to infection or tissue damage [12], [86], [87]. We have shown here that the helminth-derived cathelicidin-like peptides also regulated the innate immune response to TLR stimulation by inhibiting TNF release from macrophages stimulated with bacterial LPS. While the helminth HDMs were not as effective as their mammalian homologues at the lowest concentration tested, they were correspondingly more potent as their concentration was increased towards quantities that are likely secreted during infection. It is probable that the function of these secreted HDMs is similar to the predicted role for the mammalian HDPs, i.e. prevention of an excessive inflammatory response, which acts to prevent the expulsion of the worm and to protect the host from exacerbated tissue damage.
The role of HDPs in regulating the adaptive immune response has been less extensively studied. Recent studies have shown that LL-37 decreased the production of IgG2a from murine B cells stimulated with LPS and IFNγ [76], and that CRAMP increased the amount of IgG1 in response to IL-4 [75]. These results are consistent with the suggestion that HDPs are engaged in the process of infection resolution and wound healing, as autoreactive IgG1 antibody production is central to tissue repair processes [88], while IgG2a are associated with IFNγ/Th1-mediated inflammatory responses. Consistent with these reports, the helminth-derived peptides performed in a similar manner to their mammalian counterparts, successfully enhancing the production of IgG1. At high concentrations, the helminth-derived peptides were far more effective than mammalian HDMs at suppressing the release of IgG2a in response to IFNγ, supporting their role as potent anti-inflammatory agents.
Due to their immune modulatory activities, there is considerable interest in developing HDPs as therapeutics such as anti-inflammatory agents, adjuvants and wound healing agents. The therapeutic potential of HDPs has been demonstrated and a number of peptides are being developed as anti-inflammatory agents [89]. However, the clinical use of these peptides as injectable therapeutics has been hampered by indications of toxic side-effects on mammalian cells and their ability to lyse eukaryotic cells [90], [91]. This had led to intense research into understanding how HDPs function in terms of their physico-chemical properties. Among the factors that appear to influence specificity between their activity against prokaryotic and eukaryotic cells are the ability to form an amphipathic α-helical structure, hydrophobicity, overall charge distribution, and minimal peptide length [92]. At first sight, the biochemical characteristics of the helminth-derived HDMs would predict an inherent cytotoxic activity: they form amphipathic helices (figure 1), they have a comparable proportion of hydrophobic amino acids to mammalian HDPs and most of them are cationic (Table 1). However, based on the assays employed in this study, we found no correlation between the level of hydrophobicity and cytotoxic activity.
The majority of mammalian HDPs have an overall net charge ranging from +4 to +6.37 [93], implying an optimal range for biological activity. HDPs with a net positive charge of <+4 are found to be inactive, whereas increasing the net charge from +4 to +8 confers antimicrobial activity and some haemolytic activity [79]. Three of the HDMs used in this study, FhHDM-1p2, SmHDM-1p146 and SmHDM-2p58, all had a net charge <+4 which is consistent with a non-bactericidal, non-haemolytic peptide. However, despite possessing a net charge of +5, Sm16-p73 displayed neither antimicrobial nor cytotoxic activities. Recent studies have proposed that rather than a simple correlation between net charge and haemolysis it is the localisation of the positively charged amino acids within the peptide that also dictates membrane interaction and selectivity. By increasing the charge of an amphipathic HDP analog from +8 to +9, by the addition of one positive charge on the polar face, the haemolytic activity of the peptide was enhanced 32-fold [79]. Using this parameter, we calculated that Sm16 has four positively charged amino acids on its polar face compared to six for LL-37, CRAMP and BMAP-28 and seven for SMAP-29, which may provide some clue as to the difference in cytotoxicity between these peptides. Likewise, all the helminth HDMs used in this study have less positive charges on their polar face compared to LL-37, which was the least cytotoxic of all the mammalian peptides examined in this study.
It is generally accepted that the cytoplasmic membrane is the main target of many mammalian HDPs, whereby peptide accumulation in the membrane causes increased permeability and a loss of barrier function, resulting in the leakage of cytoplasmic components and cell death [94]. However, the cytotoxic concentrations of HDPs are higher than the concentrations required for the destruction of microbes, which, some authors suggest, reveals a cell-selective killing mechanism [85], [94]. The physiological concentration of HDPs at mucosal sites is typically less than 2 µg/ml [12], [80], well below the concentration that is cytotoxic to mammalian cells in vitro. We have found that helminth HDMs are abundant molecules within the secretions of helminth parasites, which during a multi-parasite infection would likely be at relatively high concentrations in circulation. Therefore, it is essential for the success of the parasite that these peptides do not possess cytotoxic activity, whilst at the same time retain the beneficial immune modulatory properties. The complete absence of antimicrobial activity by helminth-derived peptides is likely linked to this need to prevent host cell death during infection.
The extraordinary capacity of helminths to regulate the immune response is central to their longevity in the mammalian host and thus underpins their success as parasitic organisms [77], [78]. Therefore, it is perhaps unsurprising that helminth secretory products contain homologues of components of the host immune system that target the same mammalian pathways. In addition to the HDMs identified here, helminth parasites express highly conserved cytokine gene families that, like their mammalian counterparts, ligate specific receptors on immune cells. Brugia malayi and Ancylostoma ceylanicum express homologues of the mammalian cytokine macrophage migration inhibitory factor (MIF) [95], and in a Th2 environment, such as that activated by helminth infection, Brugia MIF synergises with IL-4 to induce the development of regulatory M2 macrophages [96]. Helminths also express members of the Tumour Growth Factor-(TGF)β and TGF-β receptor superfamilies [97]–[99], and similar to the mammalian cytokine, Heligmosomoides polygyrus TGF-β homologue has been shown to directly induce the differentiation of regulatory T cells, demonstrating a key role in parasite immune regulation [100].
The cysteine protease inhibitors, cystatins, are an ancient and conserved family of peptides in the animal and plant kingdoms [101]. The cystatins of parasitic worms differ substantially from those produced by free-living nematodes with regard to their immune modulatory properties [102]. In particular, the acquisition of an asparaginyl endopeptidase site, similar to that of vertebrate cystatin C, confers an ability to reduce the activation of host T cell responses by directly inhibiting the presentation of antigen by dendritic cells [103], [104], suggesting a specific adaption to regulate host immune responses [102], [105]. Similar to the cystatins, HDPs are conserved in all organisms, including plants, animals and humans [106]. However, we show here that while the helminth-derived HDMs are effective immune modulators, they display no bactericidal activity. These observations would suggest that like the cystatins, HDMs have become specifically adapted to support a parasitic lifestyle, losing the more ancient property of direct antimicrobial killing but acquiring the ability to regulate immune responses in order to promote their survival within the mammalian host. It is possible that HDM immune modulation arose in trematodes following their divergence from the chordate lineage (as acoelomates) and their subsequent specialisation to an endoparasitic lifestyle, distinct from the free-living acoelomate turbellarian flatworms from which HDM homologues have yet to be identified (unpublished observation).
The immune-modulatory properties of mammalian HDPs, and in particular their ability to prevent excessive inflammatory induced pathology associated with bacterial sepsis, has attracted interest in exploiting these as anti-infectives. However, their cytotoxicity, as also shown in the present study, has presented a major drawback for their in vivo use. Accordingly, the absence of cytotoxicity and retention of immune-modulatory activity observed for the helminth-derived HDMs offer an opportunity to design novel immunotherapeutics to combat microbial pathogens and immune-related disorders.
|
10.1371/journal.pbio.1002224 | The Dynamics of Incomplete Lineage Sorting across the Ancient Adaptive Radiation of Neoavian Birds | The diversification of neoavian birds is one of the most rapid adaptive radiations of extant organisms. Recent whole-genome sequence analyses have much improved the resolution of the neoavian radiation and suggest concurrence with the Cretaceous-Paleogene (K-Pg) boundary, yet the causes of the remaining genome-level irresolvabilities appear unclear. Here we show that genome-level analyses of 2,118 retrotransposon presence/absence markers converge at a largely consistent Neoaves phylogeny and detect a highly differential temporal prevalence of incomplete lineage sorting (ILS), i.e., the persistence of ancestral genetic variation as polymorphisms during speciation events. We found that ILS-derived incongruences are spread over the genome and involve 35% and 34% of the analyzed loci on the autosomes and the Z chromosome, respectively. Surprisingly, Neoaves diversification comprises three adaptive radiations, an initial near-K-Pg super-radiation with highly discordant phylogenetic signals from near-simultaneous speciation events, followed by two post-K-Pg radiations of core landbirds and core waterbirds with much less pronounced ILS. We provide evidence that, given the extreme level of up to 100% ILS per branch in super-radiations, particularly rapid speciation events may neither resemble a fully bifurcating tree nor are they resolvable as such. As a consequence, their complex demographic history is more accurately represented as local networks within a species tree.
| The rise of modern birds began after the mass extinction of nonavian dinosaurs and archaic birds at the Cretaceous-Paleogene (K-Pg) boundary, about 66 million years ago. This coincides with the super-rapid adaptive radiation of Neoaves (a group that contains most modern birds), which has been difficult to resolve even with whole genome sequences. We reconstructed the genealogical fates of thousands of rare genomic changes (insertions of selfish mobile elements called retrotransposons), a third of which were found to be affected by a phenomenon known as incomplete lineage sorting (ILS), namely a persistence of polymorphisms across multiple successive speciation events. Astoundingly, we found that near the K-Pg boundary, speciation events were accompanied by extreme levels of ILS, suggesting a near-simultaneous, star-like diversification process that appears plausible in the context of instantaneous niche availability that must have followed the K-Pg mass extinction. Our genome-scale results provide a population genomic explanation as to why some species radiations may be more complex than a fully bifurcating tree of life. We suggest that, under such circumstances, ILS bears witness to the biological limitation of phylogenetic resolution.
| The rich biodiversity of many organismal groups is the result of bursts of rapid species diversifications, with extreme examples in angiosperms [1] and vertebrates [2]. Among the latter, birds are one of the most speciose groups with a total of >10,500 recognized species that are proposed to be the result of mostly recent accelerations of diversification rates [3]. Nevertheless, the deep roots of 95% of these species lie within the ancient adaptive radiation of Neoaves, comprising all contemporary avian lineages except Palaeognathae (ratites and tinamous) and the Galloanserae (chicken and ducks). This massive radiation exhibits the highest known diversification rate among deep vertebrate radiations [2], coincides with the Cretaceous-Paleogene (K-Pg) boundary, and gave rise to 36 extant bird lineages within <15 million years (MY) [4]. Simulations suggest that the distribution of neoavian internode lengths causes a very high probability of gene tree–species tree incongruences [5], i.e., hemiplasy derived from incomplete lineage sorting (ILS) [6]. ILS denotes the persistence of ancestral polymorphisms across multiple successive speciation events and is followed by stochastic allele fixation in each descendant lineage, potentially making phylogenetic inference at the level of individual loci problematic.
Past studies on the extent of ILS during speciation have been restricted to recent divergences because homoplasy needs to be low. For example, divergences among great apes show that ~30% of the gorilla genome exhibits nucleotide substitution patterns incongruent with the human/chimpanzee/gorilla species tree [7]. In contrast, the characteristics of ILS remain to be explored in adaptive radiations. As virtually homoplasy-free phylogenetic rare genomic changes [8,9], retrotransposed elements (REs) exhibit conflicting phylogenetic signals only when their insertions occurred on short internodes; they can thus be used to localize and quantify ILS even on very deep timescales [8,10–13].
We analyzed ~130,000 long terminal repeat (LTR) retrotransposons in the 48 recently sequenced bird genomes [4] and obtained 2,118 presence/absence patterns of insertions that occurred within the neoavian radiation and are distributed genome-wide (S1 Table, S1 Fig, S1 Data). These RE markers were obtained after visual inspection under strict criteria for coding of character states at orthologous RE loci (see Materials and Methods), because we aimed to minimize the two sources of potential homoplasy; independent RE insertion and precise excision.
Homoplasy via independent RE insertion requires the retrotransposition of the same RE subtype into precisely the same genomic location, in the same orientation, and featuring an identical target site duplication. In addition to these factors that make independent insertions very rare, the LTR retrotransposons studied here have a low copy number (e.g., 3,138 copies in the zebra finch genome), were active only for a short time period around the neoavian radiation [10], and show no target site preference among thousands of reconstructed ancestral target sequences of inserted elements (S2 Fig). We therefore propose that the probability of homoplasy caused by independent insertions among our RE markers is extremely low. Homoplasy via precise excision is the deletion of the RE insertion and one copy of the duplicated target site, but not a single bp more or less than that. These requirements make the occurrence of precise excisions very rare and we therefore visually inspected all of our markers for precise boundaries of presence/absence states and coded imprecise or poorly aligned boundaries as missing data. Altogether, we suggest that our 2,118 RE markers contain negligible homoplasy, and conflicts are instead due to ILS-derived hemiplasy.
To verify that incongruences constitute ILS-derived hemiplasy, Hormozdiari et al. [14] proposed to test for topological consistence between each RE marker and a sequence tree derived from its flanking nucleotides. However, we note that failure of this test for some of their RE markers does not equal homoplasy of RE markers. Alternative and more plausible causes for inconsistencies are homoplasy or tree reconstruction uncertainties in the flanking sequence trees and the fact that recombination may cause different topologies between adjacent loci [15]. Unfortunately, single-locus sequence trees of Neoaves have an average topological distance of 63% for introns and 66% for ultraconserved elements (UCEs) from the main Jarvis et al. tree [4]. This means that the average nonexonic locus fails to congruently resolve most of the neoavian internodes. We note that it is therefore not possible to independently verify hemiplasy in neoavian RE markers by comparison to flanking sequence trees. Nevertheless, if homoplasy was prevalent in our RE markers, we would expect to see an equal distribution of RE incongruences across all of the sampled clades of Neoaves. While we find dozens of presence/absence markers with incongruences affecting the short branches within the neoavian radiation (S1 Table; e.g., the core landbirds and core waterbirds clades), there is not a single RE incongruence in our presence/absence matrix (S1 Table) affecting well-accepted internal relationships within postradiation taxa, such as passerines, parrots, eagles, penguins, the woodpecker/bee-eater clade, the hummingbird/swift clade, and the flamingo/grebe clade. Such an imbalance of RE incongruences strongly implies that homoplasy is indeed negligible among our 2,118 RE markers.
We analyzed the RE presence/absence matrix using Felsenstein’s polymorphism parsimony [16] and obtained a single most parsimonious RE (MPRE) tree, whose branches are supported by a total of 1,373 conflict-free insertion events across the neoavian radiation (Fig 1B). The topology is very similar to previous phylogenomic estimates using mostly noncoding nucleotide data [4,10,17–21], including relationships previously strongly supported in whole-genome sequence analyses [4] (Fig 1A), such as the sunbittern/tropicbird, bustard/turaco, and mesite/sandgrouse clades. From these three groups, only the sunbittern/tropic clade was previously recovered in some multilocus analyses [19–21].
The remaining 745 retrotransposon markers show different degrees of gene tree–species tree incongruence. This is best explained by the persistence of ancestral polymorphisms across successive speciation events, followed by reciprocal allele fixation in each of the descendant lineages, i.e., ILS. We define the extent of ILS as corresponding to weak conflict (persistence across two speciation events; Fig 1C), moderate conflict (three events; Fig 1D), or strong conflict (more than three events; Fig 1E). Per-branch counts of ILS-affected RE insertion events show that incongruences are pronounced on some internodes and are nearly absent on others (Fig 1A and 1B), with greater conflict in deeper internodes. The internodes among core waterbirds exhibit weaker discordances and large amounts of conflict-free RE markers, which is in line with the observation that the RE relationships are fully congruent with the genome-level sequence analyses in Jarvis et al. [4]. Within core landbirds, the MPRE tree is fully congruent with a genome-level tree based on sequences from UCEs [4], yet discordant with the main tree from Jarvis et al. [4] with regards to the position of mousebirds. The deepest divergences of core landbirds contain many ILS-affected markers with strong discordances when mapped on the main Jarvis et al. tree [4] (Fig 1A), but slightly less so when mapped on the MPRE topology (Fig 1B). Furthermore, the placement of owls in the MPRE tree (Fig 1B) is in agreement with our preliminary analysis of owl REs [4]. However, we emphasize that an alternative grouping of owls with eagles and New World vultures received nearly as strong RE support [4], which suggests that the position of the owls may in fact approximate a trifurcation. Among the remaining neoavian divergences, nine internodes are discordant between our MPRE tree and the genome-scale sequence tree (Fig 1A and 1B), all of which are characterized by scarcity of ILS-free RE markers and dominance of RE presence/absence patterns that show complex incongruences resulting from ILS across at least four consecutive speciation events (sensu Fig 1E).
It is striking that the conflicting placements of mousebirds are well-supported in the main Jarvis et al. tree [4] on the one side and our MPRE tree and the genome-scale UCE tree [4] on the other side, respectively. One explanation for this could be hybridization of two distinct and diverged ancestral species, e.g., the ancestor of mousebirds and the ancestor of the woodpecker/bee-eater/hornbill/trogon/cuckoo-roller clade, which would lead to well-supported alternative topologies with conflicts across multiple well-supported branches. This form of hybridization would then be distinguishable from ILS by an over-representation of RE markers supporting one alternative, species tree-incongruent topology and an under-representation of markers supporting the remaining alternative topologies. Such a situation was recently suggested for the very base of the rodent phylogeny [22]. We therefore analyzed all six possible positions of mousebirds within Afroaves (core landbirds without the passerine/parrot/falcon/seriema clade) for their respective support by RE markers. We also analyzed RE support for a grouping of mousebirds as sister to the remaining core landbirds, which was previously suggested in limited RE studies that used the zebra finch genome as only query species [10,23]. The strongest support (29 RE markers) was found for mousebirds as the sister taxon of the remaining Afroaves (Fig 2A), and the six alternatives were recovered by two to eleven markers each (Fig 2B–2G), with four markers supporting the Jarvis et al. topology of mousebirds being sister to Coraciimorphae s. str. [4] (Fig 2D). The fact that we found no excess of markers supporting the main Jarvis et al. topology [4] over the other alternatives suggests that the mousebird conflict was not caused by hybridization. Instead, the nearly symmetric distribution of support among the six non-MPRE topologies indicates that the presence/absence patterns of these RE markers result from stochastic sorting of alleles after persistence of ILS across the early diversification of core landbirds. We thus suggest that the whole-genome and intron-specific sequence trees [4] recover a locally anomalous topology [15] driven by the known problematic behavior of mousebirds in sequence analyses [24]. We emphasize that the genome-scale UCE tree [4] supports exactly the same mousebird affinities as the majority of REs herein (Figs 1B and 2A), raising the question as to how the UCE phylogenetic signal was overruled by intronic signal within the main genome-scale sequence analysis of Jarvis et al. [4].
Mapping RE markers on a dated time tree of the main Jarvis et al. analysis [4] enabled us to estimate the temporal dynamics of ILS across the very short internodes of the neoavian radiation (Fig 3A). Jarvis et al. infer the onset of Neoaves diversification at around the K-Pg boundary [4], which is in stark contrast to most mitochondrial and multilocus nuclear studies (but see refs. [17,26]) that estimate the deepest neoavian divergences at >82 million years ago (MYA) [3,27] or even >100 MYA (reviewed by ref. [28]). We anticipate that this debate will persist for the upcoming years. However, given that the Jarvis et al. [4] estimates are the first based on genome-scale data, we consider these to be the most reliable molecular dates currently available. We found a negative correlation between branch length and the percentage of ILS-affected RE markers per branch (Spearman's ρ = −0.6888, p = 7.1×10−5; Fig 3C), which corroborates our assumption that ILS is indeed the driving force for most (if not all) of the observed incongruences. This is due to the fact that ILS has a higher probability of occurring if the time between consecutive speciation events is short [29,30], and we would expect no such correlation if the conflicts we refer to as ILS-derived were instead caused by homoplasy. Strikingly, the per-branch estimates of ILS (Fig 3A) suggest that all those branches (or their 95% credible interval of divergence times in ref. [4]) overlapping with the K-Pg boundary exhibit 40%–100% ILS and are mostly incongruent with our MPRE tree (Figs 1B and 3C). Furthermore, the three deepest branches within the post-K-Pg diversification of core landbirds are affected by 59%–81% ILS, including the two branches involved in the aforementioned mousebird conflict. This means that with the exception of the core waterbird/sunbittern/tropicbird branch and the core landbird branch, all branches affected by ≥40% ILS were incongruent with our MPRE tree (Figs 1B and 3C).
We then tested if temporal variation in RE insertion rates (Fig 3B) may account for some of the irresolution. While there is considerable rate variation between branches (Fig 3B), there is no correlation between branch length and RE insertion rates (Fig 3D) or between RE insertion rates and degree of ILS (Fig 3E). Altogether, this suggests that the low amount of ILS-free markers on the problematic branches is not the result of very low RE insertion rates (<10 REs per MY). This is further supported by the notion that most of the post-K-Pg branches within core landbirds and core waterbirds have similarly short branch lengths and mostly low RE insertion rates, yet also low ILS (e.g., core waterbird branch with 12 RE insertions per MY and 20% ILS). We thus propose that the high prevalence of ILS ranging from 40%–100% across the deepest relationships among Neoaves is not an issue of taxon or marker sample size and rather reflects the biology of very rapid speciation.
Complex gene tree–species tree incongruences of retrotransposon markers might be more accurately represented in phylogenetic networks where data conflicts are evident as reticulate relationships among taxa and alternative topologies are visible even within well-supported lineages [12,29]. The resultant neighbor-net [31] (Fig 4A) illustrates the differential distribution of ILS-derived incongruences across the neoavian radiation, with well-supported, low-conflict branches leading to the core landbird and core waterbird clades, respectively. Together with the relatively large amount of ILS of polymorphisms originating deep within each of these two clades, and at the very base of Neoaves (Figs 1A–1B and 3A), this conclusively reveals that neoavian evolution went through three adaptive radiations [4]. Notably, the differences in the extent of ILS among these radiations imply that the tempo or demography of speciation may have varied considerably under the circumstances of accelerated diversification. More precisely, most of the 18% ILS-affected RE insertion events within the core waterbird radiation did not sort completely across two speciation events (Fig 4B), whereas 27% of the insertions in the core landbird radiation did not sort across mostly two to three speciation events (Fig 4C). These percentages of total ILS are comparable to the 34% genome-wide ILS found among human/chimpanzee/gorilla gene trees [7]. Finally, the deepest radiation of Neoaves exhibits discordances in 73% of the RE markers, mostly explained by persistence of ILS across five to seven speciation events (Fig 4D). This is consistent with a highly reticulate network structure (Fig 4A), which is restricted to those internodes that overlap with the K-Pg transition (Fig 2A). Notably, only these neoavian relationships remained unresolvable in whole-genome sequence analyses [4], and ILS-free RE insertions are scarce on these internodes (Figs 1A–1B and 3A).
Retrotransposon loci affected by ILS are distributed across the avian genome, irrespective of the duration of ILS per intronic or intergenic marker (Fig 5A). This situation likely applies to the ancestral Neoaves genome, as the avian karyotype is unusually stable, including conserved synteny of the Z sex chromosome and ubiquitous presence of numerous microchromosomes [32,33]. The Z chromosome (no recombination in female meiosis) represents a low recombination environment [34–37] and is affected by ILS to a similar extent (34%) as the genome-wide average of 35% (S2 Table), and a tree based on RE markers from the Z chromosome yields similar topological discordances with the MPRE tree (Fig 5B and 5C, S3 Fig). This is contrary to the observation of previous studies (on nonrapid diversifications), where low-recombination autosomal regions and sex chromosomes generally exhibit less ILS due to the lower effective population size (Ne) of regions with low recombination [7,30]. Finally, it is striking that the incongruences among our chromosomal trees of rare genomic changes almost perfectly overlap with conflicts among whole-genome sequence trees derived from concatenated or coalescence-based analyses of various data partitions in Jarvis et al. [4] (Fig 5D, S4 Fig) and again yield a highly reticulate structure at the base of Neoaves (cf. Fig 4A). Taken together, this reveals that many of these deepest neoavian divergences receive considerable support in some and strong refutation in other analyses, suggesting that the consecutive arrangement of their very short internodes may potentially represent Rosenberg’s “anomaly zone” [38], i.e., observing stronger support for a gene tree than for the actual species tree [15].
The probability for the occurrence of ILS depends on Ne in relation to the time between consecutive speciation events [29,30], with Ne correlating positively and time negatively with the expected extent of ILS, respectively. The observed complex genealogical fates of ancestral RE insertion polymorphisms during the initial super-radiation (Fig 4D) therefore suggest that the onset of neoavian diversification was characterized by a large number of near-simultaneous speciation events of an ancestral species with large Ne. Considering the sheer amount of differing allelic combinations that are possible to result from stochastic sorting of ancestral biallelic genetic variation after up to 17 speciation events (Fig 4D), we hypothesize that such complex signals might overrule the underlying species tree-concordant signal, because the latter can be expected to occur rarely under the complex sorting scenario envisioned (cf. Figs 1A–1B and 3A). Considering all theoretically possible RE presence/absence patterns in a five-taxon tree (Fig 1C–1E), ILS across four speciation events requires allelic sorting in each of the descendant lineages, permitting 22 different character distributions that are discordant with the species tree (Fig 6A, S5 Table). Under the model of stochastic sorting of polymorphisms of RE presence/absence (Fig 6) or other types of biallelic variation (e.g., single nucleotides), the probability for the occurrence of hemiplasy surpasses 90% after an ILS duration of seven speciation events (Fig 6D, S5 Table). This may explain why the deepest neoavian bifurcations receive various alternative topologies in the different genome-scale sequence trees of Jarvis et al. [4] (Fig 5D). However, the high bootstrap support (>90%) for some alternative bifurcations could also mean that there are several comparably likely relationships, thus resembling a local network. Alternatively, Salichos & Rokas recently proposed that bootstrapping in phylogenomic analyses can lead to strong support for bifurcations even in the light of strong conflict [39]. Even if one of these genome-scale bifurcating trees reflects the actual neoavian species tree, the verification of such a phylogenetic hypothesis remains challenged by the underlying complex discordances. Finally, the nearly star-shaped topology of this super-radiation (Figs 4A and 5D) may reflect population complexity of the ancestral species, especially if the succession of population isolation during explosive speciation happened in disagreement with prior population structure [40,41].
We conclude that Neoaves diversification is more complex than can be shown in fully bifurcating trees and exhibits a dynamic picture of ILS. The timing of the highly ILS-affected initial super-radiation coincides with the K-Pg extinction of nonavian dinosaurs and archaic birds [42], suggesting that the abrupt availability of ecological niches [4] was followed by near-simultaneous population isolations [41] via specializations and led to several network-like relationships. The subsequent, decelerated adaptive radiations of waterbirds and raptorial [4] landbirds exhibit less ILS and likely took place after the K-Pg boundary [4,17]. Interestingly, this time span is similar to placental mammal diversification [43], which was accompanied by localized and less pronounced ILS than shown here for Neoaves [12,13]. Finally, and contrary to the expectation that complete genomes will permit full resolution of phylogenies [44], our genome-level analyses of rare genomic changes yield a broadly bifurcating species tree of Neoaves [4] with local network-like reticulations that probably lie in the anomaly zone. Our study thus provides empirical evidence for a locally confined “hard” polytomy [41], and we predict that future genome-wide studies of ILS in other adaptive radiations will reveal further examples where a fully bifurcating, universal species tree is an oversimplification of the underlying complexity of speciation.
Our taxon sampling comprises the genome assemblies of 48 recently sequenced birds [45] and thus contains the same species that were used in the genome-scale sequence analyses of Jarvis et al. [4]. We focused on identifying RE insertion events during early neoavian evolution and therefore excluded non-neoavian genomes (chicken, duck, ostrich, tinamou, turkey) from the set of query species used for extracting RE candidate loci. All neoavian genomes were utilized as queries, with the exception of close relatives of ingroup species (as they do not add much more information) such as zebra finch (i.e., four remaining passerines), white-tailed eagle (i.e., bald eagle), budgerigar (i.e., kea), and adelie penguin (i.e., emperor penguin). This taxon sampling contains representatives of all major neoavian lineages [4], and we thus consider it sufficient for estimating ILS during Neoaves diversification. We expect that the addition of more taxa via sequencing of additional genomes would not result in an improved resolution of our RE data but rather lead to an increase of missing data and the detection of additional ILS across internodes that lie outside the three neoavian radiations reported herein.
We analyzed a total of ~130,000 copies of hitchcock-related LTR retrotransposons [46] that were previously shown to be REs characteristic of early bird evolution [10,47] and constitute the majority of RE activity during the neoavian radiation [10]. After repeat annotation of the sampled genomes using RepeatMasker [48] version 3.2.9, we extracted all TguLTR5d elements for each query species, including 1-kb flanks per RE locus. These sequences were then compared to the remaining query species via BLASTn [49] (cutoff E < 10−30), followed by extraction of the BLASTn hits and generation of locus-specific alignments in MAFFT [50] (version 6, E-INS-i). These alignments were postfiltered to exclude loci exhibiting less than ten species, missing flanks, plesiomorphic RE insertions (i.e., orthologous presence among all query species), or autopomorphic RE insertions (i.e., presence only in one query species). Furthermore, we omitted loci that were redundant or potentially paralogous. Among the ~8,000 remaining candidate loci, we manually identified phylogenetically informative RE insertions (including additional RE insertions in the sequences flanking the TguLTR5d query) in ~3,000 loci. For these marker candidates, we compiled final multispecies alignments after BLASTn searches (cutoff E < 10−10) of 2-kb flanks against the full taxon sampling of 48 birds.
RE markers serve as virtually homoplasy-free estimators of ILS-derived hemiplasy after minimizing potential errors that might arise from misalignment or incorrect scoring. Therefore, we carefully inspected the 48-species presence/absence alignment of each of the ~3,000 marker candidates by eye and manually coded binary character states using strict standard criteria [10,51,52]. Character state “1” requires the presence of an orthologous RE insertion (i.e., identical insertion point, orientation, RE subtype, and target site duplication) in an orthologous genomic locus (i.e., single-copy flank regions). Character state “0” constitutes the absence state of a particular RE insertion, as indicated by the presence of a nonduplicated target site and an alignment gap precisely corresponding to the RE presence/absence boundaries. If neither of the conditions necessary for character states “1” or “0” were met, character states were treated as missing data and coded as “?”. The same was done in the case of a gap in the genome assembly or a large unspecific deletion of the insertion locus. Marker candidates that did not meet the aforementioned strict criteria were omitted. This overall procedure led to a reduction of the ~3,000 marker candidates to a final set of 2,118 RE markers. Note that from these markers, 61 were previously published as a preliminary analysis of owl retrotransposons with the focus on determining the owl sister group [4,53]. We emphasize that these 2,118 markers encompass all RE insertion events during the neoavian radiation that were identified with our screening approach, with the exception of shallower internodes because they were not the main focus of our analyses of neoavian ILS. In the latter cases, we recorded a subset of the numerous marker candidates for these internodes, namely the woodpecker/bee-eater/hornbill, hummingbird/swift, pelican/egret/ibis/cormorant, and flamingo/grebe clades, as well as internodes within these.
We manually recorded target site duplications (i.e., direct repeats of 5 bp flanking the studied LTR retrotransposons) for each of our RE markers (S1 Table). This was done by visually inspecting the left and right flanks in our 48-species marker alignments to parsimoniously infer the putative ancestral states of lineage-specific nucleotide changes in each orthologous target site. We therefore suggest that these reconstructed motifs approximate the respective target sequences at the time points of RE insertion. We analyzed the motifs in 5′–3′ orientation (relative to the LTR orientation) using WebLogo [54]. The resultant sequence logo [55] contains near-equal frequencies of the four possible nucleotides per motif position (S2 Fig), which suggests that there is no target site preference among the REs studied herein.
We analyzed the 1/0-coded presence/absence matrix of 2,118 RE markers using the Dollop program in PHYLIP [56] version 3.695 under polymorphism parsimony and standard parameters with randomized input order of species (7 times to jumble, random seed “11111”). The Dollop output contained the resultant MPRE tree and the parsimony-inferred per-branch character states for each RE marker, which we used to calculate the amount of ILS-free markers per internode, and to infer the duration of ILS across speciation events in incongruent insertions. We also ran Dollop using the main Jarvis et al. tree [4,53] as input tree under the same aforementioned parameters, which was followed by estimation of the amount and duration of ILS across internodes. Subsequently, Z-chromosomal and microchromosomal RE trees were generated using Dollop under the same parameters as the MPRE tree. Finally, Splitstree [31] version 4.13.1 was used for neighbor-net analysis of conflict within our RE presence/absence matrix and supernetwork analyses of conflict between different tree topologies based on REs (S2 Data) or nucleotide sequences [4,53].
Our phylogenetic analyses yielded a reconstruction of transitions of character states for each RE marker, thus allowing the analysis of ILS-derived hemiplasy under the assumed negligibility of homoplasy. We defined an ILS-free marker (i.e., duration of ILS across maximally one speciation event) as one that required a single step when mapped on the analyzed tree. In the Dollop output, this is coded as a single transition to the presence state (“1”) on the branch where the RE insertion occurred. If a presence/absence pattern required more than one step when mapped on the given tree, it was defined as an ILS-affected marker (i.e., duration of ILS across minimally two speciation events). Under polymorphism parsimony, such a pattern results from a polymorphic RE insertion (“P”) that occurred on a branch prior to the conflicting branches and then persisted as a polymorphism across two or more speciation events, followed by stochastic allele sorting in the descendant lineages. We thus counted the total number of transitions to the presence (“1”) and the absence (“0”) allele necessary to explain the given topological conflict as a measure for the minimal amount of independent allele fixation events. Considering that n speciation events give rise to n + 1 lineages (Fig 6), our estimates of the duration of ILS correspond to the minimum of speciation events across which ILS persisted when counting the minimal amount of lineages that must have independently sorted under the given RE presence/absence pattern. Finally, we manually counted all possible allelic fates for ILS across two to four speciation events (Fig 6A) and derived a formula to calculate the amount of species tree-incongruent presence/absence patterns theoretically resulting from any duration of ILS (Fig 6, S5 Table). Dividing this number of hemiplasious character distributions by the amount of all theoretically possible presence/absence patterns yielded the probability of occurrence of hemiplasy in a biallelic polymorphism (Fig 6C, S5 Table).
|
10.1371/journal.pgen.1003882 | Correlated Occurrence and Bypass of Frame-Shifting Insertion-Deletions (InDels) to Give Functional Proteins | Short insertions and deletions (InDels) comprise an important part of the natural mutational repertoire. InDels are, however, highly deleterious, primarily because two-thirds result in frame-shifts. Bypass through slippage over homonucleotide repeats by transcriptional and/or translational infidelity is known to occur sporadically. However, the overall frequency of bypass and its relation to sequence composition remain unclear. Intriguingly, the occurrence of InDels and the bypass of frame-shifts are mechanistically related - occurring through slippage over repeats by DNA or RNA polymerases, or by the ribosome, respectively. Here, we show that the frequency of frame-shifting InDels, and the frequency by which they are bypassed to give full-length, functional proteins, are indeed highly correlated. Using a laboratory genetic drift, we have exhaustively mapped all InDels that occurred within a single gene. We thus compared the naive InDel repertoire that results from DNA polymerase slippage to the frame-shifting InDels tolerated following selection to maintain protein function. We found that InDels repeatedly occurred, and were bypassed, within homonucleotide repeats of 3–8 bases. The longer the repeat, the higher was the frequency of InDels formation, and the more frequent was their bypass. Besides an expected 8A repeat, other types of repeats, including short ones, and G and C repeats, were bypassed. Although obtained in vitro, our results indicate a direct link between the genetic occurrence of InDels and their phenotypic rescue, thus suggesting a potential role for frame-shifting InDels as bridging evolutionary intermediates.
| Homonucleotide repeats are exceptionally prone to insertions and/or deletions of bases (InDels). However, unless they occur in a multiplicity of 3 bases, InDels disrupt the reading frame and are thus expected to be purged from coding regions. Homonucleotide repeats, however, are also vulnerable to slippage by RNA polymerases and the ribosome. Using laboratory evolution techniques, we systematically mapped the occurrence of InDels within a given gene, before and after selection. Our data indicate that frame-shifting InDels were frequently bypassed to give functional proteins at surprisingly high frequencies. Further, we found a strict correlation between the repeat length, the frequency of occurrence of InDels at the DNA level, and the likelihood of bypass by transcriptional/translational slippage. Our results suggest that frame-shifting InDels might comprise functional evolutionary intermediates, and an effective mean of sequence divergence (e.g. when an adjacent InDel restores the frame, resulting in altered sequence and, potentially, in an altered protein structure).
| InDels occur in all kingdoms of life, and in some organisms they are as frequent as point mutations [1]. Short sequence repeats, and homonucleotide repeats in particular, are prone to InDels due to misalignment of the DNA strands during replication (polymerase slippage) [2]. In coding regions, at least 2/3 of InDels disrupt the reading frame and are thus considered nonsense mutations leading to loss of function (for comparison, only ∼1/20 of point mutations result in a stop codon). Frame-shifting InDels are therefore thought to be tolerated only when a gene is freed from selection pressure. Indeed, under fluctuating environments, and/or within small populations, frame-shifting InDels in sequence repeats provide a rapid means of switching genes on and off [3]–[7].
Frame-shifting InDels are considered by default as nonsense mutations. There are, however, known precedents for their bypass to give functional proteins due to transcriptional and/or translational infidelity [8]–[15]. Frame-shifts are also recruited as a regulatory mechanism by ribosome programmed −/+1 frame-shifting [9]–[13]. In other cases, alternative proteins are encoded from the same gene via translational frame-shift [16], [17]. Overall, although rare, bypass of frame-shifts has been identified in organisms from all three domains of life and in viruses [9]–[11]. Most recorded events of bypass occur in long homo-adenine repeats, e.g. ≥10A [8], [14], but bypasses within non-repeat stretches have sporadically been reported [9], [10]. Most cases also regard explicitly evolved mechanisms for the transcriptional and/or translational bypass of frame-shifting InDels, rather than accidental slippage. It therefore remains unknown to what degree randomly occurring frame-shifting InDels can persist in coding regions under purifying selection, and whether and how the likelihood of bypass relates to the sequence contexts within which a frame-shift occurs.
Our own interest in frame-shifts followed the directed, laboratory evolution of a DNA methyltransferase M.HaeIII, towards new target DNA specificities [18]. During this process, surviving variants were identified that carried frame-shifting InDels within in their coding regions. Common to all of these variants was the location of the frame shift mutation (an ‘A’ insertion within an 8-homonucleotides repeat, Figure S1). Being unaware of the possibility of bypass, we assumed these are ‘false positives’ even though these variants did exhibit detectable level of methylation of the newly evolved target. Whilst these InDel-carrying variants disappeared in the subsequent rounds of selection, we encountered additional examples of functional variants carrying frame-shifts in the selection of other proteins. We thus became curious as to how frequent the bypass of frame-shifting InDels might be, and whether they may serve as viable evolutionary intermediates.
InDels are of particular interest as they readily create alterations in a protein's length and sequence, and thus go beyond the exchange of single side chain [19]. This is also the case with our model, M.HaeIII, is a DNA methyltransferase isolated form Haemophilus aegyptius that specifically methylates GGCC ds-DNA sites. M.HaeIII belongs to the prokaryotic restriction-methylation system that encompasses hundreds of different methyltransferases, each with a different DNA target specificity. The target recognition domains (TRDs) of DNA methyltransferases exhibit relatively of low structural order and are highly diverse, including extensive changes in length [20]. We suspected that the intense diversification of TRDs might relate to InDels. We used the “Path” algorithm that produces DNA sequence alignments by back-translation of known proteins sequences (Figure S2). In this manner, frame-shifting InDels that might have underlined the divergence of these sequences might be detected [21], [22]. As discussed in detail below, frame-shifting InDels were readily identified in the aligned TRDs. However, since protein evolution is assumed to occur via a series of functional intermediates [23], the evolutionary relevance of frame-shifting InDels depends on their potential to be rescued via bypass of transcriptional or translational errors.
Here, we have systematically mapped the accumulation of InDels in a laboratory-performed genetic drift of a single gene/protein. We were interested in measuring M.HaeIII's tolerance of InDels, given that, a priori, 2/3 are expected to be purged due to frame-shifts, and that the in-frame ones are also far more deleterious than point mutations [20]. To this end, we subjected M.HaeIII to iterative rounds of random mutagenesis in-vitro followed by purifying selection. We analyzed the gene repertoires by high-throughput sequencing; both the repertoire before selection, thus mapping the occurrence of all mutations regardless of their effect on M.HaeIII, and the repertoire after selection, thus mapping the repertoire of accepted mutations. We thereby measured, for all 987 bases along the M.HaeIII gene, the occurrence rates of InDels due to DNA polymerase slippage, and the rates of their bypass due to transcriptional/translational errors. The data indicate that the rate of bypass of frame-shifting InDels is unexpectedly high, including in relatively short homonucleotide repeats, and in repeats of nucleotides other than adenine. Foremost, we found that the propensity for the genetic occurrence of InDels, and the rates of transcriptional-translational bypass, are highly correlated.
M.HaeIII was subjected to a laboratory genetic drift, namely to repeated rounds of random mutagenesis and purifying selection that eliminated non-functional variants (negative selection, Figure S3). To this end, M.HaeIII's gene was subjected to random mutagenesis using an error-prone DNA polymerase, at an average of 2.2±1.6 mutations per gene. The ensemble of mutated genes was ligated into an expression vector using restriction sites at the very beginning of M.HaeIII's ORF, around the ATG codon, and just after the stop codon. A plasmid vector was necessary for obtaining a large number of transformants such that large repertoires (≥105 variants) could be explored. However, when driven from high copy plasmids protein, expression levels can be unrealistically high, and thus bias the level of the bypass. To minimize the levels of expression, the mutated M.HaeIII genes were cloned under the control of the tightly regulated tet promoter, with a constitutively expressed tet repressor encoded downstream. The selections throughout the drift were performed at basal expression, i.e., with no inducer added to the growth media. This basal expression level was nonetheless sufficient for complete methylation of the encoding plasmid, as well as of the genome of the E. coli host, by wild-type M.HaeIII [18].
The ligated plasmids were transformed to E. coli, such that each transformed cell incorporated a different plasmid molecule carrying a different gene variant from the library of M.HaeIII mutants. In each bacterium, the transformed plasmid is replicated, and subsequently methylated at GGCC sites, or not, depending on the functionality of the M.HaeIII variant it encoded. The transformed bacteria were grown, and the plasmid pool was subsequently extracted and treated with the cognate restriction enzyme, HaeIII. Plasmids that encoded a functional M.HaeIII variant survived the digestion and thereby could be retransformed to fresh E. coli cells and propagated for the next round of selection [18].
We maintained ≥105 transformants per round of mutagenesis-selection thus avoiding population bottlenecks and the fixation by chance of deleterious mutations. Overall, M.HaeIII's gene underwent 17 rounds of random mutagenesis and purifying selection through which the drifting population accumulated an average of 2.0±1 point mutations per gene per round.
The M.HaeIII genes encoded by the plasmid pools were subjected to high-throughput sequencing (Illumina). Sequencing was performed following the first round of random mutagenesis, thus mapping the occurrence of mutations irrespective of selection (G0, or the naive repertoire). Additionally, the pool derived after 17 rounds of mutagenesis and purifying selection as sequenced to map the repertoire of accepted mutations (G17).
The short sequencing reads (∼40 nts) were mapped to the sequence of wild-type M.HaeIII. The analyzed sequence stretch included the coding region of M.HaeIII's that was repetitively mutated and re-cloned into the selection plasmid (987 nts), as well as a plasmid region located upstream of the cloning sites that was not subjected to mutagenesis (98 nts). The latter was used to determine the background frequency of sequencing errors due to the Illumina processing, in both repertoires, G0 and G17. This background frequency was subtracted from the InDel or point mutations frequencies observed at the positions subjected to drift. This procedure allowed us to measure the mutational frequencies for all possible point mutations and InDels throughout M.HaeIII's coding region, from nucleotide position 4 (downstream the NcoI cloning site) to the stop codon (position # 993; 17 nucleotides upstream the NotI cloning site).
The frequency of a given mutation, namely a given nucleotide exchange or a given InDel, at a specific position, corresponded to the number of contigs that carried this mutation divided by the total number of sequenced contigs that covered this position. We excluded sequenced positions within contigs obtained with low accuracy score, and/or showing biases with respect to their location (e.g. positions located at the edges of the contigs, see Methods). In this manner, all InDels that occurred at a frequency above the background were identified, in both the naive and the selected repertoires (Table 1).
The in-vitro mutagenesis protocol applied here does not reproduce the factors that contribute to InDels formation in natural genomes. Nevertheless, certain patterns observed in the acquisition of mutations in natural genomes were also observed here. For instance, the point mutations to InDels ratio (S/I) in our unselected repertoire (G0) was found to be ∼16. This ratio is within the range observed in natural genomes (e.g., ∼10 in humans, or 16 in S. cerevisiae) [1]. The InDels that relate to polymerase slippage in natural genomes are typically short (≤5 bp in length) with short frame-shifting InDels (i.e., InDels of 1 or 2 nts) comprising over two-thirds [24]. Here, single nucleotide InDels were overwhelmingly represented (∼98% of all detected InDels; Table 1), and were thus all expected to result in frame-shift and loss of function. Finally, as detailed below, the tendency of InDels to occur within repeat regions of natural genomes is also observed in the in-vitro generated naive library.
The predominant mechanism generating InDels in natural genomes is polymerase slippage due to infidelity in repeat sequence pairing [2], [25], [26]. Indeed, essentially all InDels observed here occurred within homonucleotide repeats of 3–8 nucleotides (Table 2). Further, as reported for natural genomes [27]–[29], the frequency of occurrence positively correlated with repeat length (R2 = 0.97, Table 2).
Sequencing of the selected repertoire, G17, showed that the overall tolerance of InDels was, as expected, low. Accordingly, under selection, the point mutations to InDels ratio (S/I) within M.HaeIII's coding region increased from ∼16 in G0 to ∼190 in G17 (Table 1). Frame-shifting InDels in M.HaeIII were primarily found to comprise nonsense mutations. Indeed, the purging of InDels in coding regions of natural genomes is intense [20], [30], [31].
The purging of mutations, including InDels, in disordered regions including inter-domain linkers and domain termini is far less intense than within ordered domains [20], [32]. Accordingly, the purging of InDels was >30-fold less intense at the last 7 amino acids of M.HaeIII's C-terminus. This stretch, starting from amino acid position 324, or nucleotide position 969, until the stop codon is structurally disordered and has no functional role (Table 1, lower panel, marked as ‘C-terminus, amino acids 324–330’). Thus, premature stop codons, or completely altered amino acid sequences within this region have little effect on M.HaeIII stability and function.
As expected most InDels were purged, yet an expectedly high fraction was found to be tolerated. Overall, out of 337 positions in which InDels were detected in the naive repertoire, 79 positions were found to carry InDels in G17. Out of the latter, 26 positions carried InDels at a frequency ≥0.2×10−3 (≥10-fold higher than background frequency; Figure 1, Figure S4). Thus, InDels that were found at significant frequencies in G17 were consequently considered as potentially tolerated.
As observed for the occurrence of InDels in the naive repertoire (G0), the frequencies of tolerated InDels (G17 frequencies) were highly correlated with length of the repeat in which they occurred (Table 2, Figure 2). Thus, the InDels that are most prone to occur are also the ones that are most likely to be rescued by transcriptional/translational errors. Indeed, out of 337 different positions in which InDels were identified in total (Figure S4), only 9 were observed above background rates and not within homonucleotide repeats (Table S1). These were found either at the end of homonucleotide repeats whereby the inserted or deleted nucleotide differed from the repeat one, or in short repeats such as TCTCT. These non-canonical InDels might be bypassed not at the InDel position itself, but at the adjacent repeat.
To verify that the frame-shifting InDels observed under selection were indeed bypassed, we generated 15 different M.HaeIII mutants each carrying a specific InDel that had been observed in G17 at frequency above 0.2×10−3. We also tested four InDels identified with high frequencies in the unselected, G0 library yet with near-background frequencies in G17 (# 9, 12, 13 and 14; Figure 1, Table 3). The InDel-carrying variants were individually cloned into the same plasmid that was used for the drift and transformed into E. coli. Cultures derived from cells transformed with individual M.HaeIII InDel variants were grown under basal expression, or under over-expression conditions (with inducer). The functionality of individual variants was determined by the standard plasmid protection test [18], [33], i.e., by the ability of the InDel-carrying variants to protect their encoding plasmids from HaeIII digestion (Figure 3). Upon over-expression, out of the 15 frame-shifted variants that corresponded to InDels found in G17, 13 showed detectable level of protection, and hence measurable methyltransferase activity (Figure 3A). Out of the four tested InDels that were found to occur in G0 but were purged under selection, two with the lowest G17 InDel frequencies (#9 and 14) showed no activity as expected. The other two (#12 and 13) exhibited detectable level of protection only when over-expressed. At basal expression level (Figure 3A), the InDels widely differ in their effects. Nonetheless, 9 out of 15 InDels were found to be bypassed at basal levels, and these also exhibited the highest G17 frequencies (Table 3). For example, variants #5–7 corresponding to InDels within the longest homorepeat (8A repeat carrying an ‘A’ deletion, or ‘A’/‘AA’ insertions, Table 3) showed high protection levels at basal expression and also the highest frequency of occurrence in G17. This result is not that surprising as long homo-A repeats are known hotspots for transcriptional/translational slippage (usually ≥10 nucleotides) [8], [14]. However, several variants carrying InDels within shorter repeats were also found to be bypassed at basal expression levels (#1, 15, 17, and 19, that occurred within 5A, 5C, 4A, and 5T homorepeats, respectively). In fact, one of these, with a deletion within a 4A repeat, seems to exhibit the highest protection level at basal expression (#17). Further, the most active InDel-carrying variants (#8 and #17) exhibited physiologically relevant levels of methylation activity as was also indicated by their ability to fully methylate their GGCC sites in the genomes of their host E. coli cells, even at basal expression levels (>12,000 GGCC sites protected from HaeIII digestion versus 19 in the selection plasmid, Figure S5).
Further validation that these frame-shifting InDels are bypassed to yield full length proteins was provided by a Western blot using an M.HaeIII construct that carries an epitope tag at the C-terminus (Figure S6). The observed levels of full-length proteins were well correlated with the plasmid protection levels at basal expression level, and with the G17 frequency of the InDels that these variants carry.
Although showing relatively high G17 frequencies, two variants (#3, 4, insertion and deletion of ‘G’ at 6G repeat in position 204) showed no detectible methyltransferase activity, even when over-expressed. This and the fact that some InDels are only bypassed upon over-expression, does not necessarily mean that their detection of these InDels in G17 is an artifact. Due to the short contigs of Illumina sequencing, the sequence composition of the full length drifted M.HaeIII variants within which these InDels originally occurred is unknown. They may well contain compensatory mutations at the background on which these InDels were tolerated. Indeed, laboratory drifted variants accumulate global suppressor mutations at high rates [34], as was also observed in our laboratory drift of M.HaeIII (unpublished data). In fact, the acceptance of InDels in naturally drifting sequences also seems to be correlated with the acquisition of enabling point mutations [20].
The relatively high frequency of tolerated InDels revealed here reinforces the possibility that frame-shifting InDels should not be considered by default as dead-ends. Rather, the sequence context within which InDels occur most frequently may also promote their tolerance. Clearly, this and other conclusions derived from this study need to be considered in view of the in-vitro mutagenesis protocol, the laboratory selection context, and the data that relates to one gene/protein. Nonetheless, key features that are also relevant to the acquisition of InDels in natural genomes were captured – foremost, the tendency of InDels to occur within repeat regions, and the higher frequency of frame-shifting InDels relative to in-frame ones. Our experimental system mimics these two features and thereby enabled us to systematically measure the rate of occurrence of InDels within all positions of the studied gene, and in the absence and in the presence of selection.
The levels of bypass observed in our dataset might be artificially elevated as a result of enhanced gene copy number and/or expression levels. In our experimental setup, M.HaeIII was encoded by a multi-copy plasmid and under an inducible promoter. Nonetheless, a relatively low, basal expression level (i.e., without induction) was maintained due to the tight regulation of the tet promoter with constitutive over-expression of repressor from the same plasmid. The natural restriction-modification system from which M.HaeIII was derived is encoded by a chromosomal gene. However, whereas the restriction enzyme is tightly regulated, the methyltransferase is constitutively expressed [35], [36]. Whilst we have no direct comparison of the protein doses in nature and in our experiment, they are unlikely to differ dramatically (for comparison, a similar plasmid-based experimental setup showed no detectible GFP signal when inducer levels were ≤20 µg/ml [37], whereas in our drift no inducer was added).
The generation of InDels in this laboratory drift was the outcome of polymerase slippage during DNA replication (components such as DNA repair were not included in the in-vitro replication protocol). This assumption is supported by the strict correlation between repeat lengths and frequency of InDels within their positions (Figure 2; G0 line). In non-repeat positions (positions whereby the flanking bases differ from the base in the mutated position), the occurrence frequency of InDels is close to the detection limit. Indeed, extrapolating from the observed linear correlation of repeat length and log[InDel frequency] to repeat length = 1, a frequency of ∼10−4 is obtained. The InDels frequency increases by ∼2.5-fold per nucleotide as the repeat length increases (Figure 2; G0).
Frame-shift bypasses are primarily associated for homo-A repeats [8], [14]. Homo-A and homo-T repeats show higher InDel frequencies than G/C repeats, but there are no clear trends regarding composition, primarily because the different base repeats are represented in M.HaeIII's gene at very different frequencies (e.g. 34 A/T 3-nucleotide repeats versus 5 G/C repeats; Table 2). A strict correlation was also observed between the frequency of bypassed frame-shifting InDels and repeat length. Due to the purging of most InDels by the negative/purifying selection, the slope of the correlation curve is steeper: 3.4-fold higher frequency per nucleotide length. Additionally, the intercept with the Y-axis indicates that bypass at non-repeat positions is below the background level (∼0.3×10−5, Figure 2; G17).
The validation of functional variants carrying individual InDels was consisted of: (i) their well-above background frequencies in the repertoire of M.HaeIII genes that passed the selection, G17, (ii) their persistence of the GGCC methylation functionality, and (iii) the detection of full-length proteins. Overall, these data suggest that the bypass of InDels occurs by slippage of the RNA polymerase and/or the ribosome, thus shifting the reading-frame either upstream or downstream (−1 or +1 shifts, respectively) to the original frame. This mechanism accounts for the observed circularity in the generation and bypass of the frame-shifting InDels. Namely, the propensity of a given sequence stretch, in this case homonucleotide repeats, towards slippage of the DNA polymerase (the genetic InDel formation), of RNA polymerase (transcriptional bypass), or of the ribosome (translational bypass), is similar.
It should, however, be noted that because frame-shifting InDels are only partly bypassed, they impose a cost. Even when bypass produces enough full-length, functional protein (e.g. in those cases where 100% methylation of both the plasmid and the host's chromosome are observed; Figs. 2B, S5, variants #8, 17), truncated versions derived from the original frame are also produced (Figure S6), thus producing aggregated, deleterious debris. Any advantage afforded by a frame-shifting InDel therefore depends on the benefit afforded counteracting the cost associated with such debris [38]. Nevertheless, competitions of cells carrying wild-type M.HaeIII with cells carrying different InDel variants indicated no growth inhibition by frame-shifts (Figure S7). In fact, the InDel-carrying variants unexpectedly became enriched. The toxicity associated with DNA methylation could bias growth in favor of InDel-carrying variants that are less active. However, an E. coli strain was used in which methylation is not toxic [39] and the most active InDel variant (#17) showed the highest growth advantage (Figure S7B). It therefore seems that, at least for the genes tested here, and within our experimental setup, the growth disadvantage imposed by frame-shifts is undetectable, possibly because expression of wild-type M.HaeIII also produces truncated fragments at comparable levels (Figure S6).
The survival of frame-shifted variants was found to be dependent not only on the nucleotide repeat length, but also on the location of the InDel within the encoded protein. In accordance with previous findings [20], non-functional InDels (e.g. #3, #4, #9 and #14) were located within highly conserved regions, whereas the tolerated ones tend to be located between conserved motifs (Figure 1D). For example, InDels #3 and 4 comprise an insertion or deletion of a G nucleotide within a 6G repeat located in the middle of M.HaeIII's catalytic region (motif IV). These InDels were detected at high frequency in the naive repertoire, but with very low frequency in the selected, G17 repertoire (Table 3). In accordance, the variants carrying these InDels show no methylation activity, even under over-expression (Figure 3). In contrast and with agreement with the “Path” analysis for prediction for frame-shifting sites (Figure S2), InDels tolerated at high frequency tend to reside in connecting loops between conserved motifs. InDels #6–8, for example, reside in a flexible loop that connects motif V and VI (Figure S8). This tendency suggests that the bypass of most frame-shifting InDels results, as a minimum, in one point mutation. The conserved structural motifs are intolerant to substitutions and hence to InDels – to in-frame InDels [20], [40], let alone frame-shifting ones (Figure 1D, Table 3).
Altogether, our data indicate that frame-shifting InDels can be tolerated at a surprisingly high frequency. We identified at least eight readily bypassed InDels within M.HaeIII, all comprising homonucleotide repeats of 4–8 nucleotides located between the enzyme's conserved motifs. Variants carrying frame-shifting InDels within the longest 8A repeat were as expected highly tolerated, but other permissive InDels were identified in unexpected repeats, e.g. relatively short repeats (4 nucleotides), homo-G or -C repeats, and even next to repeats rather than within them (Table S1). The identification of multiple permissive InDels in M.HaeIII reinforces the possibility that shifted open reading frames are often read-through to give functional proteins (for examples see [8], [13], [14], [41]). The systematic mapping performed here indicated a strict correlation of the bypass likelihood with the repeat's length, and with its structural location. These parameters may assist the identification of ORFs carrying frame-shifts that may actually encode full-length, functional proteins.
The tolerance of a frame-shifting InDels correlates with the tendency of the position within which it occurred to acquire InDels in the first place. For the very same reason, the likelihood for reversion of an InDel, thus restoring the original frame, is very high. Reversion may occur at the same position, or at other positions within the same repeat (scenarios that are indistinguishable by sequence comparison), or, as observed here, in positions flanking the repeat (Table S1). Repeats therefore comprise hot spots for changes in length and composition, as observed in rapidly evolving proteins related to bacterial pathogenicity, or in organisms that rapidly switch on and off certain genes [3]–[7]. Similarly, the target recognition domains (TRDs) of DNA methyltransferases are highly diverse not only in sequence, but also in length [20]. Indeed, analysis of alignments of M.HaeIII and its orthologs with “Path” support the hypothesis of diversification via frame-shifting InDels (Figure S2).
The bypass of frame-shifting InDels, although transient and/or accompanied by partial loss of function, greatly increases the likelihood of occurrence of a second InDel in sequential proximity to the original one, thus restoring the frame. This may result in the diversification of both the length and composition of the entire stretch of amino acids between the two InDels, and thus, in drastic structural and functional changes occurring via functional intermediates [23]. Indeed, transcriptional and translational errors, or phenotypic mutations, may play an evolutionary role in shaping protein properties or acting as bridging intermediates [42]–[45].
The M.HaeIII wild-type gene carrying four stabilizing mutations [18] was cloned with an N-terminal His-tag into pASK-IBA3+ vector (IBA, Ampicillin resistance, using NcoI and NotI; Figure S9). Plasmids were transformed into E. coli strain ER2267 (EcoK r- m- McrA- McrBC-Mrr-) in which GGCC DNA methylation is not toxic [39]. Transformants were selected by growth on ampicillin.
Random mutagenesis was performed by PCR using an error-prone polymerase (GeneMorph Mutazyme, Stratagene) and primers than flank the M.HaeIII's ORF (pASK-F and pASK-R, Table S2). The wild-type gene with 4 stabilizing mutations [18] was used as template for the first round (G0). In following rounds, the selected pool of M.HaeIII variants from the previous round was used as a template for the next one. The PCR was optimized to an average of 2.2 mutations per gene. Each round of evolution, or generation (noted as ‘G’), included the following steps (Figure S3): (i) The pool of M.HaeIII genes from the previous round was randomly mutated, recloned using the NcoI and NotI sites, transformed to E. coli and plated on agar plates containing ampicillin. (ii) About 106 individual transformants were obtained in each round, and the cells were grown at 37°C over-night. (iii) The plasmid DNA was extracted and was digested with HaeIII (10–20 units, in 50 µl of NEB buffer 2, for 2 hours at 37°C). (iv) The plasmid DNA was purified (PCR purification kit, QIAGEN) and re-transformed for another round of enrichment. Each round of evolution included one cycle of mutagenesis and three cycles of enrichment (transformation, growth, plasmid extraction and digestion). The naive library, G0, relates to the transformed plasmid DNA derived from cloning of the repertoire of M.HaeIII genes after the first round of mutagenesis with no selection by HaeIII digestion.
The samples of the naive (G0) and the selected libraries from Round 17 (G17) were prepared in the following way: (i) The plasmid pools were PCR amplified with primers pASKXhoI-F and pASKXhoI-R that amplified the M.HaeIII's open reading frame while appending XhoI restriction sites at both ends (Figure S9, Table S2). (ii) The amplified products were purified by PCR purification kit (QIAGEN) and digested with XhoI (20 units, in 60 µl of NEB buffer 4, for 2 hours at 37°C). (iii) The digested products were isolated by gel electrophoresis and a gel extraction kit (QIAGEN). (iv) To avoid bias due to poor sequencing of the edges, the fragments were ligated using the XhoI site to give concatemers. The ligation products were purified by ethanol precipitation. Sequencing libraries were prepared and sequenced according to manufacturer's protocol at the Weizmann Institute's high throughput-sequencing core facility. The obtained sequencing reads (∼40 nts) were mapped to the reference sequence of wild-type M.HaeIII with two methods: (i) Using NCBI blastn v2.2.20 [46] with parameters: e-value cutoff 0.0001, word size 7, and while allowing up to 6 mismatches and requiring a minimal alignment length of 24 consecutive nts, as previously described [47], [48]; and (ii) Using Novoalign v2.07.00 with parameters: c 4 Hash step-size 6 [47]. Point mutations, insertions and deletions were assigned based on the mapping of the sequencing reads to the reference sequence as previously described [29], [48]. Large insertions (7 or more bases) were determined by the Blast alignments, due to Blast's ability to open long gaps by performing local-alignments of the sequences. Single nucleotide mutations and short indels (7 bases or shorter) were determined by the Novoalign alignments as they take into account the base-quality information provided by the Genome Analyzer platform (using quality threshold of Q20 for filtering both indels and point mutations). Every mismatch or gap in the reads alignment relative to the wild-type reference was recorded per each nucleotide position, and further analyzed using custom Perl scripts. Only InDels that were uniformly distributed along the 40 bp reads were included. Indeed, InDels that were detected with a high bias towards the edges of reads were individually tested and found to be artifacts and were manually removed (Figure S10). InDel frequencies were determined per nucleotide position as the number of reads with a given InDel(s) divided by the total number of reads that mapped this position.
Individual InDels were introduced by all-around PCR using the pASK encoding wild-type M.HaeIII as template and phosphorylated primers harboring each InDel (Table S3). The PCR products were gel purified, ligated (blunt-end ligation; 10 units of T4 ligase, NEB, 2 hours at room temp.) and transformed to E. coli. Transformants were selected on ampicillin, and InDel incorporation was confirmed by sequencing. Appending of the C-terminal HA-tag was performed by PCR using individual InDel constructs as template and primers pASK-F and XhoCtFus-R (Table S2). The PCR products were digested with NcoI and XhoI and ligated into a modified pASK vecor containing an in-frame C-terminal HA-tag (Figure S9).
Sequence-verified InDel variants were transformed to E. coli and grown in LB media in the presence of ampicillin to OD600∼0.6. The cultures were then split: 200 ng/ml of the expression inducer anhydrotetracycline (AHT) was added to one half and the second was kept growing as is. Cultures were grown over-night at 37°C. The plasmid DNA was extracted, treated with HaeIII restriction enzyme (10–20 units, 2 hours at 37°C) and analyzed by gel electrophoresis.
|
10.1371/journal.pgen.1004162 | Genomic Networks of Hybrid Sterility | Hybrid dysfunction, a common feature of reproductive barriers between species, is often caused by negative epistasis between loci (“Dobzhansky-Muller incompatibilities”). The nature and complexity of hybrid incompatibilities remain poorly understood because identifying interacting loci that affect complex phenotypes is difficult. With subspecies in the early stages of speciation, an array of genetic tools, and detailed knowledge of reproductive biology, house mice (Mus musculus) provide a model system for dissecting hybrid incompatibilities. Male hybrids between M. musculus subspecies often show reduced fertility. Previous studies identified loci and several X chromosome-autosome interactions that contribute to sterility. To characterize the genetic basis of hybrid sterility in detail, we used a systems genetics approach, integrating mapping of gene expression traits with sterility phenotypes and QTL. We measured genome-wide testis expression in 305 male F2s from a cross between wild-derived inbred strains of M. musculus musculus and M. m. domesticus. We identified several thousand cis- and trans-acting QTL contributing to expression variation (eQTL). Many trans eQTL cluster into eleven ‘hotspots,’ seven of which co-localize with QTL for sterility phenotypes identified in the cross. The number and clustering of trans eQTL—but not cis eQTL—were substantially lower when mapping was restricted to a ‘fertile’ subset of mice, providing evidence that trans eQTL hotspots are related to sterility. Functional annotation of transcripts with eQTL provides insights into the biological processes disrupted by sterility loci and guides prioritization of candidate genes. Using a conditional mapping approach, we identified eQTL dependent on interactions between loci, revealing a complex system of epistasis. Our results illuminate established patterns, including the role of the X chromosome in hybrid sterility. The integrated mapping approach we employed is applicable in a broad range of organisms and we advocate for widespread adoption of a network-centered approach in speciation genetics.
| New species are created when barriers to reproduction form between groups of organisms that formerly interbred freely. Reduced fertility or viability of hybrid offspring is a common form of reproductive isolation. Hybrid defects are caused by negative interactions between genes that have undergone evolutionary change within each subgroup. Identifying genetic interactions causing disease or trait variation is very difficult, consequently there are few known hybrid incompatibility genes and even fewer cases where both interacting genes are known. Here, we combined mapping of gene expression levels in testis with previous results mapping male sterility traits in hybrid house mice. This new approach to finding genetic causes of reproductive barriers enabled us to identify a large number of hybrid incompatibilities, involving genomic regions with known roles in hybrid sterility and previously unknown regions. Understanding the number and type of genetic interactions is important for developing accurate models used to reconstruct speciation events. The genetics of hybrid sterility in mice may also contribute to understanding basic processes involved in male reproduction and causes of human infertility.
| To understand patterns of biodiversity, it is essential to characterize the processes by which new species arise and are maintained in nature, including ecological specialization, population differentiation and reproductive isolation. Genetic dissection of reproductive isolation has proven to be an especially powerful strategy for revealing mechanisms of speciation. Many genomic regions and even specific genes that contribute to hybrid defects have been identified by genetic mapping in recombinant populations [1]–[7]. Divergence in gene regulation is expected to contribute to reproductive isolation between nascent species, and studies with F1 hybrids support this prediction [8]–[13]. Importantly, these two approaches – genetic mapping and measurement of genome-wide expression patterns in hybrids – have yet to be combined directly in the context of speciation.
Hybrid sterility and hybrid inviability frequently result from negative epistasis between mutations at interacting genes [14]–[16]. This “Dobzhansky-Muller model” predicts that disruptions in gene networks should be common in hybrids. By integrating organismal phenotypes and genotypes with gene expression patterns, this prediction can be tested. Despite the identification of hybrid incompatibility genes in several species and the prevalence of the Dobzhansky-Muller model, the nature and complexity of hybrid incompatibility networks remains poorly understood. Do hybrid incompatibilities generally involve two loci or are higher order interactions common? Are incompatibilities independent or do they share some common loci? Is the genetic architecture of hybrid defects similar among taxa? Known incompatibility genes have provided the first hints about these questions, particularly in Drosophila [6], yet too few genes and taxa are represented to determine whether there are generalities underlying the speciation process. A network perspective should provide insights into the genetics of reproductive isolation that are difficult to obtain using a gene-by-gene approach.
The house mouse (Mus musculus) is an excellent model for investigating speciation from a network perspective. Genomic resources are abundant for the house mouse, and reproductive processes are well characterized because the mouse is the premier model for fertility research in humans [17]. House mouse subspecies are in the early stages of speciation, showing significant but incomplete reproductive isolation. Evidence for hybrid male sterility in laboratory crosses [5], [18]–[22] and in natural zones of hybridization [23], [24] suggests it is a primary isolating barrier between these nascent species.
Studies of sterility in F1 hybrids between Mus musculus domesticus and Mus musculus musculus (subsequently referred to as domesticus and musculus) revealed an important role for the X chromosome and identified several contributing autosomal loci [4], [5], [25], [26]. One of these loci is Prdm9, a histone methyltransferase [27]. Hybrids with some alleles of Prdm9 from domesticus show pachytene arrest of meiosis. The effects of sterile Prdm9 alleles appear to be due to mutations in the protein-coding sequence and there is evidence for downstream regulatory effects, but the incompatibility network involving Prdm9 has not been revealed.
Genetic mapping of sterility phenotypes in F2 hybrids between M. m. domesticus and M. m. musculus recently identified an additional set of autosomal loci, which are predominantly recessive and thus contribute to sterility in second generation and subsequent hybrids. Genetic architectures of F2 sterility traits are complex, involving a moderate number of loci with a range of phenotypic effect sizes [1].
Genome-wide studies of gene expression in testis of F1 hybrids provide evidence that sterility is associated with disrupted expression [9], [10]. Like sterility phenotypes, expression patterns in hybrids depend on the origins of parental strains, and the direction of the cross. In many cases, testis expression in hybrids is intermediate between parental strains [9]–[11]. However, extensive misexpression (expression outside the range observed in parental strains) has been documented in a few crosses. Comparison of testis gene expression patterns between reciprocal F1 musculus-domesticus hybrids showed that many X-linked genes are overexpressed in sterile but not in fertile F1s [10]. To our knowledge, gene expression patterns in testes from F2 and later generation hybrids have not been described.
Here, we integrate analysis of genome-wide expression in testis from F2 musculus-domesticus hybrids with results from a previous study mapping sterility phenotypes in the same individuals [1]. We show that sterility is associated with large-scale alterations in gene expression in F1s and F2s, and we identify quantitative trait loci (QTL) that cause X chromosome-wide overexpression in hybrids. We report expression quantitative trait loci (eQTL) for a large number of transcripts. We compare the locations of eQTL with sterility QTL, and identify disrupted processes during spermatogenesis based on affected networks. Using a conditional mapping approach, we pinpoint genetic interactions affecting expression. We highlight candidate pathways, processes, and interactions for several loci, which provide insight into the mechanisms underlying their contributions to sterility.
We measured levels of misexpression in F1 and F2 hybrids to identify major alterations in gene expression pattern associated with sterility in M. m. domesticus (WSB/EiJ; hereafter domesticusWSB) - M. m. musculus (PWD/PhJ; hereafter musculusPWD) hybrids. Sterility is asymmetric in these crosses: F1 males with musculusPWD mothers (hereafter MxD F1s) are almost always completely sterile whereas F1s with domesticusWSB mothers (hereafter DxM F1s) are fertile [1]. MxD F1 males showed significant differences from both parents for all reproductive traits measured. By contrast, all traits in DxM F1s (except seminiferous tubule area) were within the range observed in the parental lines. Trait measurements in MxD F1s and DxM F1s provide ‘fertile’ and ‘sterile’ examples that are useful for assessing trait distributions in F2s.
Misexpression was markedly higher in testis of MxD F1s (18.8% transcripts; Fig. 1A) than in DxM F1s (1.6%). In both F1s, levels of misexpression were higher for X-linked transcripts than autosomal transcripts. On the X chromosome, the number of overexpressed transcripts in MxD F1s was much higher than the number of underexpressed transcripts (25.9% over, 4.4% under). The level of underexpression was higher on autosomes, but the difference between levels of over- and underexpression was smaller (7.1% over, 11.3% under). These results are consistent with previously reported differences in expression patterns between sterile and fertile F1s [10].
Misexpression in F2s varied from 0.9–39.0% transcripts (median 2.1%; Fig. 1A), encompassing the levels observed in fertile and sterile F1s. There was substantial overlap between transcripts misexpressed in MxD F1s and in >5% of F2s (Fig. 1B) yet a large proportion of transcripts were misexpressed only in F1s or F2s. The relatively continuous distribution of misexpression in F2s and lack of recapitulation of the full F1 misexpression pattern indicates multiple genetic factors contribute to misexpression. Misexpression unique to F2s suggests some contributing loci act recessively.
A large proportion of X-linked transcripts were negatively correlated with testis weight (lower testis weight = higher expression) – opposite of the pattern for autosomal transcripts, a majority of which was positively correlated with testis weight (Fig. 1C). This result suggests that – as in sterile F1s – the X may be broadly overexpressed in sterile F2s.
To determine whether the level of misexpression was consistent throughout spermatogenesis, we compared patterns of expression in F1 and F2 hybrids among genes identified as specific/enriched to different spermatogenic cell types in previous studies [28]. Autosomal transcripts expressed in meiotic and post-meiotic cells are underexpressed in sterile MxD F1s, and transcripts specific to somatic and mitotic cells are overexpressed (Table S1). This pattern is consistent with reduced spermatogenesis, as expected based on sterility phenotypes. The X chromosome is transcriptionally silenced during meiosis (meiotic sex chromosome inactivation MSCI; [29], [30]), and thus lacks transcripts associated with meiotic cells. X-linked transcripts associated with other testis cell types showed patterns consistent with autosomal transcripts; somatic and mitotic transcripts tended to be overexpressed and the few underexpressed transcripts were predominantly postmeiotic. Misexpression patterns across spermatogenic cell types in F2 hybrids were consistent with patterns in sterile F1s.
Next, we investigated the genetic basis of gene expression variation in individual transcripts. We identified 16,705–36,753 eQTL, depending on the significance criterion (Table 1). We used a permissive threshold, based on permutation of a single transcript, for downstream analyses because our goal was to identify genome-scale patterns. It is important to note that the false-positive rate among individual eQTL identified using this criterion is high, particularly for trans eQTL.
The genomic positions of the eQTL and the affected transcripts are shown in Figure 2. eQTL located near the quantitative trait transcript (QTT) comprise the prominent diagonal stripe, a pattern typical of eQTL studies [32]–[34]. These proximal eQTL are likely to be cis regulatory elements [33], [35]. We refer to proximal eQTL as cis eQTL for convenience, although it is possible that they might not act solely in cis (by regulating alleles only if they are on the same DNA strand). We classified eQTL with peaks within 5 cM of the transcript (probe) position as cis eQTL and eQTL located on a different chromosome from the transcript as trans. We ignored eQTL>5 cM on the same chromosome, because this class might include long-distance cis eQTL in addition to trans eQTL. We identified cis eQTL for 60% of transcripts (14,807; Table 1) and at least one trans eQTL for 56.7% (13,997) transcripts. The number of trans eQTL identified per transcript ranged from one (8,092; 32.8% transcripts) to seven (3; 0.01% transcripts).
We next examined eQTL dominance and effect size. Most cis eQTL (93.8%; Fig. S1A) were additive (mean for heterozygotes is intermediate and >2 standard errors from both homozygous means – see Methods). In contrast, a substantial proportion of trans eQTL were dominant (37.1%), underdominant (9.2%), or overdominant (8.6%). Curiously, musculusPWD alleles were more likely to be dominant among cis (473/859; 55.1%) and trans eQTL (2,850/4,580; 62.2%). We cannot think of an experimental or biological explanation for this bias. The two categories of eQTL differed in effect size (Fig. S1B). The difference in expression level between genotype classes was larger on average for cis eQTL than for trans eQTL (t = 72.3 (d.f. = 15931), P<2.2×10−16). The difference in effect size is also apparent when comparing the peak LOD scores of cis (mean = 25.05) and trans eQTL (mean = 5.94).
We tested for clustering of trans eQTL, which is commonly observed in eQTL analyses [36]–[38]. Some of these ‘trans hotspots’ are visible as vertical bands in the eQTL heatmap (Fig. 2). We identified 12 genomic regions significantly enriched for trans eQTL using a sliding window analysis (P<0.05, permutation test; Table 2). Two adjacent hotspots on chromosome 10 were combined for simplicity in downstream analyses. The most striking pattern was observed for the X chromosome: most of the X was significantly enriched for trans eQTL and 8,286 autosomal transcripts (34.6%) had eQTL mapped to the proximal X hotspot (0–42 cM). We discuss the massive effect of the X on gene expression in detail below, and relate this pattern to the known importance of the X in hybrid male sterility.
The genomic distribution of eQTL we identified, as well as differences in dominance and effect sizes between cis and trans eQTL, are broadly consistent with patterns previously described in eQTL studies performed in a variety of (non-hybrid) organisms (e.g. humans: [37], [39]; C. elegans: [36], [40], [41]; Arabidopsis: [32], mice: [42]. This consistency indicates that misexpression and differences in expression level due to altered cell-composition associated with sterility phenotypes were not so severe that they obscured quantitative expression differences between musculusPWD and domesticusWSB.
The Dobzhansky-Muller model predicts that each hybrid sterility locus will have one or more interaction partners. Mapping of genetic interactions generally requires sample sizes larger than the 305 F2s analyzed here. To increase power, we treated trans eQTL hotspots as candidate hybrid sterility loci and searched for interactions involving them. We performed conditional mapping of eQTL, using genotypes at candidate loci one at a time as covariates. Genotype covariates included the marker closest to the peak of each of the nine autosomal trans eQTL hotspots, and five markers in the X chromosome trans hotspots (Table 5). For each covariate, mapping was performed twice, including an additive effect or both an additive and interactive effect; eQTL from the full model that showed a significant increase in LOD score over the additive model were classified as significant interaction eQTL.
Clustering of interaction eQTL identified by conditional mapping was even more pronounced than clustering of trans eQTL in the initial (no covariate) eQTL analysis (Fig. S3). We identified ‘interaction hotspots’ using significance thresholds from permutation for each genotype covariate. Integrating results from the conditional mapping analyses reveals a complex epistatic network showing several general patterns (Fig. 4). The large number of interactions involving the X is consistent with its substantive effect on expression pattern and sterility phenotypes. There are many interactions between loci in trans hotspots, and between trans hotspots and sterility QTL, suggesting that some incompatibilities contribute to multiple phenotypes. Overall, a large proportion of interactions are associated with sterility loci. It is important to note that many interactions may be associated with variation in gene expression unrelated to hybrid sterility.
The interactions we identified include X-autosome pairs previously associated with hybrid sterility. We identified interaction hotspots in the proximal region of chromosome 17, which encompasses Prdm9, from conditional mapping using all X-linked genotype covariates; conversely, mapping conditional on Chr17@13 cM identified a hotspot on the proximal X (Fig. S4). Previous mapping of sterility phenotypes conditional on X genotypes revealed interactions between the X and six autosomal regions on four chromosomes (3, 5, 7, 10), contributing to five sperm morphology phenotypes [1]. We found interaction hotspots involving at least one X-linked covariate overlapping each of these autosomal regions.
Each trans hotspot identified in the original analysis overlapped at least one interaction hotspot mapped with an autosomal covariate, indicating autosome-autosome interactions contribute substantially to expression variation. All of these interactions are novel. Interactions between regions with sterile alleles from the same subspecies are prevalent (Fig. 5), suggesting incompatibilities involving more than two loci are common.
Conditional mapping revealed additional associations between gene expression variation and sterility. Some sterility QTL that did not overlap a trans hotspot identified in the original analysis showed evidence for interaction with one or more hotspot regions (Fig. S4). We also found interactions with sterility QTL for each of the trans hotspots that do not overlap sterility QTL. The relative contribution of loci to expression variation with detectable marginal effects versus eQTL identified only when incorporating interactions varied (Table 5). The structure of the interaction network provides additional support for the important roles of chromosomes X and 17, the major players in F1 sterility (Figs. 4; 5). By contrast, the chromosome 6 region plays a prominent role in the interaction network (Fig. 5), which was unanticipated on the basis of relatively modest enrichment of eQTL in the trans hotspot and the lack of sterility phenotype QTL on chromosome 6.
We identified several novel loci that interact with multiple trans hotspots but did not have previous evidence for involvement in sterility (Fig. S4). Regions on chromosomes 7 (50–52 cM; 122.63–125.77 Mb), 13 (32–36 cM; 68.47–75.96 Mb), 14 (40–44 cM; 87.59–97.00 Mb) and 16 (0–4 cM; 11.20–20.02 Mb) had overlapping interaction hotspots identified by mapping with genotype covariates from trans hotspots on at least three chromosomes. These results indicate that some loci in the interaction network have marginal effects undetectable using single-QTL models and permutation thresholds.
The Dobzhansky-Muller model of reproductive isolation has been well accepted for decades but relatively few incompatible loci and even fewer interactions are known. Due to the central role of negative epistasis in hybrid defects, disruptions in gene networks are likely to be common in hybrids [45]–[47]. Inspired by recent ‘systems genetics’ studies that integrate phenotype, genotype, and gene expression data to reconstruct gene networks and infer relationships between perturbations in networks and deleterious traits [48], [49], we mapped expression traits in an F2 cross between house mouse subspecies. We combined expression-mapping results with knowledge of QTL for sterility phenotypes in the same cross to identify altered expression patterns reflecting disruptions in networks causing sterility.
The importance of evolutionary changes in transcriptional regulation for adaptation has long been recognized [e.g. 50]–[53]. Recent studies of gene expression in hybrids suggest regulatory evolution may also be an important cause of reproductive isolation between diverging populations. Misexpression has been reported in hybrids from many animal and plant taxa including Drosophila [8], [12], [54], mice [9]–[11], [55], African clawed frogs [13], [56], whitefish [57], copepods [58], maize [59], ragwort [60] and Arabidopsis [61]. Furthermore, several known hybrid incompatibility genes affect transcription of other genes, including OdsH [12] and the mouse sterility gene Prdm9 [27]. Our expression data from F1 and F2 hybrids show male sterility is associated with major alterations in genome-wide expression patterns. Clustering of trans eQTL is much less pronounced when mapping is restricted to fertile mice (Fig. S2), indicating trans hotspots in particular are associated with sterility. Each of the trans hotspots we identified overlaps a sterility QTL and/or interacts with at least one region containing a sterility QTL. One interpretation of this pattern is that divergent alleles with major effects on expression patterns are likely to cause hybrid incompatibilities. Trans regulators of gene expression must coordinate properly with cis regulators and other trans factors. The number and broad genomic distribution of regulated genes and co-factors provide many potential opportunities for incompatible interactions resulting in deleterious phenotypes in hybrids. Misexpression of a gene could result from a change in the set of positive or negative regulatory factors, or a mismatch in the spatiotemporal availability of these factors and the timing of expression. This hypothesis suggests genes in interacting regions with large cis eQTL and/or major alterations in spatiotemporal expression pattern between subspecies should be prioritized as candidates.
Numerous studies of F1 hybrid sterility and evidence for reduced gene flow in hybrid zones have shown that the X chromosome plays a central role in hybrid male sterility in house mice [5], [62]–[65]. Our expression mapping results in F2s show that the X has a massive effect on testis gene expression, providing support for an important role of the X beyond the F1 generation. Most of the X chromosome is significantly enriched for QTL affecting expression of autosomal genes.
The musculusPWD allele in the proximal X hotspot (10.16 Mb–101.19 Mb) has effects on expression suggestive of sterility (Table 2), consistent with the well-documented role of the musculus X in F1 sterility. This region harbors the largest-effect QTL identified for testis weight, sperm count, abnormal sperm head morphology, and number of offspring in X introgression experiments [25], [66]. Genes with functions related to fertility (sexual reproduction, fertilization, flagellum) were enriched among the QTT with low expression caused by the musculusPWD allele (Table 3).
By contrast, the distal X hotspot shows little similarity to patterns observed in sterile F1 males. The distal hotspot overlaps several sterility QTL identified in Xmusculus introgression experiments (Supp. Table S2), but the domesticusWSB allele at hotspot eQTL is associated with the sterile expression pattern. These results reveal the presence of at least one novel locus on the X contributing to expression variation and potentially F2 sterility (Tables 2, S2). Fertility of DxM F1s, which carry the domesticusWSB X, and lack of enrichment of the distal hotspot QTT for transcripts misexpressed in F1s, indicate this locus interacts with one or more recessive musculusPWD autosomal loci. DNA-binding genes are enriched among QTT with higher expression, raising the possibility that the distal locus controls expression of regulatory genes, and the role in sterility is indirect.
Variation within the trans hotspots on the X suggests each may harbor more than one sterility gene. The number of eQTL mapped, and the proportions of QTT with sterility-related characteristics, varied within the proximal and distal hotspots (Table S2). Furthermore, comparison of conditional mapping results using different markers on the X as covariates reveals differences in interaction patterns (Fig. S5).
We identified a region on chromosome 17 with major effects on gene expression. Several lines of evidence implicate the known sterility gene Prdm9 as the underlying causative gene. First, the QTL for overexpression of X-linked transcripts (18.46 Mb) and the peak in number of trans eQTL within the chromosome-17 hotspot (14.69 Mb) are near Prdm9 (15.68 Mb; Fig. 3B). Second, eQTL in the chromosome-17 hotspot largely show under- or overdominant effects, in contrast to trans eQTL elsewhere in the genome, which are mostly additive or dominant (Fig. 3A). This pattern is consistent with results from F1 crosses showing the most severe sterility phenotypes occur in males heterozygous at Prdm9 [67]. Finally, we find evidence for interactions between the chromosome-17 region and a musculusPWD allele on the proximal X chromosome, consistent with F1 studies [4].
If Prdm9 is the causative gene, our eQTL results provide novel insights into its role in hybrid sterility and gene regulation. In addition to the known interaction with the X chromosome, we find evidence for interaction with each autosomal locus used as a mapping covariate (Figs. S4; 5). The large number of interacting loci suggests that the DNA-binding function of Prdm9, which regulates recombination hotspots globally [73], [74], might be directly related to its role in sterility. Each Prdm9-binding site represents a potential incompatibility partner. Alternatively, disrupted regulation caused by Prdm9 might have cascading effects resulting in altered expression genome-wide.
Although Prdm9 is predicted to have broad regulatory effects, previous evidence for effects on expression levels was limited to a small set of genes directly regulated by Prdm9 [27]. The combination of eQTL in the chromosome-17 hotspot (without covariates; Table 2) and eQTL dependent on interactions with eight autosomes and the X chromosome (Table 5) identifies 5,467 unique transcripts directly or indirectly affected by the region encompassing Prdm9.
Chromosome 17 harbors a second, more distal sterility locus, Hstws, from musculus [18]. Hstws is necessary, in addition to the sterile Prdm9domesticus allele and the musculus X, to observe complete meiotic arrest, the most severe F1 phenotype [67]. We identified interactions between both the Prdm9 region and a distal chromosome 17 region with chromosomes 2, 5, 10, and X (Fig. S4), suggesting loci on those chromosomes may be involved in the Prdm9- Hstws incompatibility.
Overlap of sterility QTL with trans hotspots and/or interaction hotspots can refine estimates of the QTL position in some cases. For example, the trans hotspot on chromosome 17 is smaller than the coincident QTL for sperm count and testis weight (Fig. 3B). Moreover, the peak in number of trans eQTL is at the position closest to Prdm9. Chromosomes 5 and 10 are cases where trans eQTL and interaction eQTL patterns appear particularly useful in narrowing lists of candidate genes (Fig. S2)
Functional annotation of QTT identifies affected pathways and processes associated with some hotspots, and provide clues about the mechanisms underlying sterility. Chromatin-related genes were overrepresented among QTT with lower expression associated with the sterile domesticusWSB allele at the chromosome 11 hotspot (Table 3). Mouse knockout models for two additional genes with eQTL in this region have spermatogenesis defects that might be related to chromatin; males with null alleles at the transcription factor Crem (cAMP responsive element modulator) showed defective spermiogenesis with aberrant post-meiotic gene expression [75]. Lmna (lamin A) knockouts have severely impaired spermatogenesis associated with failed chromosomal synapsis [76]. These patterns suggest prioritizing genes in the chromosome 11 hotspot with related functions. For example, 42 genes are involved in transcriptional regulation (Table 4). One of these genes (Hils1) is involved in chromatin remodeling during spermatogenesis and has evolved rapidly within rodents [77]. Males with hypomorphic Rad51c alleles are infertile due to arrest of spermatogenesis in early meiotic prophase I related to failed double-strand break repair by recombination [78].
Interactions between novel loci and better-characterized regions point to some promising candidates. For example, the chromosome-10 hotspot interacts with the proximal X and the chromosome-17 region containing Prdm9, the two loci with the most dramatic effects on expression. A gene within the chromosome-10 hotspot, Dnmt3l (DNA methyltransferase 3-like), plays a key role in epigenetic programming during spermatogenesis. Males carrying null alleles at Dnmt3l show phenotypes similar to those documented in F1s associated with the X-17 interaction, including hypogonadism, asynapsis during meiosis, abnormal formation of the sex body, and deregulation of X-linked and autosomal genes [79]–[82]. Dnmt3l does not have methyltransferase activity but shows sequence similarity to Dnmt3a and Dnmt3b, with which it interacts to promote de novo DNA methylation [83]. Misexpression of Dnmt3a was reported previously in sterile F1 hybrids [10]. Prdm9 is a histone methyltransferase; while speculative, an interaction between Dnmt3l and Prdm9 is a promising lead. Dnmt3l is essential for several epigenetic processes occurring at different stages of spermatogenesis, including paternal imprinting, transcriptional regulation, chromatin morphogenesis through meiosis, and the histone-protamine transition during spermiogenesis. Interestingly, Dnmt3l interacts with heterochromatin [80], similar to the Drosophila sterility gene OdsH [84].
Conditional mapping revealed several genomic regions involved in the interaction network that did not have previous evidence for involvement in sterility or expression (Fig. 5), indicating this mapping approach can uncover incompatibility loci without detectable marginal effects. Some of these interaction loci are very small, containing few enough genes that targeted functional evaluation would be feasible. For example interaction hotspots mapped with covariates on chromosomes 2, 5, and X overlap on distal chromosome 7 (Fig. S4). This region spans 3.1 Mb, encompassing 14 characterized RefSeq genes.
We focused here on genome-wide patterns. Detailed characterization of individual loci, and analysis of gene co-expression networks including all related QTT, will yield additional information useful in pinpointing the disrupted pathways causing sterility and prioritizing candidates.
The rate of accumulation of Dobzhansky-Muller incompatibilities and the evolution of reproductive barriers between incipient species depend on the genetic architecture of isolating traits. Theoretical models of DMI evolution assume that incompatibilities act independently on barrier traits [85], [86]. The complex pattern of interactions we report here violates this assumption: some sterility loci are involved in multiple incompatibilities. This aspect of the network we characterized is most consistent with branched developmental pathways [45] and gene networks models [47]. Theory that incorporates this non-independence as well as other biological characteristics of incompatibilities should continue to be pursued [45]–[47], [87].
Network characteristics are also key determinants in accurate modeling of gene-flow dynamics in zones of hybridization. Non-independence of incompatibilities due to interactions of sterility loci with multiple partners is likely to result in stronger selection and slower introgression at those loci because sterility phenotypes are expressed on a variety of genomic backgrounds. Future cline theory should incorporate epistasis with multiple partner loci.
Remarkable progress in understanding the genetic basis of speciation has emerged from identification of a growing list of hybrid incompatibility genes [6] over the past 20 years. However, identification and functional characterization of hybrid incompatibility genes is feasible in only a few model organisms, and tremendous effort, time and resources are needed to identify a single gene. If this gene-by-gene approach continues as the standard in speciation genetics, it will be a long time before the number of genes and interactions identified is sufficient to reveal generalities of the speciation process. Moreover, general features of incompatibility networks, including the number and dominance of loci, types of interactions, and possibly particular developmental/regulatory pathways, are more likely to be shared among taxa than are specific incompatibility genes.
The house mouse features a rich set of sophisticated genetic tools and resources, which facilitates collection of reliable genome-scale data and ultimately will enable functional characterization of candidate incompatibility genes. Although identification and characterization of reproductive barrier genes is not feasible in most species, the integrated mapping approach we employed is applicable in a broad range of organisms. For species pairs that can be crossed in the laboratory, a similar F2 intercross can be performed and sterility or inviability phenotypes can be measured. Informative marker discovery is straight-forward and relatively low cost using RADseq [88], and RNAseq or custom microarrays can be used to collect expression data from species without commercially available platforms. Functional annotation and nomination of candidate processes/pathways is possible if a genome sequence of the focal species or even a relatively distantly related taxon is available [89]. Even in species with very limited available gene annotation, the number of incompatibility loci and the nature of interactions between them can be estimated. Consequently, we suggest that network-centered approaches are powerful and have promise to substantially advance understanding of speciation.
Mice were maintained at the University of Wisconsin School of Medicine and Public Health mouse facility according to animal care protocols approved by the University of Wisconsin Animal Care and Use Committee.
Reciprocal crosses of wild-derived inbred strains of M. m. domesticus (WSB/EiJ; domesticusWSB) and M. m. musculus (PWD/PhJ; musculusPWD) were performed to generate F1 hybrids. A total of 305 F2 males were generated by mating F1 siblings (294 from domesticusWSB female×musculusPWD male crosses and 11 from musculusPWD female×domesticusWSB male crosses). Male F2s were euthanized at 70 (±5) days of age. Five sterility phenotypes were quantified: testis weight, sperm count, sperm head shape, proportion of abnormal sperm, and seminiferous tubule area (see White et al 2011 for detailed methods). The left testis was flash-frozen in liquid nitrogen upon dissection and stored at −80°. Testes from musculusPWD (n = 8), domesticusWSB (n = 8), musculusPWD×domesticusWSB F1s (n = 6), and domesticusWSB×musculusPWD F1s (n = 4), were dissected using the same procedure to provide controls for expression analyses. Frozen testis samples were transferred to RNAlater-ICE buffer (Invitrogen, Grand Island, NY, USA), shipped to the Max Planck Institute in Plön and stored at −80° until processing.
To identify classes of genes enriched among QTT, we used the DAVID functional annotation tool [43], [44], which integrates gene annotation information from several resources. Functionally related genes are clustered based on biological process, cellular compartment, molecular function, sequence features, protein domains, and protein interactions. To account for multiple comparisons, we used a significance threshold based on the false discovery rate (Benjamini) calculated within DAVID.
We identified candidate genes in trans hotspots and among QTT that have roles in male reproduction and/or regulation of gene expression using reviews of male fertility [17] and meiosis [105] and gene ontology (GO) terms related to male reproduction, meiosis, or the regulation of gene expression: 0001059; 0001060; 0001109; 0001121; 0003006; 0006351; 0006352; 0006353; 0006354; 0006355; 0006360; 0006366; 0006383; 0006390; 0006396; 0006412; 0007127; 0007135; 0007140; 0007285; 0009008; 0009299; 0009300; 0009302; 0009304; 0010216; 0010468; 0010608; 0010628; 0010629; 0022414; 0023019; 0030724; 0030726; 0032775; 0032776; 0036206; 0040020; 0040029; 0042793; 0043046; 0043484; 0044030; 0045132; 0045835; 0045836; 0045892; 0045893; 0048133; 0048136; 0048140; 0048515; 0048610; 0050684; 0051037; 0051257; 0051604; 0070192; 0070613; 0070920; 0080188; 0090306; 0097393; 1901148; 1901311; 2000232; 2000235; 2000241; 2000242; 2000243. Many genes identified as candidates in publications were not annotated with related GO terms, highlighting the limitations of gene ontology. Moreover, genes causing sterility might not have functions obviously related to reproduction.
|
10.1371/journal.pgen.1003349 | Human Spermatogenic Failure Purges Deleterious Mutation Load from the Autosomes and Both Sex Chromosomes, including the Gene DMRT1 | Gonadal failure, along with early pregnancy loss and perinatal death, may be an important filter that limits the propagation of harmful mutations in the human population. We hypothesized that men with spermatogenic impairment, a disease with unknown genetic architecture and a common cause of male infertility, are enriched for rare deleterious mutations compared to men with normal spermatogenesis. After assaying genomewide SNPs and CNVs in 323 Caucasian men with idiopathic spermatogenic impairment and more than 1,100 controls, we estimate that each rare autosomal deletion detected in our study multiplicatively changes a man's risk of disease by 10% (OR 1.10 [1.04–1.16], p<2×10−3), rare X-linked CNVs by 29%, (OR 1.29 [1.11–1.50], p<1×10−3), and rare Y-linked duplications by 88% (OR 1.88 [1.13–3.13], p<0.03). By contrasting the properties of our case-specific CNVs with those of CNV callsets from cases of autism, schizophrenia, bipolar disorder, and intellectual disability, we propose that the CNV burden in spermatogenic impairment is distinct from the burden of large, dominant mutations described for neurodevelopmental disorders. We identified two patients with deletions of DMRT1, a gene on chromosome 9p24.3 orthologous to the putative sex determination locus of the avian ZW chromosome system. In an independent sample of Han Chinese men, we identified 3 more DMRT1 deletions in 979 cases of idiopathic azoospermia and none in 1,734 controls, and found none in an additional 4,519 controls from public databases. The combined results indicate that DMRT1 loss-of-function mutations are a risk factor and potential genetic cause of human spermatogenic failure (frequency of 0.38% in 1306 cases and 0% in 7,754 controls, p = 6.2×10−5). Our study identifies other recurrent CNVs as potential causes of idiopathic azoospermia and generates hypotheses for directing future studies on the genetic basis of male infertility and IVF outcomes.
| Infertility is a disease that prevents the transmission of DNA from one generation to the next, and consequently it has been difficult to study the genetics of infertility using classical human genetics methods. Now, new technologies for screening entire genomes for rare and patient-specific mutations are revolutionizing our understanding of reproductively lethal diseases. Here, we apply techniques for variation discovery to study a condition called azoospermia, the failure to produce sperm. Large deletions of the Y chromosome are the primary known genetic risk factor for azoospermia, and genetic testing for these deletions is part of the standard treatment for this condition. We have screened over 300 men with azoospermia for rare deletions and duplications, and find an enrichment of these mutations throughout the genome compared to unaffected men. Our results indicate that sperm production is affected by mutations beyond the Y chromosome and will motivate whole-genome analyses of larger numbers of men with impaired spermatogenesis. Our finding of an enrichment of rare deleterious mutations in men with poor sperm production also raises the possibility that the slightly increased rate of birth defects reported in children conceived by in vitro fertilization may have a genetic basis.
| Male infertility is a multifaceted disorder affecting nearly 5% of men of reproductive age. In spite of its prevalence and a considerable research effort over the past several decades, the underlying cause of male infertility is uncharacterized in up to half of all cases [1]. Some degree of spermatogenic impairment is present for most male infertility patients, and, in its most severe form, manifests as azoospermia, the lack of detectable spermatozoa in semen, or oligozoospermia, defined by the World Health Organization as less than 15 million sperm/mL of semen. Spermatogenesis is a complex multistep process that requires germ cells to (a) maintain a stable progenitor population through frequent mitotic divisions, (b) reduce ploidy of the spermatogonial progenitors from diploid to haploid through meiotic divisions, and (c) assume highly specialized sperm morphology and function through spermiogenesis. These steps involve the expression of thousands of genes and carefully orchestrated interactions between germ cells and somatic cells within the seminiferous tubules [2]. It is likely that a large proportion of idiopathic cases of spermatogenic failure are of uncharacterized genetic origin, but measuring the heritability of infertility phenotypes has been challenging.
Known genetic causes of non-obstructive azoospermia (NOA) include deletions in the azoospermia factor (AZF) regions of the Y chromosome [3], Klinefelter's syndrome [4], and other cytogenetically visible chromosome aneuploidies and translocations [5]. Beyond these well-established causes, which are observed in 25–30% of cases, the genetic architecture of spermatogenic impairment is currently unknown. One might expect a priori that rare or de novo, large effect mutations will be the central players in genetic infertility, and indeed other primary infertility phenotypes like disorders of gonadal development, isolated gonadotropin-releasing hormone deficiency, and globozoospermia, a disorder of sperm morphology and function, appear to be caused by essentially Mendelian mutations operating in a monogenic or oligogenic fashion [6], [7], [8]. Similarly, recurrent mutations of the AZF region on the Y chromosome are either completely penetrant (AZFa, AZFb/c) or highly penetrant (AZFc) risk factors for azoospermia. Our working model at the start of this study was that additional “AZF-like” loci existed in the genome, either on the Y chromosome or elsewhere, and that, much like recent progress in the analysis of developmental disorders of childhood, a large number of causal point mutations and submicroscopic deletions could be revealed in idiopathic cases by the appropriate use of genomic technology.
In this paper, we employ oligonucleotide SNP arrays as discovery technology to conduct a whole-genome screen for two rare genetic features in men with spermatogenic failure. First, we extract and analyze the probe intensity data to find rare copy number variants (CNVs). A growing number of CNVs have been associated with a host of complex disease states [9] including neurological disorders [10], [11], [12], [13], several autoimmune diseases [14], [15], type 2 diabetes [16], cardiovascular disease [17], and cancer [18], [19], [20], [21]. Now, a role for CNVs in male infertility is beginning to emerge [22], [23], [24], [25].
As a second approach to identify rare genetic variants, we use a population genetics modeling framework to identify large homozygous-by-descent (HBD) chromosome segments that may harbor recessive disease alleles. When applied to consanguineous families, so-called “HBD-mapping” has been an unequivocal success in identifying the location of causal variants for simple recessive monogenic diseases [26]. HBD analysis can also be used to screen for the location of rare variants in common disease case-control studies of unrelated individuals, using either a single-locus association testing framework or by testing for an autozygosity burden, frequently referred to as “inbreeding depression”: an enrichment of size or predicted functional impact of HBD regions aggregated across the genome. This approach has produced results for a growing list of common diseases, including schizophrenia [27], Alzheimer's disease [28], breast and prostate cancer [29].
In this study, we screened three cohorts of men with idiopathic spermatogenic failure in an attempt to identify rare, potentially causal mutations, and to better understand the genetic architecture of the disease (Table 1). We found a genomewide enrichment of large, rare CNVs in men with spermatogenic failure compared to normozoospermic or unphenotyped men (controls). We also identify a number of cases with unusual patterns of homozygosity, possibly the result of recent consanguineous matings. Our results show that spermatogenic output is a phenotype of the entire genome, not just the Y chromosome, place spermatogenic failure firmly among the list of diseases that feature a genomewide burden of rare deleterious mutations and provide a powerful organizing principle for understanding male infertility.
First, we attempted to find evidence for undiscovered dominant causes of spermatogenic failure by studying the genomewide distribution of CNVs in our primary cohort from Utah: 35 men with idiopathic non-obstructive azoospermia, 48 men with severe oligozoospermia, and 62 controls with normal semen analysis. All cases had previously tested negative when screened for canonical Y chromosome deletions. Samples were assayed with an Illumina 370K oligonucleotide array that provides both SNP and CNV content. There was no detectable difference in the average number of CNVs called per sample among the three groups (mean = 20, azoospermic; 19.5, oligozospermic; 20, normozoospermic), however, the majority of variants (61% on average) in any one sample were common polymorphisms.
When restricting our analysis to CNVs with a call frequency of less than 5%, a subset likely to be enriched for pathogenic events, we observed pronounced differences among groups (Table S1). Azoospermic and oligozoospermic men have nearly twice the amount of deleted sequence genomewide when compared to controls (p = 1.7×10−4, Wilcoxon rank sum test), and a nonsignificant 12% increase in the number of deletions per genome. When examining the even more restricted set of rare CNVs larger than 100 kb (Dataset S1), these associations are more pronounced: the rate of deletions in cases was twice that of controls (1.12 vs. 0.55, p = 9.7×10−4) and the amount of deleted sequence 2.6 times greater in cases (p = 8.8×10−4).
In order to replicate these initial findings, we assayed two additional cohorts – one group of 61 Caucasian men with severe spermatogenic impairment and 100 ethnicity-matched, unphenotyped controls, both collected at Washington University in St. Louis (WUSTL), and a larger case cohort of 179 Caucasian men with idiopathic azoospermia, primarily from medical practices in Porto, Portugal, matched to an unphenotyped control set of 974 Caucasian men collected by the UK National Blood Service (NBS, [30]). Although using different array platforms (Text S1), we observed replication of our initial association (Table S2 and Table S3); in the WUSTL cohort a 20% increase in the rate (p<0.05) and in the Porto cohort a 31% increase in rate (p<5×10−3). We excluded several artifactual explanations for this burden effect, including specific batch phenomena or population structure (Text S1, Figures S1, S2, S3, S4, S5). To better characterize these genomewide signals, we set out to search for clustering of pathogenic mutations on specific chromosomes.
We focused first on the Y chromosome as it is the location of most known mutations modulating human spermatogenesis (Figure 1, Figure S6). Y-linked microdeletions of the AZFa, AZFb, and AZFc regions are well-established causes of spermatogenic impairment, and thus we excluded from this study cases with AZF microdeletions visible by STS PCR. In the array data, we found no significant difference in the frequency of rare Y deletions between case and controls groups; however rare duplications were more abundant in Porto cases compared to the NBS controls (a 3-fold enrichment in Porto cohort, p = 1.9×10−3). We could classify the majority (>90%) of our samples to major Y haplogroups using SNP genotypes (Text S1), and, as expected, most of these samples fall into the two most common European haplogroups: I (22%) and R (70%). The observed duplication burden was not an artifact of differences in major Y haplogroup frequency between cases and controls, as association was essentially unchanged when only considering samples with haplogroup R1 (p = 3.3×10−3). Due to low probe coverage, only one Y-linked duplication was called in the Utah cohorts (in a control individual) and two in the WUSTL cohort (both in cases), so this burden of Y duplications was not replicated.
Next we turned to the X chromosome, which is highly enriched for genes transcribed in spermatogonia [31]. In the Utah cohorts there were 71 gains and losses with a frequency of less than 5% on the X chromosome, cumulatively producing three times as much aneuploid sequence in azoospermic and oligozoospermic men compared with normozoospermic men (89 kb/person azoo, 45 kb/person oligo, 27 kb/person normozoospermic men, all cases versus controls p<0.03). This burden was strongly replicated in the Porto samples, which displayed a 1.6 fold enrichment of rare CNV on the X (p = 5×10−4) and the WUSTL samples (31% of cases with a rare X-linked CNV versus 16% of controls, p = 0.02 by permutation).
The genome-wide signal of CNV burden was not driven solely by sex chromosome events: considering only autosomal mutations in Utah samples there was an enrichment of aneuploid sequence in large deletions in azoospermic men (268 kb/person) and oligozoospermic men (308 kb/person) compared to control men (189 kb/person, p = 9.8×10−3), and an enrichment in the rate of deletions in all cases when considering just events >100 kb (1.9 fold enrichment, p = 6×10−3). In the Porto cohort, there was modest evidence for a higher rate of rare deletions of all sizes in azoospermic men (1.27 fold enrichment, not significant) as well as an increase in total amount of deleted sequence (345 kb/case vs. 258 kb/control, p<0.003).
In order to cleanly summarize our findings across all cohorts, we fit logistic regression models for each cohort, regressing case status onto CNV count for different classes of CNV. We also fit a linear mixed-effects logistic regression model to the total dataset for each CNV class, treating cohort as a random factor (Figure 1). In each regression model we controlled for population structure by including eigenvectors from a genomewide principal components analysis (Methods). On the basis of the combined analysis, we estimate that each rare autosomal deletion multiplicatively changes the odds of spermatogenic impairment by 10% (OR 1.10 [1.04–1.16], p<2×10−3), each rare X-linked CNV (gain or loss) by 29%, (OR 1.29 [1.11–1.50], p<1×10−3) and each rare Y-linked duplication by 88% (OR 1.88 [1.13–3.13], p<0.03).
Deletions of the AZF regions of the Y chromosome are often mediated by non-allelic homologous recombination (NAHR) between segmental duplications and are the most common known cause of spermatogenic failure. Because of their prognostic power and high rate of recurrence in the population, screening for AZF deletions is a standard part of the clinical workup for azoospermia. It would be of high clinical value if additional azoospermia susceptibility loci with significant recurrence rates could be identified.
We screened all cohorts for large (>100 kb) rearrangements flanked by homologous segmental duplications capable of generating recurrent events by NAHR [32]. There was no significant enrichment of gains or losses in cases across these hotspot regions when considered as an aggregate. Due to small sample sizes we found no single-locus associations, at these hotspot loci, or elsewhere, that met the strict criteria of genomewide significance in both the discovery and replication cohorts. Many of our single-cohort associations from one platform lack adequate probe coverage on other platforms for robust replication (Text S1). However, several loci were significant on joint analysis of all cohorts.
The best candidate for a novel locus generating NAHR-mediated infertility risk mutations is a 100 kb segment on chromosome Xp11.23 flanked by two nearly identical (>99.5% homology) 16 kb segmental duplications containing the sperm acrosome gene SPACA5 (Figure 2a, Figure S7). We identified 9 deletions of this locus spread across all patient cohorts (3 in PT, 1 in UT, 5 in WUSTL) compared to 8 in the pooled 1124 controls (2.8% frequency versus 0.7%, odds ratio = 3.96, p = 0.005, Fisher exact test). We genotyped the deletion by +/− PCR in an additional cohort of 403 men with idiopathic NOA from Weill Cornell, and observed an additional 3 deletions (Figure S8, Text S1). In a prior case-control study of intellectual disability, investigators using qPCR estimated the allele frequency of this deletion to be 0.47% (10/2121) in a large Caucasian male control cohort [33]. Combining these data, we estimate the allele frequency of the deletion to be 1.6% in Caucasian cases, compared to 0.55% in Caucasian controls (OR 3.0, 95% CI 1.31–6.62, p = 0.007). The deleted region contains the X-linked cancer-testis (CT-X) antigen gene SSX6; the CT-X antigen family is a highly duplicated gene family on the X chromosome comprising 10% of all X-linked genes and is expressed specifically in testis. After controlling for differences in coverage across the array platforms used in this study, we find a significant enrichment of rare deletions of CT-X genes in all cases (p = 0.02); this finding did not extend to duplications or CT antigen genes on the autosomes (Table 2).
When analyzing all cohorts jointly, our strongest association (genomewide corrected p-value <0.002) is to both gains and losses involving a 200 kb tandem repeat on Yq11.22, DYZ19 (Figure S6, Figure S9), a human-specific array of 125 bp repeats first discovered as a novel band of heterochromatin in the Y chromosome sequencing project [34]. Tandem repeat arrays are often highly unstable sequence elements that can mutate by both replication-based and recombination-based (e.g. NAHR) mechanisms. In our data there were 9 gains and 11 losses at DYZ19 in 323 cases (combined frequency 6.1%), compared to 3 gains and 12 losses in 1136 controls (combined frequency 1.3%). While this finding may ultimately require painstaking technical work to conclusively validate, we have several reasons to believe the association is real. First, we have previously shown that it is possible to identify real copy number changes at VNTR loci using short oligonucleotide arrays [35]; second, copy number changes at this locus were identified by multiple platforms in the current study; third, the association is nominally significant in both the Utah and Porto cohorts; fourth the locus is within the AZFb/c region. The direction of copy number changes does appear to track with haplogroup – while 12/13 duplications occur on the R1 background, 14/15 deletions for which haplogroup could be determined occur on I or J background. Haplogroup assignments for the carriers of these CNVs were confirmed by standard short tandem repeat analysis (Text S1). The strong association between haplogroup and direction of copy number change is noteworthy; it may indicate that DYZ19 CNVs are merely correlated with other functional changes on these chromosomes, or perhaps the structure of these chromosomes predisposes them to recurrent gains (R1) or losses (I/J).
The gene DMRT1 is widely believed to be the sex-determination factor in avians, analogous to SRY in therians, and may play the same or similar role in all species that are based upon the ZW sex chromosome system [36]. DMRT1 encodes a transcription factor that can activate or repress target genes in Sertoli cells and premeiotic germ cells through sequence-specific binding [37]. In humans, DMRT1 is located on 9p24.3 in a small cluster with the related genes DMRT2 and DMRT3. Large terminal deletions of 9p are a known cause of syndromic XY sex-reversal, and although the role of the DMRT genes in the 9p deletion syndrome phenotype has not yet been defined, mouse experiments have shown that homozygous deletion of DMRT1 causes severe testicular hypoplasia [38], [39], [40].
We found two, perhaps identical, 132 kb deletions spanning DMRT1 in the Utah cohort in men with azoospermia, and a 1.8 Mb terminal duplication of 9p, spanning these genes, was seen in a single normozoospermic control from Utah (Figure 2b). All three of these rearrangements were validated by TaqMan assay (Figure S10, Text S1). Both men were recruited into the study in Salt Lake City, UT between 2002 and 2004. They self-reported their ancestry as Caucasian, and in both cases this assumption was clearly verified by principal components analysis of their genetic data (Figure S2). There was no evidence that the two deletion carriers were closely related upon comparison of their whole-genome SNP genotypes. Testis biopsies were performed on both men; these indicated apparent Sertoli cell only syndrome in the first and spermatocytic arrest in the second. Both men exhibited apparently normal male habitus and virilization with no phenotypic similarities to 9p deletion syndrome.
We obtained Affymetrix 6.0 array data from a previously published genomewide association study of idiopathic NOA in Han Chinese [41] comprised of 979 cases and 1734 controls (Text S1). After processing these samples with our CNV calling pipeline, we observed an additional 3 deletions of DMRT1 exonic sequence in cases (0.3%) and none in controls (Figure 2B, Figure S11). From these combined array data we estimate a frequency of DMRT1 exonic deletion of 0.38% (5/1306) in cases and 0% (0/2858) in controls (OR = Infinity, [2.0-Inf], p = 0.003). We obtained the two largest control SNP array datasets in the Database of Genomic Variants (DGV), representing CNV calls from 4519 samples typed with platforms of equal or higher probe density to the ones used here [42], [43]. None of these samples contained CNV of any sort affecting DMRT1. Finally, we screened an additional set of 233 idiopathic NOA cases from Weill Cornell, and 135 controls with the TaqMan validation assay and identified an additional 3 deletions (2 in cases, 1 in controls, Text S1, Figure S12). As this qPCR assay interrogates intronic sequence, the functional consequences of these 3 deletions are unclear. Our array data have revealed some of the smallest coding deletions of DMRT1 reported to date in humans, and should help to clarify the critical regions of 9p involved in testicular development and function.
Notably, using a bespoke reanalysis of the intensity data, we did not see evidence for CNVs involving the gene PRDM9, a recently characterized zinc finger methyltransferase that appears to control the location of recombination hotspots in a diversity of mammalian species. Heterozygosity of PRDM9 zinc finger copy number has been shown to cause sterility in male hybrids of Mus m. domesticus and Mus m. musculus due to meiotic arrest [44].
The identification of functional or physical annotations enriched in case-associated CNVs can be a powerful step in constructing models to classify pathogenic variants. We searched for significant case-specific aggregation of CNVs in several classes of functional sequence, including 195 genes previously shown to result in spermatogenic defects when mutated in the mouse [45], all protein and non-protein coding genes, and 525 testis genes that are differentially expressed during human spermatogenesis (Text S1). Deletion of X- or Y-linked exonic sequence conferred the strongest risk (OR = 1.87 [1.30–2.68], p<1×10−3). Very similar risk was associated with deletion of exonic sequence from testis genes differentially expressed during spermatogenesis, despite the fact that only 15% of these genes are located on the sex chromosomes (OR = 1.85 [1.01–3.39], p<0.05). Deletion of any exonic sequence was also associated with disease (OR = 1.25 [1.07–1.46], p<5×10−3). Deletion of miRNAs was not associated, nor was deletion of the 195 mouse spermatogenic genes [45], which were very rarely deleted in either cases or controls.
We hypothesized that at least some of the functional impact of CNV burden on fertility was a result of disruption of haploinsufficient (HI) genes, as has been demonstrated for neuropsychiatric and developmental disease [46]. For each singleton deletion in our collections we used a recently described modeling framework to calculate the probability that the deletion is pathogenic due to dominant disruption of a haploinsufficient gene [47]. Much to our surprise, HI scores from deletions in infertility cases were much smaller than those from cases of autism and developmental disorders and in fact indistinguishable from controls (mean HI score −1.16 in controls, −1.02 in all spermatogenic impairment cases, p = 0.49 by Wilcoxon rank sum test; Figure 3). Likewise there was no enrichment of large rearrangements within 45 known genomic disorder regions in cases [46]. In contrast to previously described diseases that feature CNV burden, spermatogenic impairment may be more likely to result from large effect recessive mutations, or perhaps the additive effect of deleterious mutations across many loci. We sought to uncover support for recessive mutation load in our cases by assessing the impact of inbreeding, or elevated rates of homozygosity, on disease risk by applying a population genetic approach to the SNP genotype data from our samples [48].
The major genetic side effect of consanguineous mating is a genome-wide increase in the probability that both paternal and maternal alleles are homozygous-by-descent. This probability is often summarized as the inbreeding coefficient, F, and can be estimated from analysis of pedigree structure or by direct observation of genomewide SNP genotypes.
Due to differences in demographic history and culture, the extent of background homozygosity in the genome is expected to vary when comparing diverse populations throughout the globe. The haplotype modeling algorithms implemented in the software package BEAGLE estimate the background patterns of linkage disequilibrium and homozygosity across a set of samples, allowing population-specific information to be used to assess the evidence that any given section of a genome is likely to be homozygous-by-descent (HBD). During the course of our study we concluded that standard PCA-based approaches to stratification are insufficient to correct for population structure during the analysis of inbreeding, even when using population genetic methods like BEAGLE (Text S1, Figure S13). The problem comes not from spurious identification of HBD, but from spurious association of HBD with disease status when case and controls are sampled from groups with different levels of background relatedness. For instance, in a recent survey of 17 Caucasian cohorts, estimates of the average inbreeding coefficient, F, varied from 0.09% to 0.61%, with UK-based cohorts showing the lowest F and the one Portuguese cohort showing the highest [27]. While PCA-based methods traditionally detect and correct for differences in allele frequencies among groups, we believe that they do not detect differences in inbreeding that can be readily incorporated into a case-control testing framework. In the following section, we use data from 622 healthy adults from Spain, who we believe form a more appropriate control group for the Porto case cohort (Methods, Text S1, Figure S13).
Analyzing each cohort separately, BEAGLE identified 5343 chromosome segments likely to represent HBD regions (HBDRs) across all samples. We excluded low-level admixture as a spurious source of HBD (Figure S3). Only three of these segments were identified as apparent artifacts induced by large heterozygous deletions (287 kb, 817 kb, and 877 kb in size) and were removed before subsequent analyses. As expected, the distribution of HBD across all samples was L-shaped, with the majority of HBDRs shorter than 1 Mb and a few intermediate and very large events observed (Figure 4b). The largest HBDR identified spanned all of chromosome 2 in an azoospermic individual, indicative of uniparental isodisomy of the entire chromosome. Clinical reports of UPD2 are extremely rare – there are 7 previous reports of UPD2 that have been ascertained through association with an autosomal recessive disorder [49]. In each of these cases a recessive disorder that lead to clinical presentation was identified. There is currently no proof of imprinted genes on chromosome 2 from either mouse or human data. We performed whole exome sequencing on this individual, and using a simple scoring scheme based on functional annotation and population genetic data, identified a homozygous missense mutation of the INHBB gene as the most unusual damaging homozygous lesion in the genome of this individual (Figure 5, Text S1). The biology of the INHBB gene product strongly implicates this mutation as a causal factor but without additional functional or epidemiological evidence such a conclusion is speculative (Figure 6).
Setting aside this case of UPD2, we found only modest evidence for an enrichment of homozygosity in men with spermatogenic impairment (Figure 4a, Table 3). Our hypothesis was that, if a large percentage of cases of azoospermia were attributable to large-effect autosomal recessive Mendelian mutations, we would see a corresponding increase in the proportion of cases with large values of F. The average inbreeding coefficient was numerically higher in each case cohort compared to its matched control cohort (Table 3). We used a logistic regression mixed model framework to test for association between autozygosity and disease, while controlling for population structure, fitting models that treated autozygosity as both a categorical variable (e.g. inbreeding coefficient >6.25%, yes or no) and a continuous variable (F, Methods). While the estimated effect of inbreeding on disease risk was positive in every model that we tested, the corresponding odds ratios did not differ significantly from 1 in any version (Table 3). There were fewer than 10 HBD regions shared by 2 or more cases, supporting the model that spermatogenic efficiency has a polygenic basis. We also tested for case-specific aggregation of HBD segments using the same association framework as that used for CNVs. We did not identify any significant patterns. Based on published analyses of small-effect recessive risk mutations in other complex diseases, we believe our current sample size would be underpowered to detect association between very old inbreeding (e.g. due to shared ancestors 15 generations ago). It is possible that large cohorts, consisting of over 10,000 cases, may be needed to accurately estimate the relationship between low-level variation in inbreeding (F values smaller than 0.1) and azoospermia risk, as well as map specific risk alleles [27], [50].
We report here the largest whole genome study to date investigating the role of rare variants in infertility, examining data from 323 cases of male infertility and 1,136 controls. These data demonstrate that rare CNVs are a major risk factor for spermatogenic impairment, and while confirming the central role of the Y chromosome in modulating spermatogenic output, our risk estimates for autosomal and X-linked CNVs indicate that this phenotype is influenced by rare variation across the entire genome. The controls from two of the cohorts were unphenotyped, and given the estimated prevalence of azoospermia (1%), we may have underestimated the risk associated with these large rearrangements.
We observed 5 deletions of DMRT1 coding sequence in cases and none in over 7,000 controls. These deletions ranged in size from 54 kb to over 2 Mb (Table 4). DMRT1 is situated in a region of chromosome 9p that has been identified as a source of syndromic and non-syndromic forms of XY gonadal dysgenesis (GD). The deletions of this region that are associated with syndromic forms of GD are usually 4–10 Mb in size, while isolated GD has been reported for deletions smaller than 1 Mb [40], [51], [52]. Despite frequent involvement of DMRT1 in these putative causal mutations, there is variability in both the phenotypic outcome affiliated with each deletion and the extent of DMRT1 coding sequence contained therein. At least two cases of GD have been linked to deletions near but not overlapping DMRT1 – one 700 kb mutation 30 kb distal to DMRT1 in a case of complete XY GD that was inherited from an apparently normal mother, and a second 260 kb de novo deletion about 250 kb distal to DMRT1 [39], [40]. Both of these deletions overlapped the genes KANK1 and DOCK8. On the other hand, two smaller deletions, one a 25 kb deletion of DMRT1 exons 1 and 2, and one a 35 kb deletion of exons 3 and 4, have been observed in patients with complete GD and bilateral ovotesticular disorder of sexual development, respectively [51], [52]. Based on the clinical records of patients in our current study, there is no chance that our DMRT1 deletion carriers could represent misdiagnosis of a condition as severe as complete XY GD, which presents with the appearance of female genitalia. Indeed, two of our DMRT1 deletion carriers were subject to testicular biopsies. Our observations here suggest that hemizygous deletion of DMRT1 is a lesion that shows variable expressivity that may depend on the sequence of the undeleted DMRT1 allele, variation in other sequences on chromosome 9p, and the state of other factors in the pathways regulating testicular development and function. Strictly speaking, statements that hemizygous deletions of DMRT1 are “sufficient” to cause GD or spermatogenic failure need to be qualified at this point until we gain a better understanding of the effects of genetic background. For instance, in most studies of DMRT1 deletion, the undeleted DMRT1 allele is rarely sequenced. Is the mode of action dominant or recessive?
Deletions of the Y chromosome have long been appreciated as a cause of azoospermia, and we have now shown here that Y-linked duplications are also significant risk factors for spermatogenic failure. The precise definition of the duplication sensitive sequences awaits further investigation. Historically, Y duplications have been much less studied than Y deletions, as +/− STS PCR is the standard assay for assessing Y chromosome copy number variation in both the clinical and research setting. Quantitative PCR methods for measuring Y chromosome gene dosage have been described in the literature, and applied almost exclusively to studying the phenotypic effects of duplication of genes in the AZFc region [53]. Results of these investigations are conflicting, with studies of Europeans reporting no association between AZFc partial duplication and spermatogenic impairment [54], while reproducible associations have been reported in east Asian cohorts [55], [56]. Notably, we identified some duplications on the Y chromosome greater than 2.5 Mb in size, all spanning the AZFc locus (Figure S6), in 8/179 cases (those typed on Affymetrix 6.0), compared to 13/972 controls (OR 3.45 [1.21–9.12], p<0.01). Rearrangements of this size on the autosomes confer staggering risk for other forms of disease; for example, by one recent estimate CNVs larger than 3 Mb have an OR of 47.7 for intellectual disability and/or developmental delay [46]. Our results suggest that Y chromosome structure may be more dosage sensitive than previously appreciated, and we speculate that some genes and non-coding sequences of the Y chromosome may be under stabilizing selection for copy number [57].
Three recent studies have used array-based approaches to characterize CNVs in men with azoospermia. Our finding of an X-linked CNV burden in men with spermatogenic failure has been replicated and described elsewhere [58]. In a second study, Tuttelmann et al. evaluated 89 severe oligozoospermic, 37 azoospermic, and 100 normozoospermic control men using Agilent 244K and 400K arrays and identified a number of CNVs potentially involved in male infertility [24]. Third, Stouffs et al. assayed nine azoospermic men and twenty control samples using the 244K array and followed-up CNVs of interest by q-PCR in up to 130 additional controls [25]. Using the criterion of at least 51% reciprocal overlap, we have identified a number of CNVs in the current study that overlap with case-specific CNVs in the Tuttelmann and Stouffs studies. The majority of these CNVs appear to be relatively common polymorphisms and not case-specific in our larger dataset; however several noteworthy CNVs overlap between studies and are absent, or are present at a very low frequency in controls. For example, Tuttelmann et al. identified a private duplication on Xq22.2 in an oligozoospermic man [24], and we identified an overlapping duplication in an oligozoospermic man from the present study (ChrX:103065826–103205985, NCBI36). These duplications alter the copy number of a small number of testis-specific or testis-expressed variants of histone 2B (H2BFWT, H2BFXP, H2BFM). No CNVs in this region were identified in more than 1600 controls. Tuttelmann et al. also identified an azoospermic man with a deletion and another with a duplication on 8q24.3, encompassing the genes PLEC1 and MIR661 [24]. We identified an oligozoospermic man with a duplication of the same region, affecting the same functional elements (chr8:145064091–145118650, NCBI36). CNVs of this locus are very rare, with a frequency of about 0.005% in our controls and 0.0025% in controls used for a recent study of developmental delay [46]. It is important to note that new variants will frequently be discovered whenever a discovery technology such as array CGH is applied to a new sample set, and the observation that a variant is patient-specific is not in itself remarkable, especially when one is investigating very small sample sizes.
Our observation of low deletion HI scores in cases raises a number of considerations for future studies of the genetics of spermatogenic impairment. We interpret low HI scores in cases as evidence against a widespread role for dominant, highly penetrant deletions in spermatogenic failure. It is possible that our case recruitment, which pre-screened for normal karyotype, may have removed all large HI score events; however our identification of two large HI deletions of WT1 and MAPK1 indicate otherwise (Figure 3). A second concern is that the data used to train the haploinsufficiency prediction algorithm is in part based on features of deletions known to cause dominant pediatric disease, and that an analogous approach trained on fertility phenotypes may lead to different conclusions. There are few examples of dominant loss-of-function mutations causing isolated infertility in humans and only 5 of the >200 mouse infertility mutants described in a previous review showed a phenotype in heterozygous form [45], so fitting a model of a dominant infertility mutation may be challenging in the short term. Nonetheless, developing disease-specific pathogenicity scores for infertility phenotypes should be a priority.
Despite the differences between the genetic signatures of spermatogenic impairment and severe developmental disease noted above, there are connections in their epidemiology. Recent results estimate a 9.9% rate of birth defects in children conceived by intracytoplasmic sperm injection (ICSI), the technology typically employed for assisting cases of severe male factor infertility, which is an OR of 1.77 compared to unassisted reproduction [59]. Among several possible explanations for this finding, our data raise the possibility that mutations that compromise gonadal function may act pleiotropically to disrupt development in other tissues. A better understanding of the genetic basis of male infertility is urgently needed in order to improve risk assessment for couples considering assisted reproduction.
Clinical genomics is a paradigm in need of robust applications, and our finding of a large CNV burden in cases suggest that some infertility mutations may have the high penetrance required for clinical utility. Indeed some mutation screens are already used clinically in the management of male infertility. Although the presence of azoospermia can be easily assessed using a standard laboratory test, many men with azoospermia will have sperm production within the testis and be candidates for testicular sperm retrieval. We have already identified that the specific AZF deletion (a, b or b/c) has a dramatic effect on the prognosis of sperm retrieval (vs. AZFc-deleted males) [60]. In the present study, we have identified deletion of DMRT1 coding sequence as a genetic event that appears highly predictive of spermatogenic failure. In depth characterization of carriers is now needed to understand how this mutation affects the prognosis of sperm retrieval. Similar whole genome tests may provide critical prognostic information that can help to characterize the chance of successful treatment for couples with non-obstructive azoospermia, avoiding expensive and needlessly invasive interventions, while potentially providing guidance for new therapeutic interventions.
All DNA samples used in this study were derived from peripheral blood lymphocytes collected from individuals giving IRB-approved informed consent. The following IRBs were involved: INSA Ethics Committee and Hospital Authority (Portugal), University of Utah IRB, and Washington University in St. Louis IRB (#201107177). All samples of genomic DNA to be analysed in this study i) belong to DNA banks that have been established throughout the years; ii) are coded; and iii) each individual has signed a declaration of informed consent before donating his genomic DNA for analysis, authorizing molecular studies to be performed with this material.
All cases were deemed idiopathic following a standard clinical workup, which included screening for Y chromosome deletions. Controls from the Utah cohort were men with normal semen analysis, remaining controls were not phenotyped on semen quality. Full details of the source and diagnosis of samples in this study are available in Supplemental Methods. When using SNP arrays, CNV analysis is more sensitive to experimental noise than SNP genotyping, and we used different sample QC metrics to inform CNV and SNP stages of our project. As a result, we have slightly larger sample sizes for the HBD analyses than for the CNV analyses.
The individuals studied here were sourced from diverse geographic locations (Table 1, Text S1). All primary samples (e.g. 323 cases and 1133 control samples subjected to whole-genome genetic analysis) were of self-reported Caucasian ancestry, but it was necessary to take additional steps to control for population structure in all aspects of the analysis. First, genetic ancestry of each sample was assessed by principal components analysis and ethnicity outliers were removed (Figure S2, Figure S3). Second, eigenvectors generated by this principal components analysis were used as covariates in both CNV association and inbreeding coefficient association analyses. For analyses focusing on the Y chromosome, we performed analyses conditioning on Y haplogroup to provide the most stringent possible correction for population structure with available data. Lastly, we conducted alternate association analyses with the Porto case cohort using a smaller, but more geographically proximal Spanish control cohort (Figure S5).
Three array platforms were used for CNV discovery: Illumina 370K (Utah), Illumina OmniExpress (Washington University), and Affymetrix 6.0 (Porto, Cornell, Nanjing). Full details of sample processing and array experiments are available in Supplemental Methods. Three CNV calling algorithms were used to generate CNV maps for each individual typed with Illumina technology: GADA, a sparse Bayesian learning approach [61]; PennCNV, a Hidden Markov Model (HMM)-based method originally designed for the Illumina platform [62]; and QuantiSNP 2.0, another HMM-based method for Illumina [63]. CNVs called by 2 of 3 algorithms were retained for analysis. CNV calling for Affymetrix 6.0 was performed with Birdsuite [64]. Due to the complexity of calling CNVs on the sex chromosomes, for all array datasets we implemented a bespoke normalization and calling procedure that used only the GADA algorithm to call CNVs from the X and Y chromosomes. For full details of CNV calling see Supplemental Methods.
Regions of homozygosity-by-descent (HBD) were identified using BEAGLE 3.0 [48]. SNPs with no-call rates >5% were removed prior to HBD analysis. As BEAGLE uses a model for background linkage disequilibrium that is fit from the data, cases and controls from each cohort were analyzed simultaneously and separately to assess cohort-specific biases in calling HBD. Prior to downstream analysis, we identified and removed a small number of reported HBD regions that corresponded to rare, large hemizygous deletions.
Inbreeding coefficients for each individual were calculated from their HBD data using the formula:
Due to differences in array content, CNV frequencies were determined on a per-platform basis. All CNV calls made on a given platform, in both cases and controls, were combined into CNV regions using a threshold of 50% reciprocal overlap to defined two events as the same ([35]). We defined the CNV frequency as the proportion of all samples (cases and controls) containing that CNV.
We constructed several statistical tests to measure differences between cases and controls. We used Mann-Whitney U tests to test for differences in the total amount of aneuploid sequence per genome. We used standard logistic regression to test for CNV load on chromosome compartments (e.g. the autosomes, X chromosome) and a small number of functional features (genes, miRNA, etc). To control for population structure these models included the first 10 principal components from PCA analysis of the SNP genotype data from all cohorts (Figure S2). We used a permutation strategy for genomewide, locus-by-locus testing for association at all genes and in 500 kb non-overlapping genomic windows. The permutation strategy, implemented with the software package PLINK, calculates nominal and genomewide p-values by permuting case-control labels [65]. To present consistent summaries of CNV burden for the entire study (all cohorts combined), we used linear mixed-effects logistic regression, treating cohort as a random factor and compared these to effect size estimates for each cohort separately using standard logistic regression (Figure 1). The mixed effects modeling framework controls for SNP platform as each case-control cohort was typed on a different platform; a similar use of mixed-effect modeling was recently described in a meta-analysis of schizophrenia SNP data [27].
Analogous tests were conducted on HBD segments from the original discovery cohort and the combined primary and replication datasets.
We performed validation and replication analyses of DMRT1 deletions with and assay based on Taqman PCR. Copy number was assessed using a pre-designed assay #Hs06833797_cn within the DMRT1 gene against an RNase P reference (assay # 4403326; both assays from Applied Biosystems, Carlsbad, CA, USA) according to manufacturer's recommendations.
|
10.1371/journal.pcbi.1004018 | Protein-Protein Docking with Dynamic Residue Protonation States | Protein-protein interactions depend on a host of environmental factors. Local pH conditions influence the interactions through the protonation states of the ionizable residues that can change upon binding. In this work, we present a pH-sensitive docking approach, pHDock, that can sample side-chain protonation states of five ionizable residues (Asp, Glu, His, Tyr, Lys) on-the-fly during the docking simulation. pHDock produces successful local docking funnels in approximately half (79/161) the protein complexes, including 19 cases where standard RosettaDock fails. pHDock also performs better than the two control cases comprising docking at pH 7.0 or using fixed, predetermined protonation states. On average, the top-ranked pHDock structures have lower interface RMSDs and recover more native interface residue-residue contacts and hydrogen bonds compared to RosettaDock. Addition of backbone flexibility using a computationally-generated conformational ensemble further improves native contact and hydrogen bond recovery in the top-ranked structures. Although pHDock is designed to improve docking, it also successfully predicts a large pH-dependent binding affinity change in the Fc–FcRn complex, suggesting that it can be exploited to improve affinity predictions. The approaches in the study contribute to the goal of structural simulations of whole-cell protein-protein interactions including all the environmental factors, and they can be further expanded for pH-sensitive protein design.
| Protein-protein interactions are fundamental for biological function and are strongly influenced by their local environment. Cellular pH is tightly controlled and is one of the critical environmental factors that regulates protein-protein interactions. Three-dimensional structures of the protein complexes can help us understand the mechanism of the interactions. Since experimental determination of the structures of protein-protein complexes is expensive and time-consuming, computational docking algorithms are helpful to predict the structures. However, none of the current protein-protein docking algorithms account for the critical environmental pH effects. So we developed a pH-sensitive docking algorithm that can dynamically pick the favorable protonation states of the ionizable amino-acid residues. Compared to our previous standard docking algorithm, the new algorithm improves docking accuracy and generates higher-quality predictions over a large dataset of protein-protein complexes. We also use a case study to demonstrate efficacy of the algorithm in predicting a large pH-dependent binding affinity change that cannot be captured by the other methods that neglect pH effects. In principle, the approaches in the study can be used for rational design of pH-dependent protein inhibitors or industrial enzymes that are active over a wide range of pH values.
| Through tightly controlled cellular pH, posttranslational modification by protons regulates biological function [1]. Cellular pH can vary from highly-acidic in the lysosomes (∼pH 5) to basic in the peroxisomes (∼pH 8) [2], profoundly influencing biomolecular folding and assembly processes [3], [4]. pH effects are especially critical in protein-protein binding, and binding-induced protonation state changes contribute to the association energy of most protein-protein complexes [5], [6]. However, computational protein-protein docking algorithms often ignore the pH effects. In this paper, we develop a pH-sensitive protein-protein docking algorithm and demonstrate that it can improve prediction accuracy and recover pH-dependent binding effects.
Computational docking algorithms are playing an increasingly influential role in driving large-scale protein-protein interactions (PPI) surveys [7], [8] and genome-wide interactome studies [9], but they need to accommodate sensitivity to local environment pH for improved reliability. Although pH effects on protein-small molecule complex calculations are well studied (e.g., refs. [10]–[15]), efforts to incorporate pH effects in computational protein-protein complex calculations have just begun. For example, Spassov et al. [16] recently demonstrated a pH-sensitive binding prediction method with an aim to prolong the half-life of therapeutic antibodies. HADDOCK [17] determines the missing protonation state of the histidine residues in the input protein complex using the WHATIF server [18] before the start of the docking simulation. However, in real systems protonation states are affected not only by the solution pH but also the change in the local environment of the ionizable surface residues due to the receptor-ligand interactions during binding. pKa calculation studies (e.g. [19]) stress the importance of simultaneously evaluating both favorable residue side-chain conformations and their preferred ionization states. Similarly, in docking algorithms, residue pKa values vary depending on the conformations of the ligand relative to the receptor. Hence dynamic evaluation of the protonation states during docking using pKa calculation algorithms on-the-fly is more true to the physical process of binding and may improve prediction accuracy.
Current computational pKa calculation algorithms have been collectively assessed by the scientific community recently to improve their accuracy [20]. One of the primary aims of the pKa calculation methods is to identify and improve the deficiencies of the energy function, particularly the electrostatics [21]. Despite the deficiencies, pKa calculations by many algorithms are within a root-mean-square deviation (RMSD) of 1 pH unit from the experimental pKa values (except in extreme cases with very large pKa shifts [22]–[24]). Hence unless the solution pH is very close to the shifted pKa values of the ionizable residues, current algorithms can in principle reasonably estimate the relevant pH-sensitive protonation state during docking. Since computational protein-protein docking algorithms typically generate hundreds to several thousand target conformations, effective use of the protonation state data requires pKa calculations to be fast, accurate and compatible with the docking methodology. Unfortunately, the most rigorous physics-based pKa calculation methods prohibitively require several minutes to hours to calculate a single pKa value, and the faster empirical methods are not currently compatible with the docking frameworks.
We previously created Rosetta-pH [25], a fast and efficient pKa calculation algorithm with a focus on the use of the protonation state data in protein structure prediction and design. After we added a pH-sensitive score term to the standard (pH-independent) Rosetta score function and calibrated the electrostatic and solvation score terms, Rosetta-pH achieved a RMSD of 0.83 pH units from the experimental pKa values. Since we built Rosetta-pH using the object-oriented Rosetta biomolecular modeling suite [26] which forms the basis for the protein-protein docking algorithm RosettaDock [27], [28], we were able to fuse the methods to create, to our knowledge, the first pH-sensitive protein-protein docking algorithm.
In the remainder of this article, we describe our fast pH-sensitive docking algorithm (pHDock) that can sample side-chain protonation states of five ionizable residue types (Asp, Glu, His, Tyr, Lys) on-the-fly during the docking simulation. After combining the Rosetta-pH and RosettaDock frameworks, we recalibrate the pHDock score function to accommodate the new pH-sensitive score term. We use local docking studies to test pHDock's performance on a dataset of protein-protein complexes [29] and compare it to RosettaDock. We also study the effects of incorporating backbone flexibility in pHDock using a backbone conformational ensemble for docking a subset of the complexes. Finally, we explore a case study to investigate the efficacy of pHDock in the prediction of large pH-dependent binding affinity change in a protein complex [30].
We developed pHDock, a multi-scale Monte Carlo (MC) algorithm based on the RosettaDock framework [27], [28] with modifications to allow dynamic sampling of the residue protonation states during simulation. Residue protonation states at the environment pH are constantly updated during multiple side-chain packing steps throughout the protocol by explicitly sampling both protonated and deprotonated versions of the side chains from a discrete rotamer library [31].
The pHDock algorithm is illustrated in Fig. 1. In the first pre-packing step, the protein complex side chains are idealized, and the residue ionization states are equilibrated with the solution pH using Rosetta-pH [25]. Then, following the standard RosettaDock low-resolution stage, the residue side chains are represented by coarse-grained centroid atoms. This stage comprises i) a random initial perturbation of the partners, and ii) rigid-body ligand moves relative to the receptor which are accepted/rejected based on the Metropolis criteria. In the high-resolution stage, the side-chain centroid pseudo-atoms are replaced by the side-chain atoms from the initial unbound conformation. The high-resolution stage involves i) repacking the residue side chains with simultaneous evaluation of the most favorable residue protonation states at the environmental pH, and ii) minimization of the side-chain torsion angles and rigid-body orientation of the ligand relative to the receptor with an accompanying Metropolis criteria check. One thousand candidate structures, or models, are generated for each target and then ranked according to their interface scores, and the top-ranked model is picked as the final prediction.
To test the performance of the algorithm, we use both standard RosettaDock (henceforth referred to as simply ‘RosettaDock’) and pHDock to generate local docked models starting from a dataset of unbound structures from the curated Docking Benchmark 4.0 [29]. For pHDock, we assume the crystallization pH of the corresponding bound complex as the solution pH. In the following sections, we first illustrate the docking performance analysis of the new algorithm using a sample protein complex. Next, we compare the performance of pHDock to RosettaDock over the complete benchmark dataset using several metrics and inspect a few predictions in greater detail. We later focus on the effects of backbone flexibility on the docking accuracy. Finally we use a case study to demonstrate pHDock's performance in the prediction of pH effects on binding affinities.
Performance of structural docking algorithms can be analyzed by studying the distribution plots of the free energies or score function vs. the deviation from the starting native bound complex. The native complex is assumed to be at the free energy minimum, hence structural models generated using the docking algorithm with receptor-ligand orientation close to the native structure are expected to have lower energies compared to the structures farther away. To create a set of models sampling both near-native and non-native conformations, starting positions of the ligand relative to the receptor are perturbed by up to 3 Å translation and 8° rotation around the axis joining the centers of the two partners.
Fig. 2 shows sample plots for the Triticum aestivum xylanase inhibitor-I (TAXI-I) in complex with Bacillus subtilis xylanase crystallized at a pH of 4.6 (PDB: 2B42 [32]). The y-axis represents the interface score (Isc), an approximation of the binding free energy, normalized by the difference between the 5th and 95th percentile scores. The x-axis quantifies deviation from the native complex using interface RMSD (Irmsd). Each point on the plot represents a single docking model and is colored based on the CAPRI structural quality rating [33] (see Methods). The interface of the top-scoring pHDock-generated structure (Fig. 2B) is just 1.7 Å from the native interface, compared to 4.7 Å for the RosettaDock-generated structure (Fig. 2A). While RosettaDock does not generate any structures better than acceptable quality, pHDock produces a structure with higher native residue-residue contact recovery qualifying as medium quality.
We quantified the docking performance using a discrimination score [34] (shown in bottom right in the docking score plot), which captures the extent to which the low-rmsd models have lower energies compared to the high-rmsd (incorrect) models. The discrimination score is calculated by dividing the x-axis using multiple Irmsd cut-offs and averaging the energy gaps between the lowest scoring structure on the left and right of each cut-off (see Methods). A lower discrimination score is an indicator of better docking performance, with a negative score indicating a successful docking prediction. The additional side-chain protonation state sampling helps pHDock produce a successful and more pronounced docking funnel (discrimination score: −1.19) compared to RosettaDock (discrimination score: 0.16).
Fig. 2C compares the interfaces of the crystal structure and the top-ranked pHDock model for the xylanase–TAXI-IA complex. Experimental studies [32], [35] discussed the importance of the strong salt bridge between the positively charged imidazole side chain of TAXI-IA His-374 (spheres) with the negatively charged Asp-37. This ionic interaction is critical for binding, and the pH optimum of the xylanase (determined by the pKa value of Asp-37) is reported to directly influence the affinity of the enzyme–inhibitor complex, with a lower Asp pKa value leading to stronger binding. The top-scoring pHDock model not only captures this interaction through precise prediction of the positively charged His-374 side-chain rotamer but also recovers all the xylanase active-site-residue side-chain rotamers. RosettaDock, which assumes a neutral His side chain, fails to capture the interaction. Overall, while the top-scoring RosettaDock model recovers just 13% of the native interface contacts, the pHDock model recovers 49% of all the interface residue-residue contacts.
For a large-scale docking performance analysis, we tested pHDock over a dataset of diverse protein–protein complexes from the curated Docking Benchmark 4.0 [29]. On average, 25% of the interface residues in the dataset complexes are ionizable (Asp, Glu, His, Tyr, Lys) (S1 Figure). Fig. 3 compares the discrimination scores of the docking funnels generated using pHDock and RosettaDock. pHDock produces successful docking funnels (discrimination score ≤0) in approximately half (79/161) the structures from the dataset, including 19 cases where RosettaDock fails to produce a successful prediction. Based on the discrimination score, pHDock outperforms RosettaDock in approximately 60% of the targets (94/161) (Table S2), and the improvements are statistically significant (paired t-test, p = 0.039). Additionally, since models are generated stochastically, we performed bootstrap case resampling [36] to quantify the variation of the discrimination scores. The bootstrap mean discrimination scores µ(D) (S2 Figure) again show that pHDock produces successful funnels [µ(D) ≤0] in half the targets (79/161) including 17 cases where RosettaDock fails. Hence the results are robust to the stochastic sampling noise. The average standard deviation of the discrimination scores [σ(D): 0.07] is approximately 4% of the total observed µ (D) range.
As pHDock has access to nonstandard residue protonation states unlike RosettaDock, we examined the prevalence of such protonation states and their effect on docking accuracy. In docking funnel plots in S9 Figure, structures with nonstandard residue protonation states are distinguished. pHDock produces models with nonstandard protonation states for all the target complexes (S3 Figure), with a majority of the nonstandard protonation states observed in complexes with docking pH within one pH unit of the residue intrinsic pKa values (S4 Figure). Overall, pHDock outperforms RosettaDock in 67% (20/30) of the cases where the top-ranked pHDock model recovers a nonstandard protonation state observed in the native bound complex (S5 Figure). pHDock also performs better than RosettaDock in 64% (7/11) of the cases where the top-ranked pHDock produces a nonstandard protonation state different from the one observed in the native bound complex illustrating the importance of dynamic protonation states.
Since pHDock is a stochastic docking algorithm that generates several candidate models, the performance of the algorithm broadly depends on (i) the quality and diversity of the generated ensemble of models, or ‘sampling’, and (ii) the ability of the final score function to discriminate native-like models from non-native-like models, or ‘scoring’. To test the sampling performance of pHDock, we examined the lowest-Irmsd models for all the complexes in the dataset. The Irmsd distribution for pHDock is similar to RosettaDock (Fig. 4A), and in 92% of the docking targets, it generates at least one model within 4 Å from the native interface. Out of 1000 models generated for each target, pHDock creates on average 1.9, 18.5, and 90.8 high-, medium-, and acceptable-quality models, respectively. In comparison, RosettaDock samples 7–12% fewer medium- and high-quality models (S6 Figure). To test the scoring performance of pHDock, we calculated the Irmsd and fnat distributions of the top-scoring models for each target (Figs. 4B–C). pHDock generates top-ranked models within 4 Å in 57% of the targets (RosettaDock 51%), and 52% of the time these models recover more than 30% of the native residue-residue contacts (RosettaDock 46%).
To further assess the quality of the predicted top-ranked structures, we examined the receptor-ligand interface hydrogen bonds (henceforth referred to as simply ‘interface hydrogen bonds’). Previous surveys found 8–13 interface hydrogen bonds in each protein–protein complex [37], [38]. Using Rosetta's hydrogen bonding definition, the native crystal complexes in our dataset contain 6.4±3.5 interface hydrogen bonds on average (Fig. 5A). In comparison, the top pHDock models are involved in 5.1±2.5 interface hydrogen bonds, while the top RosettaDock models form only 3.4±2.1 interface hydrogen bonds. As pHDock primarily focuses on ionizable residues, we also calculated the number of interface hydrogen bonds containing such residues as donors or acceptors. The native complexes contain 3.5±2.6 ionizable interface hydrogen bonds (Fig. 5B). Encouragingly, the top pHDock models are found to form an identical 3.5±2.4 ionizable interface hydrogen bonds, while the top RosettaDock models form only 2.1±1.6 hydrogen bonds.
The analysis of the total number of interface hydrogen bonds shows significant pHDock improvements in generating models with a larger receptor-ligand hydrogen bond network. However, such an analysis does not reveal the accuracy of the generated interface hydrogen bonds. So we also examined the fraction of the native interface hydrogen bonds recovered in the top-ranked models. pHDock recovers more than one-fifth of the native interface hydrogen bonds in only 33% of the targets from the dataset, while RosettaDock performs worse, recovering the same fraction in just 22% of the targets (Fig. 5C). The results are similar for the fraction of recovered ionizable interface hydrogen bonds. pHDock recovers more than one-fifth of the ionizable interface hydrogen bonds in 32% of the targets, while the performance of RosettaDock drops further to just 19% of the total targets in the dataset (Fig. 5D). In summary, while pHDock generates more interface hydrogen bonds, only a minor faction of these hydrogen bonds match those seen in the native complex.
Finally, to test the effects of hydrogen bonding accuracy on docking results, we examined a few sample cases in greater detail. The tumor susceptibility gene 101 protein–ubiquitin complex (1S1Q; pH 4.6 [39]) has four native interface hydrogen bonds. The top pHDock model recovers three of them and forms a total five interface hydrogen bonds, while the top RosettaDock model exhibits three interface hydrogen bonds but none of them are native. The docking plots for both pHDock and RosettaDock (discrimination score −0.19 vs. −0.01) show success based on discrimination scores, but the docking funnel is clearly more pronounced in pHDock (Fig. 6A). Although the near-native sampling in both pHDock and RosettaDock is comparable, the additional recovered native hydrogen bonds help pHDock in the final scoring, and the top model interface is only 1.4 Å away from the native interface. The improved performance is likely due to a protonated interface histidine (His-66) in ubiquitin. In a second case, the PPARgamma+RXRalpha–GW409544+co-activator peptide complex (1K74; pH 7.5 [40]) has five interface hydrogen bonds. The top pHDock model exhibits eight interface hydrogen bonds, three of them being native, while none of the ten hydrogen bonds found in the top RosettaDock model are native (Fig. 6B). In this case, pHDock (discrimination score −0.35) outperforms RosettaDock (discrimination score −0.12) in both sampling and scoring (Fig. 6B). The top-scoring pHDock model is a high-quality prediction just 0.93 Å from the native interface. In this case, the interface residues are all in their standard protonation states; we infer that the improvement must be due to kinetic effects during the Monte Carlo docking search. The larger number of interface hydrogen bonds in pHDock models do not always translate to improvements in docking predictions. For example, the CDK2 kinase–cell cycle-regulatory protein CksHs1 complex (1BUH; pH 7.5 [41]) has four native hydrogen bonds. Again, the interface residues in the top-ranked pHDock model are predicted to be in their standard protonation states. Neither top pHDock nor RosettaDock models recover any of the native interface hydrogen bonds although they form nine and one interface hydrogen bonds, respectively. As shown in the docking plots in Fig. 6C, pHDock scoring favors a false-positive docking prediction with a large number of interface hydrogen bonds more than 12 Å from the native interface.
Inclusion of backbone flexibility in protein-protein docking is critical to capture the conformational changes during the binding event [42]. Within RosettaDock, backbone flexibility mimicking both conformer selection (CS) and induced fit (IF) binding models increases native contact recovery, although the computational costs are higher and there is a risk of false positive predictions [43]. Thus we tested whether the addition of backbone flexibility further improved native contact recovery in pHDock. We chose a subset of 14 complexes common among the published study and the curated Docking Benchmark 4.0 used for pHDock (Table S3). We then used the RosettaRelax [44], [45] protocol to generate an ensemble of unbound backbones. RosettaRelax, an MC algorithm, employs a cycle of small backbone dihedral (φ, ψ) perturbations, residue side-chain packing and score function minimization along the gradient in the torsion space to generate a backbone ensemble typically within 1 Å Cα RMSD of the starting structure. We generated 500 models starting from the ligand unbound coordinates for each of the complexes and picked the ten top-scoring models for docking.
S10 Figure compares the docking funnels generated using RosettaDock, pHDock and ensemble pHDock. The ligand backbone flexibility helps ensemble pHDock generate better docking funnels (based on discrimination score) in 11 targets compared to pHDock. The Irmsd values of the lowest-Irmsd models generated using ensemble pHDock are not significantly better compared to pHDock. However, there is a noticeable improvement in the quality of the receptor-ligand interfaces in the top-ranked models. The top-ranked models generated using ensemble pHDock outperform pHDock in native contact recovery with comparable or better fnat values in 12 targets. Encouragingly, the top-ranked models also recover comparable or more native interface hydrogen bonds in all the targets compared to pHDock and RosettaDock (Table S3). To summarize, the additional backbone flexibility further improves the docking funnel quality in a majority of the targets and generates top-ranked models that recover more native contacts and hydrogen bonds.
pHDock simulates the complexes at solution pH and relies on dynamic residue protonation state sampling. To assess the individual contribution of these two components, we performed control docking experiments using a subset of complexes (same 14 complexes used for ensemble pHDock). First, to test the robustness of the docking predictions to changes in the solution pH, we used pHDock at physiological pH (pH 7.0). Second, to test the benefits of employing dynamic residue protonation states, we docked the complexes with fixed residue protonation states obtained from the lowest energy rotamer state of the starting partners at the solution pH (fix-pHDock).
Of the cases where both RosettaDock and pHDock either fail (four targets) or succeed (eight targets), the fix-pHDock and pHDock at pH 7.0 runs perform similarly (see docking funnel plots, S11 Figure), showing, as might be expected, an insensitivity to pH effects. There are two cases in this test set where RosettaDock fails and pHDock produces a successful docking funnel. In the α-chymotrypsin–eglin C complex (1ACB; pH 6.5 [46]), pHDock produces a discrimination score of −0.24 at pH 6.5, and RosettaDock a discrimination score of 0.01. pHDock at pH 7.0 produces a weaker funnel (discrimination score: −0.1) while fix-pHDock fails (discrimination score: 0.09) due to a false positive model 7 Å Irmsd away from the native complex. Similarly, in the Fab D44.1–lysozyme complex (1MLC; pH 6.0 [47]), pHDock generates a discrimination score of −0.11 while RosettaDock, pHDock at pH 7.0, and fix-pHDock all fail (discrimination scores 0.13, 0.07, 0.33, respectively). Thus, in these two cases where RosettaDock fails, both pHDock at pH 7.0 and fix-pHDock fail to completely capture pHDock's success. These cases suggest that accurate knowledge of the solution pH and the dynamic protonation states are vital for maximum pHDock accuracy.
In the discussion so far, we analyzed pHDock's performance at the solution pH and compared it to RosettaDock (no pH dependence) over a large dataset of protein complexes. However, such an analysis does not test pHDock's performance in predicting effects of subtle environmental pH changes on a single protein-protein complex. In previous work, we and other groups have used RosettaDock interface scores in correlating binding affinities [48], [49] and in predicting relative affinities [50]. The neonatal Fc receptor (FcRn) binds maternal immunoglobulin G (IgG) from ingested milk in the gut at acidic pH (pH≤6.5) and releases it in the bloodstream of the newborn at basic pH (pH 7.4) [51]. This process is facilitated through a drastic drop in the binding affinity by more than two orders of magnitude as the pH changes from 6.0–6.5 to 7.0–7.5 [51], [52]. The Fc–FcRn system has been previously used for a pH-dependent binding calculation [16], but to our knowledge, there are no existing pH-sensitive docking studies.
To test the efficacy of pHDock in predicting pH effects on binding affinities, we used the pHDock algorithm to dock the murine Fc–FcRn complex (1I1A [30]) at various environmental pH values. We tested all integral pH values between 3.0 and 11.0, and used a finer interval of 0.25 pH units for the relevant pH range of 6.0–8.0 where the striking binding affinity change is observed. We used the interface scores (I) of the top-scoring pHDock models to approximate the binding affinity at different pH values. Fc–FcRn complex shows a binding minimum at pH 6.25 (IpH6.25: −13.99 Rosetta Energy Units (REU)), and thereafter the affinity rapidly weakens as the environment pH increases to 7.50 (IpH7.50: −11.82 REU) (Fig. 7A). Converting the binding energies to equilibrium constants using the relation we estimated the ratio of equilibrium constants at pH values 6.25 and 7.50 as where KpH6.25 and KpH7.50 are the equilibrium binding constants at pH 6.25 and 7.50, respectively, and kBT is 0.59 kcal/mol at 298K. The equation yields a 40-fold drop in the binding affinity as the pH increases from 6.25 to 7.50, which is similar to the 50 to 100-fold drop from experiments [52]. Interestingly, the docking plots show successful energy funnels for both pH values (Fig. 7B). However, the energy funnel is more pronounced at pH 6.25 (discrimination score −0.96) than pH 7.50 (discrimination score −0.47), indicating a site-specific binding event at both pH values, but with markedly different affinities.
Previous studies [30], [51] attribute the pH-dependence of Fc–FcRn binding to the titration of interface histidine residues with pKa values in the range of binding affinity transition (6.5≤pH≤7.0). The Fc–FcRn interface has three salt bridges with the residues His-310, His-435, and His-436 from Fc interacting with Glu-117, Glu-132, and Asp-137 from FcRn. The proposed mechanism involves titration of all the three histidine residues disrupting the binding as the environment pH increases, but studies have shown two buried titratable salt bridges are sufficient to confer pH dependence. Encouragingly, the top-scoring pHDock-generated models at different pH values successfully capture the titration event. While His-310 remains protonated in the models at both pH values, His-435 and His-436 are protonated at pH 6.25 and deprotonated at pH 7.50 and are involved in salt bridges with Glu-132 and Asp-137, respectively (Fig. 7C). Thus, pHDock not only predicts the relative Fc–FcRn binding affinities at different pH values, but also captures the expected physical mechanisms responsible for the different affinities.
We have created pHDock, the first pH-sensitive protein-protein docking algorithm that samples residue protonation states dynamically during the search. The algorithm integrates the Rosetta-pH pKa calculation method [25] with the RosettaDock framework using the object-oriented design of the Rosetta modeling suite [26]. Local docking studies show that pHDock outperforms RosettaDock in 60% of the docking targets and also performs better than control cases involving docking at pH 7.0 or using fixed, predetermined protonation states. pHDock also shows encouraging improvements in the quality of the generated candidate predictions. On average, the top-ranked pHDock structures have lower interface RMSDs and recover more native residue-residue contacts and hydrogen bonds. While pHDock is designed to improve docking predictions by accounting for environmental pH effects, the successful prediction of a large pH-dependent binding affinity change in the Fc–FcRn complex suggests that it can be further exploited to improve affinity predictions.
pHDock improves docking primarily by enhancing the scoring in the docking high-resolution stage, as the improved score function finely tuned for pKa predictions is active only during the high-resolution steps involving dynamic protonation states. Although there are few cases where pHDock samples conformations closer to the native compared to RosettaDock, the similarity of the interface RMSD distributions of the closest-sampled models (to the native complex) shows that its sampling quality is largely unchanged, likely because it retains the RosettaDock low-resolution stage which is largely responsible for model diversity. Over the complete dataset, pHDock generates at least one high-quality model in 25% of the complexes (41 targets), slightly higher than RosettaDock (34 targets). ReplicaDock [53], which uses a set of temperature replicas, overcomes the kinetic barriers and improves sampling in the low-resolution docking stage. Further work can thus focus on combining the principles of ReplicaDock with pHDock to improve the model diversity in the low-resolution centroid phase. Also, availability of even sparse biochemical information [54] can be used as an alternative to constrain the conformational search space and circumvent the sampling concerns in the centroid phase to improve docking accuracy.
Although the top-ranked pHDock models show significant advancements in recovering native contacts, the hydrogen bonding performance is mixed. The geometry of interface hydrogen bonds is less optimal than intra-chain hydrogen bonds, but they are nevertheless critical for protein-protein binding [37]. The top pHDock models exhibit more hydrogen bonds than RosettaDock on average. The increase is especially evident in the case of ionizable residues where the pHDock hydrogen bond distribution matches the native distribution. However, many of the pHDock interface hydrogen bonds are non-native, i.e., they are not observed in the bound crystal complexes. In fact, in two-thirds of the targets, pHDock fails to recover more than one-fifth of the native interface hydrogen bonds, a shocking number revealing the limitations still present in the hydrogen bonding model.
There are a few possible explanations for the poor hydrogen bond performance. First, pHDock uses an implicit solvation model and thus fails to capture the water-mediated interface hydrogen bonds. Although the water-mediated hydrogen bonds are excluded from native hydrogen bond calculations, ignoring the water molecules during docking can result in the compensation of unsatisfied hydrogen bond donors/acceptors through formation of non-native hydrogen bonds. Second, pHDock ignores protein backbone flexibility and uses the unbound coordinates of the protein partners for docking, hence any resulting backbone inaccuracies can shift the hydrogen bond network. Accounting for backbone flexibility using a conformational ensemble for a small subset of complexes improves hydrogen bond recovery compared to pHDock, but the top-ranked models still recover just a quarter of the native interface hydrogen bonds. Further studies to improve hydrogen bond recovery can focus on calibrating the score function using the bound coordinates of the complex to minimize the errors introduced due to the rigid backbone assumption and the inaccuracies in the receptor-ligand orientation in the docking models. However, work will be needed to reconcile the changes with the docking score function that is tuned for recovering native-like structures.
We tested pHDock's ability to capture the large pH-dependent binding affinity change in the Fc–FcRn complex. Since the binding changes are a result of protonation state shifts in the interface histidine residues, any docking algorithm ignoring environment pH will fail to capture the effect. pHDock predicts a 40-fold drop in the binding affinity due to the increase in the environment pH, and the top-scoring model captures the resulting disrupted salt bridges at the Fc–FcRn complex interface. The accuracy of the affinity prediction suggests that pHDock can be expanded to power computational protein design studies such as those that recently began to exploit the pH-dependence for regulating protein binding activity [55]. Previously during the CAPRI rounds 20–27 [56], we used pHDock for the blind prediction of the g-type lysozyme–PliG inhibitor complex [50]. Lysozyme operates in a low pH environment [57] and hence provided an opportunity to test pHDock's performance. Docking the complex at pH 6.2 (crystallization pH of the unbound lysozyme) generated a medium-quality prediction just 2.0 Å from the interface of the native complex. The encouraging performance of pHDock proves that it can be effective in capturing environment-pH effects on both docking and binding.
Recent efforts have begun to capture structural details of protein interactions in complete cellular environments [58]–[60]. There is tremendous scope for computational docking algorithms to power such studies, but the methods must be versatile and include the effects of environmental conditions. Since intracellular pH is strictly regulated across multiple eukaryotic cellular compartments and is critical for protein interactions [61], accounting for pH effects can boost prediction accuracy. The results in this paper contribute to the community effort to simulate protein-protein interactions in the complete cell with all environmental factors.
The Protein-Protein Docking Benchmark 4.0 by Hwang et al. [29] is a set of 176 non-redundant protein-protein complexes with both bound and corresponding unbound crystal coordinates from the Protein Data Bank [62]. The dataset comprises 121 ‘rigid-body’, 30 ‘medium’, and 25 ‘difficult’ targets based on the interface backbone conformation variation between bound and unbound coordinates [63].
We curated the benchmark dataset in multiple stages. First, we removed water and all non-peptide molecules containing heteroatoms from the complex structures. Since Rosetta pH does not currently predict protonation states of non-peptide molecules, we excluded complexes with such molecules at the interface. We also eliminated structures in which Rosetta was unable to resolve the steric clashes in the starting atomic coordinates due to the conformational changes between bound and unbound complexes, leaving 161 test complexes for the study. Second, we truncated both the unbound and bound structures to the same amino-acid sequences for Rosetta scoring consistency. Third, we collected the crystallization pH values in the PDB coordinate file for each bound complex to determine the docking environment pH. For structures missing pH information in the PDB files, we used the pH value from the corresponding original research article if available. For the remaining structures, we assumed a physiological pH of 7.0 (Table S2).
Rosetta-pH [25] is a Metropolis Monte Carlo algorithm in which the protonation state of the lowest energy conformation is evaluated using the Rosetta-pH score function at intervals of pH to estimate pKa values. The Rosetta-pH score function is based on the standard Rosetta score function with additional terms including:
i) Protonation potential based on the probability of protonation of individual amino acid residues at a given pH. The probability of protonation of an amino acid is and the protonation potential (EpH) is
where pH is defined by the environment, and IpKa is the unperturbed intrinsic pKa value of the model compound in solution (4.0 for Asp, 4.4 for Glu, 6.3 for His, 10.0 for Tyr and 10.4 for Lys). kBT is assigned a value of 0.59 kcal/mol, corresponding to T = 298K. Cys protonation state changes (intrinsic pKa 8.5) are ignored due to the complications of coupling between pKa and redox equilibrium [64].
ii) Coulomb electrostatic potential with a distance-dependent dielectric (ε = 10r) for gradual shielding at increasing interatomic distances [65], and
iii) Recalibrated solvation reference energies for the non-standard protonation variants in the Lazaridis–Karplus implicit model for solvation [66] (See [25] for details).
Rosetta pHDock uses the object-oriented design of the Rosetta biomolecular modeling suite [26] to implement the environment pH effects in the RosettaDock protocol. The pHDock development workflow can be broadly classified into three stages:
i) In the first stage, we incorporated explicit protonation state sampling from Rosetta-pH [25] into the RosettaDock algorithm. RosettaDock accounts for residue side chain flexibility in the prepacking step and the later high-resolution stage with full-atom side chains. The sampling of the side-chain χ-angles is discrete based on a backbone-dependent rotamer library [31]. Rosetta pHDock augments the sampling by allowing variable residue ionization states to be simultaneously sampled during every side-chain packing step and picking the most favorable residue protonation state based on the residue's local interactions and the solution pH. For neutral His, both possible tautomers (with proton on either Nδ1 or Nε2 atoms) are sampled. The conformational degeneracy in the protonated variants of Asp and Glu (with H atoms on either of the terminal Oδ and Oε atoms, respectively) is also explicitly incorporated by accommodating both possible protonated versions for the residues during sampling.
ii) In the second stage, we generated a dataset of structures and evaluated the contributions of the individual score terms (including e_pH) to the total interface score. We first generated 1000 models (for each complex) using the standard RosettaDock local docking routine [28] on a subset of 60 randomly-selected bound complexes (∼1/3 of the total docking benchmark set). We then repacked each model (sampling both side chains and protonation states) at the crystal pH of the bound complex and calculated the interface contribution of each score term aswhere is the contribution of the score term i in the repacked complex, and is the score term contribution in each separate binding partner j after repacking the ionizable interface residues at the crystal pH of the bound complex. Repacking the ionizable residues is required for accurate score term estimation, as separation of the binding partners exposes the previously-buried interface residues to the solvent affecting their preferential protonation state.
iii) In the third stage, we parameterized the pHDock score function. Reweighting is mandatory since the original RosettaDock score function had a minimal weight on electrostatics, and the new electrostatic weight and pH reference term must be rebalanced against the hydrogen bonding and solvation contributions. Similar to prior parameterization of the RosettaDock score function [27], we sought to maximize the free energy gap between ‘near-native’ and ‘non-native’ models. Models in the top 5% based on CAPRI rating [33] (high, medium and acceptable-quality in that order) with repulsive van der Waals scores lower than the 80th percentile are classified as near-native models. Models with the same CAPRI rating are ordered based on the fnat values (higher fnat is better). We classified the remaining models as non-native models. We then derived the score term weights using a generalized linear regression to maximize the free energy gap between the near-native and non-native model clusters. The free energy gap (ΔE) iswhere wi is the weight for score term Ei and . The score terms include an attractive van der Waals score (Eatr), a repulsive van der Waals score (Erep), an implicit solvation score (Esol) [66], a hydrogen bonding score (Ehb) [67], rotamer probability term (Edun) [31], a statistical residue pair term for ion-ion interactions (Epair) [68], a Coulomb electrostatic term (Eelec), and a term for the pH effects (EpH) [25].
Table S1 compares the optimized pHDock weights to the RosettaDock weights. The new pHDock weights for the dominant score terms Eatr, Esol, and Ehb show small deviations compared to RosettaDock (0.377, 0.225, and 0.249 versus 0.338, 0.242, and 0.245). Besides the new addition of pH-sensitive score term EpH (weight 0.21), the major changes in the score function are in the score term weights for Epair, Eelec, Edun, and Erep. The Epair term is completely absent and is balanced by the increased Eelec weight (0.319 compared to 0.026 in RosettaDock). While the Edun weight also increases (0.036 to 0.080), the Erep weight drastically drops from 0.044 to 0.005 demonstrating that the repulsive van der Waals score does not aid in docking model discrimination. The exceptionally small Erep weight however creates two issues. First, the algorithm produces structures with steric clashes during the rigid-body minimization step in the docking high-resolution stage (Fig. 1). RosettaDock [27] addresses this issue by increasing the Erep weight during minimization using a multiplier. We followed the same strategy and raised the Erep weight to match the RosettaDock weight during minimization. Second, some structures with unfavorable sterics are ranked higher during the final model discrimination. To address this, we eliminated the worst 5% percent of the pHDock structures sorted by their Erep scores. For a balanced comparison, we also omitted the worst 5% of the RosettaDock structures sorted by their interface scores.
In local docking, the input complex consists of unbound partners (orientation determined by superimposing on the coordinates of the bound complex) and the starting positions are generated by randomly perturbing the ligand relative to the receptor by up to 3 Å translation and 8° rotation around the axis joining the centers of the two partners. Both pHDock and RosettaDock use local docking to generate a diverse set of models sampling both near-native (Irmsd <4 Å) and non-native (Irmsd>4 Å) conformations around the binding site.
The CAPRI structural quality rating [33] classifies docking predictions as incorrect, acceptable-, medium-, or high-quality based on a combination of the metrics Lrmsd, Irmsd, and fnat. L_rmsd is defined as the root-mean-square deviation (RMSD) of the ligand Cα atoms after superposition of the receptor chains of the predicted and the native bound complexes. Irmsd is the Cα-atom RMSD after superposition of the interface residues (residues <4.0 Å from the binding partner) with coordinates from the bound complex. fnat is the fraction of the residue-residue contacts (<5.0 Å all-atom distance) in the native bound complex that are recovered in the predicted complex. CAPRI ratings depend on multiple criteria, but models are considered to be at least acceptable quality if they are within 4 Å from the native interface and recover at least 30% of the native contacts (fnat) [33].
A ‘docking funnel’ derives its name from the funnel-like appearance of the target score vs RMSD plots where the near-native models have better scores than non-native models. It is often used as a measure to determine the success of a docking simulation. We used two different metrics to quantify docking funnels.
i) N5: As defined by Chaudhury et al. [28], N5 is the number of models with an Irmsd of at most 4.0 Å among the five top-scoring structures based on interface score. A docking result is considered a success if N5≥3. We performed bootstrap case resampling (1000 models per target with replacement) to compare correlation between the mean µ(N5) and calculated N5, and to quantify the inherent noise within set of models using the standard deviation σ(N5) (S7 Figure).
ii) Discrimination score (D): Applying the formulation by Conway et al. [34] to docking, we first normalize the model interface scores (Î) using the 5th and 95th percentile scores as the reference by assigning them values of 0 and 1, respectively. The models are then divided into clusters based on Irmsd with cut-offs from = {1.0, 1.5, 2.0, 2.5, 3.0, 4.0, 6.0} in Ångstroms. Discrimination score (D) is defined as the normalized interface score difference of the lowest-energy model below and above each cut-off r∈, averaged over the number of cut-offs (Nr):
A docking result is considered a success if D≤0. We performed bootstrap case resampling (1000 models per target with replacement) to quantify the inherent noise within the set of models using the standard deviation σ(D) (S2 Figure).
pHDock is part of the Rosetta biomolecular modeling suite (www.rosettacommons.org) which is freely available for academic and non-profit use. The Supporting Information includes the complete list of structures from the docking benchmark dataset with the corresponding pH values and the command-line syntax for using pHDock method in Rosetta. Component methods and objects are also available in the PyRosetta libraries (www.pyrosetta.org) [69].
|
10.1371/journal.pcbi.1000316 | Intrinsic Disorder in Protein Interactions: Insights From a Comprehensive Structural Analysis | We perform a large-scale study of intrinsically disordered regions in proteins and protein complexes using a non-redundant set of hundreds of different protein complexes. In accordance with the conventional view that folding and binding are coupled, in many of our cases the disorder-to-order transition occurs upon complex formation and can be localized to binding interfaces. Moreover, analysis of disorder in protein complexes depicts a significant fraction of intrinsically disordered regions, with up to one third of all residues being disordered. We find that the disorder in homodimers, especially in symmetrical homodimers, is significantly higher than in heterodimers and offer an explanation for this interesting phenomenon. We argue that the mechanisms of regulation of binding specificity through disordered regions in complexes can be as common as for unbound monomeric proteins. The fascinating diversity of roles of disordered regions in various biological processes and protein oligomeric forms shown in our study may be a subject of future endeavors in this area.
| Traditionally, protein structure is believed to determine function. Recently, it was observed that many proteins contain regions without well-defined structure (intrinsically disordered regions), including a large fraction of eukaryotic proteins. Intrinsic disorder has been associated with particular functions including cell regulation; signaling; and protein, DNA, and ligand binding. Many proteins are intrinsically disordered in native form and fold upon binding, following the conventional paradigm. Accordingly, disorder in a protein may facilitate binding to multiple partners. However, in some cases disorder has also been found in the bound state. To gain clearer insight into the functional importance of disorder regions in protein complexes, we perform a large-scale analysis of disorder using protein structures in complex and in unbound forms. We show that disorder in protein complexes is rather common and pinpoint changes that occur upon protein binding at interaction interfaces. By illustrating a variety of functional roles for disorder in specific proteins, we emphasize the versatility and importance of this phenomenon.
| Many proteins and protein regions have been shown to be intrinsically disordered under native conditions; namely, they contain no or very little well-defined structure [1]–[6]. Intrinsically disordered proteins (IDPs) have been found in a wide scope of organisms and their disorder content was shown to increase with organism complexity [7]–[11]. Comparative analysis of the functional roles of disordered proteins suggest that they are predominantly located in the cell nucleus; are involved in transcription regulation and cell signaling; and also can be associated with the processes of cell cycle control, endocytosis, replication and biogenesis of cytoskeleton [10],[12].
IDPs have certain properties and functions that distinguish them from proteins with well-defined structures. 1) IDPs have no unique three-dimensional structure in an isolated state but can fold upon binding to their interaction partners [1], [4], [13]–[18]. 2) Conformational changes upon binding in proteins with unstructured regions are much larger than those in structured proteins [1]. 3) The conformations of disordered regions in a protein complex are determined not only by the amino acid sequences but also by the interacting partners [1],[19]. 4) IDPs can have many different functions and can bind to many different partners using the same or different interfaces [20]. 5) IDPs can accommodate larger interfaces on smaller scaffolds compared to proteins with well-defined structure [14],[21],[22]. 6) IDPs typically have an amino acid composition of low aromatic content and high net charge as well as low sequence complexity and high flexibility [2],[10],[23]. 7) Intrinsic disorder provides for a rapid degradation of unfolded proteins, thereby enabling a rapid response to changes in protein concentration (regulation through degradation) [24]. 8) Finally, intrinsic disorder offers an elegant mechanism of regulation through post-translational modifications for many cellular processes [20],[25].
Predictions of disorder in proteins take into account the characteristic features of unstructured proteins and have been shown to be rather successful, especially in the case of large regions. According to the results of CASP7 (7th Community-Wide Experiment on the Critical Assessment of Techniques for Protein Structure Prediction), the best prediction groups successfully identified 50–70% of the disordered residues with false positive rates from 3% to 16% [26]. Prediction methods aim to identify disordered regions through the analysis of amino acid sequences using mainly the physico-chemical properties of the amino acids [23], [27]–[36] or evolutionary conservation [12], [37]–[39].
As protein interactions are crucial for protein function ([40], references within), the biological role of disordered proteins should also be studied in this context. Indeed, folding of disordered proteins into ordered structures may occur upon binding to their specific partners [1], [4], [13]–[17] which may allow disordered regions to structurally accommodate multiple interaction partners with high specificity and low affinity [1], [41]–[43]. Moreover, it has been shown that the binding mechanism, whether binding occurs between folded or unfolded chains, depends on the structural characteristics, interface properties, and degree of minimal frustration of monomers [21],[44]. Binding through unfolded or partially unfolded intermediates can provide a kinetic advantage through the “fly-casting” mechanism [19]. According to this mechanism a dimensionality reduction occurs when the folding of a disordered protein is coupled with binding, thereby speeding up the search for specific targets.
A database of continuous protein fragments (Molecular Recognition Features or MORFs) has been compiled from the Protein Data Bank to include short protein chains (with fewer than 70 residues) bound to larger proteins [45],[46]. It has been argued that MORFs participate in the coupling of binding and folding, a hypothesis that was supported by the analysis of the composition and predicted disorder of MORF segments. As a result of studying the subtle structural differences of the same proteins in different conditions and functional states, many so-called “dual personality” protein segments were found able to exist in both ordered and disordered states [47]. There is a continuous range between completely structured and completely disordered proteins in which intermediate cases are rather common [24]: proteins that are disordered but compact, multi-domain proteins with disordered linkers, and ordered proteins with some local disorder.
Examples of proteins with intrinsically disordered regions which exhibit coupling between folding and binding have been described in the literature previously [1], [4], [13]–[18]. Nevertheless, the universality of this phenomenon and functional importance of many disordered regions remains unclear. The question can be expanded further to how much intrinsic disorder do protein complexes contain and what is its functional importance? To answer these questions we examine observed and predicted disorder in protein complexes and unbound proteins using a large-scale dataset of protein structures. The atomic details of structures and the conserved binding mode analysis introduced earlier [48] allow us to monitor changes happening on or near interaction interfaces and to infer their functional importance.
Figure 1 presents a flowchart of the assembly of the dataset. From the Protein Data Bank (PDB) [49] we selected X-ray structures with resolution better than 3Å. We assigned domains from the Conserved Domain Database (CDD) [50] on each protein structure chain using RPS-BLAST [51] with default parameters (E-value≤0.01). As we focus on protein-protein interactions (interactions between different protein chains) we ensured that each chain has only one CDD domain which covers at least 70% of the full chain sequence. Among overlapping domain assignments, the domain with the longest footprint was chosen where the footprint region extends from the first to the last residue in the alignment mapping a CDD family to a given chain.
Once CDD families are assigned, we identify all interacting chains within a PDB entry. Two chains qualify as interacting if they have at least 5 residue-residue contacts. A contact takes place between a residue from one chain and a residue from the other when the distance between any non-hydrogen atom of one residue is within 6 Å of any non-hydrogen atom of the other residue. The set of residues which make contacts between the chains form the interface. To ensure that interactions are biological and not spurious, such as from crystal packing, we remove interactions that are not confirmed with additional instances of the same family pair interacting in the same orientation, so-called Conserved Binding Modes (CBM) [48]. These CBMs are defined using structural alignments between different structural instances of the same interacting family pair to confirm overlap of at least 50% of interface residue positions (Figure 2). Two definitions of conserved binding modes (CBMs) have been used: in one case confirmation of a binding mode can occur only between different non-redundant structures; in the other case recurrent interactions might occur within one structure. We refer to a dimer of interacting chains with a distinct CBM as a “complex” although it includes only pairwise interactions and several such “complexes” can be found in one PDB entry.
While analyzing disorder in dimer complexes, we also compare their disorder content with the fraction disorder of the protein in a monomeric state (Figure 1). Monomer and complex chains (as defined in PDB) corresponding to the same domain family were aligned to ensure 100% sequence identity in the non-gapped alignment. Their alignment was extended beyond the CDD footprint region as far as possible. In 95% of all cases the alignment was extended to include the entire shorter chain and in 75% of cases the alignment was extended to include both entire chains from monomer and complex structures (within 1–2 residues from both ends). The alignments are more extensive than footprint regions and cover footprint regions plus C- and N- terminal sequence regions which often do not have coordinates. Biological unit assignments were taken from the PDB asymmetric unit (ASU) assignments and from PISA predictions of multimeric states (which are based on calculation of stability of multimeric states inferred from the crystalline state) [52].
We cannot directly investigate the disorder on the interfaces in complexes as complexes are defined through residue contacts so those interface residue coordinates must be present in PDB files (see definitions of disorder below). As shown in Figure 2, disorder on the interfaces can be inferred by exploiting monomeric states of proteins, using their alignment to map the interface region from a complex onto the monomers. Given the overall numbers of disordered and non-disordered residues in the alignment, the number of residues on the mapped interface and the number of disordered residues on the interface, we can estimate the probability of observing a given number (or higher) of disordered residues on the mapped interfaces purely by chance. Using the binomial test we calculated p-values for all complexes with at least five disordered residues in the footprint or aligned regions and at least one disordered residue on the mapped interface (altogether there are 55 complexes for which interface p-values can be calculated).
After excluding those cases where interfaces are entirely outside of the alignment, our data set contained 4,884 dimer complexes and 418 unique monomer structures. Since multiple protein chains can be found in the same PDB entry (on average four chains per PDB entry from our test set) and these chains may belong to the same family, we performed an averaging of all observed quantities over the members of the family and conserved binding modes. Namely, as shown in Figure 2, disorder content observed in family type X was averaged over all instances (structures) of family X interacting with family type Y through a specific CBM. Hereafter we refer to them as “CBM interactions” or merely “interactions”. Overall, we ended up with 588 CBM interactions (“test588”). To compare disorder content in monomeric and complex states we used the more strict definitions for both binding modes and oligomerization states (see previous section). If we use the more strict CBM definitions and restrict the monomeric states by PISA (those structures which are monomeric in ASU are also predicted to be monomeric by PISA) the set is reduced to 149 interactions (“test149”). Also, for each protein used in our test set we retrieve the Gene Ontology (GO) functional annotations [53]. All structures, protein families, disorder content, GO functional annotations and other relevant information are provided in the Supporting Information.
Disordered regions were defined as those regions with missing coordinates in X-ray-resolved structures. This is the most direct way to observe intrinsically disordered regions although largely disordered proteins may be underrepresented in PDB because of the difficulties in their crystallization [5]. Disordered regions were also predicted as those with low packing density using the FoldUnfold described previously [31],[32]. Some advantages of the FoldUnfold method are that the program was not trained on the missing coordinates in PDB and that it reports a very high specificity (small number of false positives). Its performance has been shown to be comparable to other disorder prediction methods [31],[54]. (See also Table S2). According to FoldUnfold, an average packing density observed in structures was computed for each of the 20 amino acid residues. These values were considered to be the expected packing density for the same type of residues in a query protein (with or without known structure). Using a sliding window of 11 residues, the center residue of each window is predicted to be disordered if the mean packing density of the window falls below a threshold. We performed disorder predictions for all proteins in our data set.
To differentiate between ordered regions (hinge-like movements or “wobbly” domains, for example) with missing PDB coordinates and true disordered regions, we annotated those regions which are both predicted to be disordered and at the same time have missing coordinates in PDB. They will be referred hereafter as “confirmed disordered regions”. To quantify the disorder content, we calculated the “fraction disorder” as a ratio of the number of residues in disordered regions and the number of residues in the footprint or aligned regions. To see all computed values of fraction disorder consult Dataset S1 (missing coordinate definition) and Dataset S2 (confirmed disordered regions).
Analysis of fraction disorder in different families shows that one quarter of our test complexes do not have any disorder while others can have as much as one third of their residues in the disordered state (Figure 3). The three quarters of complexes with non-zero disorder have on average 4.3% disorder in the aligned regions and about 1.6% in the footprint regions. Confirmed disordered regions have similar disorder content for pairs with non-zero disorder and drops to about 1% if all 588 interactions are included. The reason is that disordered regions with missing coordinates sometimes do not overlap with the predicted disordered regions. There are also families that exhibit rather wide variation in fraction disorder among different members of these families (a ratio of standard deviation over the mean value of fraction disorder is greater than 1); they constitute 13% of all cases.
Table 1 shows several cases of complexes with disorder that were confirmed by experimental studies to be functional. Proteins from these families are found to function in dimer, tetramer and other oligomeric states. Their disordered regions play important roles in regulating the specificity of interactions between the dimer complexes and their interacting partners, in establishing the links between different residues upon allosteric regulation, and possibly in kinetics. In this table we highlight the generality of this phenomenon for many different proteins including enzymes, chaperones and others. As can be seen from this table, all cases (except for the last one) constitute homodimer complexes and, as will be shown in the next section, homodimers have a tendency to contain larger fractions of disordered regions compared to heterodimers. References for Table 1 can be found in Table S1(a).
Here we describe in detail one example from the table: a complex of heat shock protein hsp31 which has chaperone activity and functions as a homodimer in solution (1PV2 [55]) (Figure 4). The complex contains four dimers in a triclinic cell exhibiting a conserved symmetrical homodimer binding mode. Structures of the homodimers show significant fraction disorder of about 8–9% in both aligned and footprint regions. Disordered regions D2 and D3 are found at positions 27–49 and 109–115 and part of the first and the entire second region are also predicted to be disordered by the sequence-based method [32]. These regions have particular functional importance as they are located close to the dimer interface and at high temperatures become disordered and expose a large hydrophobic interface area that helps in binding to client proteins [55]. When the temperature decreases, D2 and D3 lock in certain conformations and facilitate the removal of the client protein from the hydrophobic patch.
We performed an analysis separating all interacting pairs from our test set into homo- (535 complexes) and heterodimers (53 complexes), where both chains in a pair are classified as belonging to the same or different families respectively. Similarly, the prevalence of homodimers over heterodimers in a cell was reported previously [56]. All homodimers were separated into symmetrical and non-symmetrical classes (“isologous” and “heterologous” according to [57]). We define symmetrical homodimers as those that use more than 80% of the same surface in both subunits for binding (316 complexes); all other homodimer arrangements were defined as non-symmetrical (266 complexes). Some homodimer families have structures belonging to both symmetrical and non-symmetrical classes (near the 80% cutoff) but such cases are rare. Eleven families form both homo- and heterodimers. The majority of such cases are examples of larger complexes where the same protein participates in homo- and hetero-interactions within the same complex.
Figure 5 shows average fraction disorder in different classes of homo- and heterodimers. As can be seen from this figure, fraction disorder in complexes decreases as the interaction interface deviates more from being a symmetrical homodimer interface. Fraction disorder in heterodimers is almost two times smaller compared to symmetrical homodimers and the difference is statistically significant (p-value<0.001). The observed trend for hetero- and non-symmetrical homo-complexes to contain smaller disordered regions was confirmed by the disorder prediction analysis, although the trend is not as pronounced for predicted disorder in aligned regions. We did not find significant differences in fraction disorder between homo- and heterodimers for proteins that participate in homo- and hetero-interactions within the same complex.
In studying disorder in protein complexes, we can use the monomer states of the proteins as references. First we would like to check whether the disorder-to-order transition may occur upon binding; and second, to analyze if this transition happens on binding interfaces. In this section we compared fraction disorder of proteins in their monomer and complex states. By definition, binding interfaces should involve only residues with coordinates and therefore can introduce bias toward ordered regions in the complexes (complexes with the entire interface disordered are not considered in the analysis). Therefore, for fair comparison between monomers and complexes we subtracted the number of disordered residues in a monomer which are mapped onto interfaces in a complex from the overall number of disordered residues in a monomer.
Figure 6 shows fraction disorder in aligned regions for monomer and complex structures of the same interaction using the “test588” and “test149” sets. As can be seen from this figure, there exist three types of behavior: cases with higher fraction disorder in a monomer compared to the complex, cases with higher fraction disorder in a complex and, finally, those interactions with no preference towards disordered or ordered states in a monomer or a complex. It should be mentioned that different ways of averaging over structures or using confirmed disorder regions does not change the overall result, namely, that there are three groups and that the sizes of the first and second groups are comparable.
While in the previous section we focused on the disordered regions spanning the whole aligned or footprint regions, here we will focus on disorder in the interface regions. Since the interface in complexes is ordered by definition, we looked at disordered regions in monomers which are aligned to the interface region of the same protein in a complex. The monomer reference state gives us an opportunity to analyze the disorder in the regions of a monomer which form the interface upon binding. We found that the mapped (inferred) interface regions can be up to 50% disordered in a monomer and for 42% of the complexes (23 out of 55 complexes for which p-values can be calculated, see Methods), there is a statistically significant bias toward the disorder on inferred interface regions with p-values of less than 0.05. We observed similar fractions of cases with significant disorder on inferred interfaces if we use confirmed disorder regions (see Methods). Additional restriction of monomeric states by PISA indicates 75% of the cases have significant disorder on interfaces (9 out of 12 complexes from “test149” used for p-value calculation).
Several cases with significant disorder on inferred interfaces are listed in Table 2 (and in Table S1(b) to include references). Their disordered regions predicted by FoldUnfold and by five other methods are highlighted in Table S2. Figure 7 shows one example of ubiquitin C-terminal hydrolase in two states: monomeric (1UCH [58]) and in complex (1XD3 [59]) with ubiquitin vinylmethylester, a ubiquitin-based active site-directed probe. Ubiquitin C-terminal hydrolase catalyzes the hydrolysis of the isopeptide linkage between the C-terminal glycine of ubiquitin and a lysine of the target polypeptide. The structure of the free form of this enzyme has 4–6% fraction disorder in footprint and aligned regions compared to only 0–0.9% fraction disorder in the complex with ubiquitin. The disordered region in 1UCH constitutes a 20 residue loop (147–166) which is also predicted to be disordered (region 150–164) by the sequence-based method [32]. This disordered loop is positioned just over the active site cleft and becomes ordered upon binding to ubiquitin vinylmethylester. The interaction interface mapped from complex structure to monomer shows that 30% of the interface is disordered in a monomer (binomial p-value<10−8) which points to the coupling between folding and binding. It was suggested earlier that this disordered loop might prevent access to the active site for larger substrates and affect substrate specificity as larger substrates could only be accommodated in the active site by peeling away this loop from the active site cleft [58],[59].
Our large-scale study of disordered regions in proteins and protein complexes underscores a fascinating diversity among the biological processes that make use of protein disorder. Analysis of GO functional annotations of complexes reveals a variety of categories where intrinsic disorder can play an important functional role, the most frequent of them being nucleic acid binding proteins, enzymes, ATP binding proteins, receptor binding proteins and other ligand binding proteins (see Dataset S3). In addition to well-documented cases of signaling and transcription related proteins, we detect and describe intrinsic disorder in a large variety of enzymes and other proteins. In accordance with the conventional view that folding of disordered regions occurs upon binding to the interaction partners, we find many such cases in our analysis where ordering occurs upon complex formation. Moreover, we investigated the details of protein interaction interfaces and deduced changes occurring on the interfaces in disorder-to-order transitions. We find that in 42–75% of interactions (for which statistical significance could be estimated), there is evidence that disorder-to-order transition occurs on binding interfaces.
Many complexes in our dataset have significant amounts of intrinsic disorder. The role of disordered regions in complexes has been analyzed in several previous studies on smaller test sets [22],[60]. In our study we find as many cases with disorder in complexes as the number of instances of disorder-to-order transition upon binding. This is a rather unusual result as many such cases until recently were largely overlooked. It has been proposed that disordered regions can be energetically beneficial in proteins and their complexes due to a number of reasons: they can provide an increase in backbone conformational entropy upon ligand binding, can accommodate sites for post-translational modifications, and can provide interfaces for binding other partners [6], [22], [60]–[65]. In addition, the formation of complexes of proteins containing functionally important disordered regions can help to increase their stability (entropy-driven complexation, see the last section) and prevent their degradation.
Many proteins perform their functions while interacting with each other in larger complexes. We argue that intrinsic disorder in complexes may play an important functional role in regulating the specificity of interactions between the dimer complexes and their interacting partners, in establishing the links between different residues upon allosteric regulation, and in possibly influencing the kinetics. For example, the mechanisms of regulation of binding specificity through disordered regions in complexes can be as common as for unbound proteins: controlling the exposure of the dimer interface or nearby regions for potential binding targets, or providing specific binding for substrates of certain sizes. The former mechanism has been recently investigated in the stable symmetrical homodimers, UmuD2 and UmuD2′, which lack secondary structure and might lock the disordered regions in conformations that facilitate further binding of other proteins [66]. In addition, the formalism of flexible folding and mechanism of the “conformational selection” model [19], [67]–[72] can be expanded to include the binding between protein complexes and their interacting partners.
Interestingly, we find that the disorder content in homodimers, especially in symmetrical homodimers, is significantly higher than in heterodimers. Indeed, many soluble and membrane-bound proteins form homo-oligomeric complexes in a cell and oligomerization can generate new binding sites at dimer interfaces to increase specificity and diversity in the formation of complexes. Indeed, intrinsic disorder in homodimers might have more pronounced functional importance compared to the disorder in heterodimeric complexes. Symmetrical arrangements in homodimers might be crucial to keep functional disordered regions close together in space to form joint binding interfaces or to form near-interface regions to regulate the accessibility of the binding partner. Moreover, from the energetic point of view, symmetrical homodimers have an advantage over non-symmetrical arrangements [73],[74]; at the same time, self-interactions between disordered parts in homodimers can be of evolutionary and functional importance [66],[75].
Another explanation comes from thermodynamics considerations. Entropy of complexation gives an important contribution to the complex stability and drives macromolecular complexes to less symmetric states. Any rearrangement of monomers that decrease complex symmetry would therefore result in a more stable complex (see Eq. 20 in [52]). The presence of disordered regions in the symmetrical homodimers will make the protomers asymmetric and change the symmetry number γ from 2 to 1 (two-fold symmetry to asymmetry) and make a favorable contribution to the free energy. At the same time disordered regions should not affect symmetry numbers in cases of heterodimers or non-symmetrical homodimers (they are asymmetric by default) and will not change their stability. Ultimately, the interplay between the binding energy and entropy contributions is important and it is not unrealistic that the entropy-driven disordered complex formation can be realized in some cases.
It is difficult to systematically account for all factors which influence the fraction disorder in proteins. The amount of disorder in crystals depends in general on crystallization conditions and crystal packing parameters. The balance between order and disorder is rather subtle and is difficult to detect but the evidence pointing to the tremendous importance of intrinsic disorder in a large variety of cellular processes is accumulating and merits further study.
|
10.1371/journal.pcbi.1005848 | Choose, rate or squeeze: Comparison of economic value functions elicited by different behavioral tasks | A standard view in neuroeconomics is that to make a choice, an agent first assigns subjective values to available options, and then compares them to select the best. In choice tasks, these cardinal values are typically inferred from the preference expressed by subjects between options presented in pairs. Alternatively, cardinal values can be directly elicited by asking subjects to place a cursor on an analog scale (rating task) or to exert a force on a power grip (effort task). These tasks can vary in many respects: they can notably be more or less costly and consequential. Here, we compared the value functions elicited by choice, rating and effort tasks on options composed of two monetary amounts: one for the subject (gain) and one for a charity (donation). Bayesian model selection showed that despite important differences between the three tasks, they all elicited a same value function, with similar weighting of gain and donation, but variable concavity. Moreover, value functions elicited by the different tasks could predict choices with equivalent accuracy. Our finding therefore suggests that comparable value functions can account for various motivated behaviors, beyond economic choice. Nevertheless, we report slight differences in the computational efficiency of parameter estimation that may guide the design of future studies.
| In economic decision theory, value is a construct that provides a metric to compare options: agents are likely to select options leading to high-value outcomes. In neuroscience, different behavioral tasks have been used to elicit the subjective values of potential outcomes, notably rating tasks, which demand an explicit value judgment on the outcome, and effort tasks, which demand an energetic expense (in order to increase outcome probability). However, it remains unclear whether the values elicited by these tasks are the same as the values that drive choices. Indeed, it has been argued that they involve different costs and consequences for the agent. Here, we compared value models that could account for behavioral responses in choice, rating and effort tasks involving the same set of options, which combined a monetary gain for the participant and a donation to a charity. We found that the most plausible model was that a same value function, with a similar selfishness parameter (relative weight on gain and donation), generated the responses in all three tasks. This finding strengthens the notion of value as a general explanation of motivated behaviors, beyond standard economic choice.
| Value (or utility) functions have been defined to account for preferences revealed in choice tasks [1]. One basic principle is that if an agent prefers A over B, then for this agent the value of A is higher than the value of B. Assuming basic axioms of expected utility theory, cardinal functions have been described, such that option values can be positioned on a numeric scale [2]. Cardinal values rely on the notion that choice probability depends on the distance between option values, as well as on their distance from a reference point [3]. Value functions can be parameterized when choice options are combinations of objective quantities, e.g., the probability and magnitude of monetary payoff. The parameters can then be estimated through fitting procedures that maximize the likelihood of observed choices under the valuation model. Fitting choices involves specifying a function relating choice probability to option values, generally a softmax rule [4]. Thus, most studies have used choice data to infer functions that assign cardinal values to any possible option.
Alternatively, a more direct approach has been used in the neuroeconomics literature, using behavioral tasks in which subjects assign cardinal values to available options, instead of inferring value functions from their choices. One possibility is to ask subjects to rate on analog scale the desirability (or likeability) of the outcomes associated to the different options [5]. Another possibility is to ask subjects to express the maximal cost (e.g. price, effort or delay) that they are willing to endure in order to obtain these outcomes [6, 7]. The aim of the present study was to compare the value functions derived from these direct cardinal measures with the value functions derived from fitting choice data. We selected, in addition to a standard binary choice task where subjects state their preference between two options, a subjective rating task where subjects score the desirability of every possible outcome and an effort production task where the probability of obtaining the outcome depends on the force produced with a handgrip. Standard models of behavior in these tasks suggest that ratings and forces can be taken as direct measures of the subjective outcome values that drive choices (see Methods).
However, there are a priori reasons why the value functions elicited by the different tasks should differ in their form or in their parameters. In our perspective, the key difference between tasks is the nature of the cost. In choice tasks, the response entails an opportunity cost, corresponding to the value of the non-selected option [8]. The response is therefore based on the value difference between the two possible outcomes, which is often called decision value. As the motor response is generally similar for the two options, there is no need to consider action costs. In effort tasks, the response is associated with a specific cost due to energy expenditure, which may be signaled through muscular pain. The response therefore aims at maximizing the net value, i.e. the trade-off between outcome value and action cost [9]. In rating tasks, the variation in action cost across the possible positions on the scale is usually negligible, although the extremes may be longer to reach. Thus, the response should be a direct expression of outcome value. As decision values, net values and outcome values may be computed by different brain systems, they may follow different functions [10].
In addition, there is a cost that may be common to all behavioral tasks, which is social reprobation. Some responses may be more socially acceptable than others, particularly if moral considerations are involved [11]. This social cost may be more salient in rating tasks, which have no other consequences and can therefore be considered as ‘hypothetical’ decisions. By opposition, choice and effort tasks are typically consequential: they determine the outcome, either deterministically or probabilistically, and therefore involve ‘real’ decisions. Hypothetical and real decisions have been compared in a number of studies using various tasks [12–16], with contrasted results and no proper model comparison. Yet it may seem intuitive that subjects in rating tasks are more likely to pretend having values they do not have, for reputation concerns, because there is no obvious costly consequence. To assess this potential difference between tasks we used options that combined money for the subject (gain) and money for a charity (donation), with the aim of triggering moral dilemma.
Also, each behavioral task may be susceptible to specific artifacts. For instance, the rating scale is somewhat arbitrary, and may yield distortions of value functions due to framing or anchoring phenomena [17], particularly if subjects are not familiar with the range of values spanned in the set of options. Effort exertion, between zero and maximal force, may be less arbitrary but susceptible to fatigue, which may increase with the number of performed trials and influence effort cost, and hence the values expressed by participants [18].
In the present study, we compared the value functions elicited by the different tasks for a same set of composite outcomes, each combining gain and donation. We found that the same valuation model provide the best fit of behavior in the three tasks, with slight differences in parameter estimates.
Subjects (n = 19) participated in three tasks aimed at measuring subjective values of bi-dimensional outcomes composed of one gain for themselves and one donation for a charity organization they selected prior to the experiment (Fig 1, top). In the rating task, participants rated how much they would like to obtain the composite outcome using a scale graduated from 0 to 10. The feedback was probabilistic and they obtained the outcome in 70% of the trials, irrespective of their ratings, which were therefore not consequential. The probabilistic contingency was adjusted so as to match that of the effort task. In the force task, subjects had to squeeze a handgrip knowing that the chance to win the outcome was determined by the ratio of the force they produced during the trial and their maximal force measured beforehand. Note that previous experiments in the lab using the grip task with similar range of incentives showed that subjects produce on average about 70% of their maximal force [19]. In the choice task, participants had to choose between two composite options, the selected outcome being obtained in 70% of trials. The choice task followed an adaptive design [20] in which options were proposed so as to optimize the parameterization of an a priori value function (linear integration of gain and donation with their interaction).
As expected, explicit ratings, forces produced and subjective values inferred from choices all increased with incentives, i.e. with both gain and donation (Fig 1, bottom). Before going into more sophisticated models, we conducted linear regressions (for ratings and forces) or logistic regression (for choices) against the two main factors (gain G and donation D) and their interaction. Regression estimates obtained for main factors were significantly different from zero in all cases: in the rating task (βR(G) = 0.07±6.10−3, t(18) = 11.5, p = 1.10−9; βR(D) = 0.06±7.10−3, t(18) = 8.2, p = 1.10−7), in the force task (βF(G) = 0.05±6.10−3, t(18) = 8.5, p = 1.10−7; βF(G) = 0.05±6.10−3, t(18) = 7.2, p = 9.10−7) and in the choice task (βC(G) = 0.16±0.03, t(18) = 5.6, p = 2.10−5; βC(G) = 0.12±0.02, t(18) = 5.4, p = 4.10−5). Interaction terms were significant for the rating and force tasks but not for the choice task (βR(G*D) = -2.10–4±9.10−5, t(18) = -2.7, p = 0.01; βF(G*D) = -3.10–5±1.10−5, t(18) = -2.6, p = 0.02; βC(G*D) = 1.10–5±2.10−4, t(18) = 0.1, p = 0.95). In none of the tasks did we find a significant difference between the weights of gain and for donation, although there was a trend in favor of selfishness (R: t(18) = 1.79, p = 0.089; F: t(18) = 1.10, p = 0.29; C: t(18) = 1.70, p = 0.11).
We also regressed the residuals of this regression against trial and session number, in order to test for fatigue effects. As none of these tests was significant (all p>0.1), we did not include any parameter accounting for fatigue in our computational models. Finally, we compared the distribution of forces and ratings, irrespective of gain and donation. As uncertainty was controlled by force production in the effort task, the distribution could be affected by risk attitude, relatively to the rating task in which uncertainty was constant. Indeed, subjects should avoid medium forces, if they are risk averse, or on the contrary favor them, if they are risk seeking. We thus fitted a second-order polynomial function to individual distributions of forces and ratings. The coefficients of quadratic regressors were significant for both tasks (F: b = -0.31 ± 0.11, t(18) = -2.75 p = 0.013, R: b = -0.21 ± 0.06, t(18) = -3.34, p = 4.10–3), with no significant difference between tasks (t(18) = -0.85, p = 0.41). There was therefore no evidence that risk attitude created a difference between forces and ratings.
However, these model-free analyses do not provide any formal conclusion about how value functions differ across tasks, so we now turn to a model-based Bayesian data analysis.
In order to further investigate how changing the elicitation paradigm could affect the subjective value of potential outcomes, we defined a set of twelve value functions that could explain the observed behavior in each task (see Methods). These value functions represent different ways of combining the two dimensions (gain and donation) composing the outcomes proposed in the tasks. They were used to generate forces and ratings with linear scaling (with slope and intercept parameters) and choices with logistic projection (softmax function with temperature parameter). All value functions were fitted on behavioral responses for every subject and task using Variational Bayesian Analysis (VBA) [21, 22]. The explained variance (averaged across subjects) was comprised between 43 and 70% in the force task, between 57 and 85% in the rating task and between 45 and 85% in the choice task. These results show that, for all three tasks, there were important differences in the quality of fit between value functions, which we compare below.
In this study, we showed that three tasks varying on several features elicited the same value function accounting for participants’ behavior. Moreover, the most critical parameter, precisely the relative weighting of gain and donation (selfishness), was similar in the three tasks. However, we found some differences in the concavity of value functions. In addition, the different tasks presented practical advantages and disadvantages that should be taken into account when selecting a particular elicitation procedure.
We showed with a Bayesian model comparison that the same value function could account for the three types of behavior. It is interesting to note that Bayesian inference enables concluding in favor of the null hypothesis, which cannot be formally validated from an absence of significant difference in classical statistical inference. The null hypothesis (no difference in value function) is consistent with subjects maximizing simple net utility functions defined as the difference between expected outcome values in the choice task, the expected outcome value minus a quadratic effort cost in the effort task, and the similarity of overt rating and covert judgment in the rating task (see Methods). This means that the computational processes used to generate the different behaviors (choice, rating, force) from underlying outcome values have no backward influence on these values. As a consequence, the results reported in the neuroeconomic literature using the different tasks, regarding the brain valuation system in particular, can be directly compared.
The winning value function, called ‘Constant Elasticity of Substitution’ [24], has been shown to provide a good account of choices made by participants in other experiments that involved sharing money with others [25], which is consistent with the present results. It has the advantage of simplicity, with only two parameters: one controlling the relative weighting of outcome dimensions (here, the selfishness parameter) and one controlling the interaction between dimensions (the concavity parameter). Note that the other value functions used in the model comparison also provided a satisfying fit of behavioral data, capturing the relative sensitivity to gain and donation. Thus, we do not wish to make a strong claim that the CES function should be used in any task assessing altruistic behavior. We simply used it in the following because it was the best candidate function to investigate the integration of outcome dimensions.
The three tasks not only shared the same value function, but also elicited similar selfishness parameters. Thus, the differences in the consequentiality of the behavioral response, and in the nature of associated costs, did not impact the effective weights assigned to the gain and donation dimensions. This may come as a surprise, given that exhibiting altruism comes for free (with no cost) in the rating task but not in the choice task (where there is an opportunity cost) or the force task (where there is an effort cost). This result suggests some stability across elicitation procedures in how dimensions are weighted. It is consistent with previous studies reporting similar values for hypothetical and real decisions [12–14]. In our data, the selfishness parameter denoted a preference for gain over donation, which is consistent with what has been observed in studies investigating altruism [26, 27]. Yet we note that our participants appeared less selfish, possibly because we asked them to select a NGO which they would give money to, instead of asking them to share money with another participant who they did not know.
We acknowledge that our demonstration of a same value function for different tasks suffers from some limitations. First, the range of costs involved in the choice and effort tasks remained reasonable. It is likely that costs should be integrated in the value function if they get more extreme (say if winning one euro for a charity demands days of work). Second, the stability of elicited value functions was assessed within subjects, which may favor consistency in behavioral responses. Results might have been more variable had we tested separate groups of subjects on the different tasks or the same subjects on different days. Indeed, the measures might be differentially sensitive to states such as mood or fatigue, which were not controlled in our design. Third, our conclusion could be specific to the particular dimensions that composed the outcomes presented in our tasks. Further experiments would be needed to generalize the result to other multi-attribute options, as in for example risky or inter-temporal choice, or to more natural multidimensional options such as food items.
Even if the same value function and the same selfishness parameter could explain the behavior in the three tasks, we found a significant difference between tasks in the concavity parameter. Indeed, the choice task did not reveal any concavity, indicating no interaction between dimensions, whereas the force task, and to a lesser extent the rating task, revealed a concavity, denoting a biased sensitivity to high monetary amounts, irrespective of the receiver. It remains difficult to conclude whether the concavity seen in rating and force tasks denotes an artifactual distortion of the actual value function or a better sensitivity to actual values, compared to the choice task which is more complex (with four numbers to be integrated). Indeed, concavity in the effort task may be higher because the effort cost function is not quadratic, as we assumed for the sake of simplicity. One may also speculate that high amounts trigger arousal responses, which may affect effort production but choice or rating. Alternatively, concavity in the choice task may be absent because in most cases, there are high amounts in both options. Note that choice options in our design were selected to optimize a value function (linear with interaction) where there was no concavity parameter. Nevertheless, even if no concavity was observed on average in the choice task, the model with a concavity parameter was favored by the Bayesian selection. This means that some subjects were better fitted with concave and others with convex value functions. This inter-subject variability possibly reflects differences in the sensitivity to equity (options with similar amount for them and for the charity).
Independently of the elicited value function, we assessed how the tasks differed in terms of precision and speed of parameter estimation. The choice and rating tasks were better fitted, with higher coefficients of determination than the force task. However, the value functions inferred from the rating and force tasks were equally capable of predicting choices. It was therefore not that the value function elicited with the force task was distorted or variable, but simply that the force data were noisier. Thus, if the objective is to predict choices, there is no reason, based on the accuracy criterion, to prefer any particular task.
On the other hand, response times recorded in the force task were shorter than in the rating task. Moreover, without design optimization, there was no significant reduction in the number of trials needed for stabilizing parameter estimation with the rating task compared to the force task. Thus, the speed criterion (total task duration) seems to be in favor of the force task. Note that this advantage could vanish if responses were mapped to ratings in a different way, for instance with one key per value. Also, the effort task requires some equipment and a calibration phase to determine maximal force, which may mitigate the gain in task duration.
Finally, for a similar precision and speed, the choice task needs an adaptive design (for the selection of choice options), which implies to posit priors on value functions and on parameters, whereas the other tasks can be run in a model-free manner. Thus, the simplest way to experimentally measure subjective value functions might not, eventually, be the binary choice task that is standard in behavioral economics.
To our knowledge, this is the first study comparing direct elicitation of cardinal values (rating and force tasks) to ordinal rankings (choice task) for a same set of options. Those tasks are widely used in neuroeconomics and it is somewhat comforting that they reveal similar value functions driving the behavior despite trivial differences. They nonetheless present different advantages and drawbacks that may guide the design of future studies.
The study was approved by the Pitié-Salpétrière Hospital ethics committee. All subjects were recruited via e-mail within an academic database and gave informed consent before participation in the study.
Participants were right-handed, between 20 and 30 years old, with normal vision and no history of neurological or psychiatric disease. They were not informed during recruitment that the task was about giving money to a charity, in order to avoid a bias in the sample. Nineteen subjects (10 females; age, 22.2 ± 1.4) were included in the study. They believed that the money won while performing the task would be their remuneration for participating, but eventually, their payoff was rounded up to a fixed amount (100€).
Subjects performed the three tasks, the order being counterbalanced across subjects for the force and rating tasks. The choice task was always performed after the two others, which were performed during MRI scanning for other purposes.
The force task was preceded by maximal force measurement for the right hand [6]. Participants were verbally encouraged to squeeze continuously as hard as they could until a line growing in proportion to their force reached a target displayed on a computer screen. Maximal force was defined as the maximal level reached on three recordings. Then subjects were provided a real-time feedback about the force produced on the handgrip, which appeared as a red fluid level moving up and down within a thermometer, the maximal force being indicated as a horizontal bar at the top. Subjects were asked to try outreaching the bar and state whether it truly corresponded to their maximal force. If not, the calibration procedure was repeated.
In the force and rating tasks, 121 trials were presented in a random order across three sessions of 40 or 41 trials. Each trial corresponds to one of the 121 combinations of the experiment design (eleven possible incentives for themselves by eleven possible incentives for charity donation: from 0€ to 100€ with steps of 10€). Subjects performed the three sessions with the right hand, with short breaks between sessions to avoid muscle exhaustion.
In the force and rating tasks, each trial started by revealing the potential outcome, composed of two monetary incentives, with the inscriptions “YOU” followed by the amount for the subject, and “ORG” followed by the amount for the charity (Fig 1, top). The outcome was displayed for a duration jittered between 4 and 6 seconds. In the force task, subjects knew that the probability to win the outcome was proportional to the force they would produce after the display of the thermometer on the screen. More precisely, the probability of winning was equal to the percentage of their maximal force that they produced in the current trial. Subjects were also instructed to manage their forces in the effort task to avoid any frustration due to potential fatigue effect, and to use breaks between sessions to recover their muscular strength. During task trials, they were provided with online feedback on the exerted force (via a fluid level moving up and down within a thermometer). They were also informed that they had to produce a minimal effort in every trial (10% of their maximal force) and that the trial would be over when they stop squeezing the handgrip. Each trial ended with the display of the final outcome of their effort, for a duration jittered between 4 and 6 seconds, via the words “WON” (with the proposed monetary earnings) or “LOST” (with null earnings for both subject and charity).
The rating task only differed at the time of the motor response. Instead of a thermometer, a vertical rating scale from 0 to 10 units appeared after presentation of the potential outcome. Subjects were asked to rate the desirability of the outcome on the screen by moving the cursor through button presses with the right hand (index and middle finger for moving the cursor left and right, and ring finger for validating the response). They were asked to use the whole scale across trials. They were also informed that their rating would have no impact on the final outcome. They were then shown the final outcome that was randomized to obtain a “WON” in 70% trials, and a “LOST” 30% of trials (i.e., a proportion similar to that obtained in the force task).
The binary choice task included 200 trials, each presenting two composite options, one on each side of the screen. After considering the two options for 2 seconds, subjects could indicate the one they would prefer to win using their right hand (index vs. middle finger for left vs. right option). This option was actually won in 70% of trials, which was indicated with a positive feedback (“WON”) accompanied by the selected earnings. In the other 30% of trials, a negative feedback (“LOST”) was shown with a null outcome (0€) for both receivers.
Given the number of options in our design, there were 1212 (14641) possible binary choices. Constraints can be applied to reduce this number: choices are informative only if options are crossed (attributes never dominate on both dimensions), if options differ on both dimensions, and if the pair of options was not previously presented. However, those constraints only reduced the number of choices to 3025. Thus, we used an online optimization design to exploit the fact that some options are more informative than others to estimate a value function. At each trial, the design was optimized over a single dimension (gain or donation). The chosen combination was the one that minimized the trace of the posterior covariance matrix over the parameters of an a priori value function defined as follows: V(G,D) = βG * G + βD * D + βGD * (G * D), corresponding to a linear integration with interaction [20]. Contrary to the force and rating tasks, the amounts for subjects and charity could vary with steps of 1€ (still between 0€ and 100€), since options were optimized for each trial and subject.
Subjects were informed that three trials would be randomly drawn (one per task) and that the average outcome would be actually implemented (including both their gain and donation). They were aware that their responses in the rating task would have no influence on the outcome, whereas they would have an impact in the effort and choice task. The uncertainty about winning the outcome was fixed to 70% in the choice and rating tasks, but controlled by the force produced in the effort task. As expected, the average forces were not significantly different from 70% (65±3%, p>0.1), and hence matched the uncertainty level of the other tasks.
|
10.1371/journal.ppat.1005972 | NK-CD11c+ Cell Crosstalk in Diabetes Enhances IL-6-Mediated Inflammation during Mycobacterium tuberculosis Infection | In this study, we developed a mouse model of type 2 diabetes mellitus (T2DM) using streptozotocin and nicotinamide and identified factors that increase susceptibility of T2DM mice to infection by Mycobacterium tuberculosis (Mtb). All Mtb-infected T2DM mice and 40% of uninfected T2DM mice died within 10 months, whereas all control mice survived. In Mtb-infected mice, T2DM increased the bacterial burden and pro- and anti-inflammatory cytokine and chemokine production in the lungs relative to those in uninfected T2DM mice and infected control mice. Levels of IL-6 also increased. Anti-IL-6 monoclonal antibody treatment of Mtb-infected acute- and chronic-T2DM mice increased survival (to 100%) and reduced pro- and anti-inflammatory cytokine expression. CD11c+ cells were the major source of IL-6 in Mtb-infected T2DM mice. Pulmonary natural killer (NK) cells in Mtb-infected T2DM mice further increased IL-6 production by autologous CD11c+ cells through their activating receptors. Anti-NK1.1 antibody treatment of Mtb-infected acute-T2DM mice increased survival and reduced pro- and anti-inflammatory cytokine expression. Furthermore, IL-6 increased inflammatory cytokine production by T lymphocytes in pulmonary tuberculosis patients with T2DM. Overall, the results suggest that NK-CD11c+ cell interactions increase IL-6 production, which in turn drives the pathological immune response and mortality associated with Mtb infection in diabetic mice.
| In the current study, we employed an experimentally induced type 2 diabetes mellitus (T2DM) model in wild type C57BL/6 mice and investigated the immune response to Mycobacterium tuberculosis (Mtb) infection. We found that natural killer (NK) and CD11c+ cell interactions in Mtb-infected T2DM mice led to increased IL-6 production, which drives the pathological immune response and reduces survival of Mtb-infected T2DM mice. We also found that IL-6 increases inflammatory cytokine production in pulmonary tuberculosis patients with T2DM. The NK-CD11c+ axis and the IL-6 pathway may be promising new targets for host-directed therapies aimed at reducing the severity of immune pathology, which drives morbidity and mortality in those infected by tuberculosis (TB).The study demonstrates for the first time that NK-CD11c+ cell interactions increase IL-6-mediated inflammation and reduce survival in T2DM mice infected with Mtb. The NK-CD11c+ cell axis appears to be a promising new target for reducing inflammation and mortality in tuberculosis patients with type 2 diabetes.
| Mycobacterium tuberculosis (Mtb) infects one-third of the world’s population and causes almost 1.3 million deaths per year [1]. Approximately 90% of those infected have a latent tuberculosis infection and develop protective immunity to contain it; however, but 10% progressive to active tuberculosis (TB) disease months or years after infection [2]. The risk for progression to TB disease is increased by acquired factors including human immunodeficiency virus (HIV) infection, alcoholism, smoking, and diabetes [3].
Developing nations are epicenters of diabetes [4]. Diabetes mellitus alters innate and adaptive immune responses and increases the risk of developing active TB [5]. In type 2 diabetes mellitus (T2DM) patients, there is a reduced association between mycobacteria and monocytes; therefore, phagocytosis via complement receptors is compromised [6,7]. Mtb-infected diabetic mice show delayed priming of the adaptive immune response, which is necessary to restrict Mtb replication [8]. Hyperactive T-cell responses and increased Th1 and Th17 cytokine production are noted in TB patients with type 2 diabetes [9]. Limited information is available about experimental models used to study the effects of T2DM during Mtb infection. Spontaneous T2DM rodent models, such as GK/Jcl rats, have a higher bacterial load and increased immune pathology than non-diabetic Wistar rats after infection with a Mtb Kurono aerosol [10]. Furthermore, T2DM guinea pigs are highly susceptible to Mtb infection; even non-diabetic hyperglycemia exacerbates disease severity [11,12]. However, a detailed understanding of the protective immune responses in type 2 diabetic hosts during Mtb infection is essential if we are to develop an adequate prophylactic or therapeutic agent.
In the current study, we employed an experimentally induced T2DM model in wild type C57BL/6 mice and investigated the immune response to Mtb infection. We found that natural killer (NK) and CD11c+ cell interactions in Mtb-infected T2DM mice led to increased IL-6 production, which drives the pathological immune response and increases mortality. We also found that IL-6 increases inflammatory cytokine production in pulmonary tuberculosis patients with T2DM.
A combination of streptozotocin (STZ) and nicotinamide (NA) induces T2DM in mice [8]. STZ (180 mg/kg of body weight) and NA (60 mg/kg of body weight) were administered intraperitoneally to C57BL/6 mice three times, with an interval of 10 days between doses. A schematic representation of T2DM induction is shown in Fig 1A. After 1 month, mice developed T2DM, as assessed by measurement of blood glucose and serum insulin levels. Blood glucose levels measured at monthly intervals for up to 8 months in STZ/NA-treated mice were consistently ≥250 mg/dl (Fig 1B). Blood glucose levels in control mice were 60–125 mg/dl.
To determine whether STZ/NA-treated mice developed insulin resistance, we next measured serum insulin levels. One and three months after T2DM induction, serum insulin levels in STZ/NA-treated mice were significantly higher than those in control mice (Fig 1C). Six months after T2DM induction, serum insulin levels in STZ/NA-treated mice were 4-fold higher than those in control mice (Fig 1C). Insulin resistance, a characteristic feature of T2DM, was confirmed by oral glucose tolerance test (OGTT) 6 months after STZ/NA injection (Fig 1D). Furthermore, serum levels of cholesterol, triglyceride, and free fatty acids were elevated by 6 months after STZ/NA treatment (Fig 1E–1G). Dyslipidemia is another characteristic of T2DM in humans and, combined with demonstrated insulin resistance, confirms the validity of our mouse T2DM model.
We next investigated TB defense in T2DM mice by aerosol challenge with Mtb as shown in Fig 2A. One and three months post-infection (p.i.), the lung bacterial burden was similar in T2DM and control mice (Fig 2B). However, by 6 months p.i., lung bacterial burden was significantly greater in T2DM mice compared to controls (Fig 2B). A similar increase in the bacterial burden was observed in the spleens and livers of T2DM mice when compared with those of control mice (data presented in Dryad Data Repository; doi:10.5061/dryad.qn42t).
Alveolar macrophages are the first immune cells that Mtb encounters in the lung [13]. To determine whether the increased bacterial growth described above was due to altered antimicrobial function of these cells, we isolated alveolar macrophages from control and T2DM mice (one, three and six months after T2DM induction) and infected them with Mtb. The CFU were quantified after 5 days. Mtb growth was similar in the alveolar macrophages of control and T2DM mice after one and three months post induction of T2DM. However, control of Mtb growth was impaired in alveolar macrophages, six months after the induction of T2DM (Fig 2C). We next determined the survival of uninfected control and T2DM mice and of Mtb-infected control and T2DM mice. By 10 months p.i. all Mtb-infected T2DM mice died, whereas only 40% of the uninfected T2DM mice and 6.6% of the Mtb-infected non-diabetic mice died (Fig 2D). In contrast, all control mice survived.
We next determined whether T2DM has any effect on pro- and anti-inflammatory responses following Mtb infection. Control and T2DM mice were infected with Mtb, and after 1 and 6 months the levels of various cytokines and chemokines were measured in lung homogenates by multiplex (23-plex) ELISA. As shown in Fig 3A, there was a significant increase in both pro- and anti-inflammatory cytokines and chemokines in the Mtb-infected T2DM mice at 1 and 6 months p.i., when compared with either uninfected T2DM mice or Mtb-infected control mice. However, the cytokine and chemokine levels in the lungs at 6 months p.i. were significantly higher than those at 1 month p.i. The levels of inflammatory cytokines (IL-6, IFN-γ, TNF-α, IL-1β) and chemokines (MCP-1) in whole-lung homogenates from T2DM mice were significantly higher than those in homogenates from Mtb-infected control mice or uninfected T2DM mice (Fig 3A). In addition, we found that interleukin (IL)-1α, -5, -9, -12 [p40], -12 [p70], and -13, and G-CSF, GM-CSF, KC, and MIP-1β, in the lung homogenates of Mtb-infected T2DM mice were significantly higher than those in Mtb-infected control mice and uninfected T2DM mice at 6 months p.i. (data presented in the Dryad Data Repository; doi:10.5061/dryad.qn42t). We also examined expression of various pro- and anti-inflammatory cytokines in whole-lung tissue using real-time PCR. Similar to the ELISA data, we observed increased expression of TNF-α, IFN-γ, IL-6, IL-1β, IL-21, IL-23, TGF-β, and IL-10 in Mtb-infected T2DM mice (data presented in the Dryad Data Repository; doi:10.5061/dryad.qn42t).
Histological analysis revealed significantly more inflammation throughout the lungs of Mtb-infected T2DM mice when compared with those of Mtb-infected control mice or uninfected T2DM mice (Fig 3B).
IL-6 is a pleotropic cytokine that regulates both pro- and anti-inflammatory cytokine production [14], and it has both protective and pathogenic roles in diabetes [14]. We found that both pro- and anti-inflammatory cytokine production is dysregulated in Mtb-infected T2DM mice compared to control T2DM mice and Mtb-infected control mice. There are conflicting reports about the role of IL-6 in Mtb infection [15,16]. IL-6-deficient mice are susceptible to Mtb infection [15], and IL-6 participates in the induction of type 1 protective T-cell responses after vaccination [17]. However, IL-6 is not required to generate specific immune responses to Mtb infection [18]. Thus, we next determined whether neutralizing IL-6 affects survival, cytokine production, or the bacterial burden in T2DM mice. Fig 4A shows a schematic representation of Mtb infection and anti-IL-6 mAb treatment in T2DM mice. One month after T2DM induction (acute diabetes), mice were intranasally infected with 50–100 CFU of Mtb. At 6 months p.i., the mice were treated with a neutralizing anti-IL-6 mAb, an isotype-matched control mAb, or PBS. As shown in Fig 4B, 65% (p = 0.05) of Mtb-infected T2DM (acutely diabetic) mice that received the isotype-matched control mAb or PBS died within 2 months. By contrast, all mice that received the anti-IL-6 mAb survived. Anti-IL-6 mAb treatment also reduced the bacterial burden in the lungs (Fig 4C), spleen (1.5 ± 0.8 × 104 vs. 10.5 ± 0.8 × 104 CFU; p = 0.0003), and liver (1.75 ± 0.25 × 103 vs. 2.7 ± 0.25 103 CFU; p = 0.03). Real-time PCR analysis of lung samples indicated that anti-IL-6 mAb treatment inhibited expression of IL-17, TNF-α, IL-10, and TGF-β (Fig 4D) when compared with that in mice treated with the isotype-matched control mAb or PBS. Histological examination of lung tissue indicated a similar degree of inflammation in PBS-treated and isotype-matched control antibody-treated mice with acute T2DM and infected with Mtb (Fig 4E). By contrast, anti-IL-6 mAb treatment significantly reduced inflammation in the lungs of Mtb-infected acute T2DM mice (Fig 4E).
We next determined whether neutralizing IL-6 affected survival, cytokine production, or bacterial burden in mice with chronic T2DM (mice were infected 6 months after T2DM induction). On the day of infection, mice received the anti-IL-6 mAb, the isotype-matched control mAb, or PBS (Fig 5A). As shown in Fig 5B, 80% (p = 0.05) of Mtb-infected T2DM mice (chronically diabetic) that received the isotype-matched control mAb or PBS died within 2 months. By contrast, all Mtb-infected chronic T2DM mice that received the anti-IL-6 mAb survived. Anti-IL-6 mAb treatment also reduced the bacterial burden in the lungs (Fig 5C), spleen (1.5 ± 0.3 × 104 vs. 4.4 ± 0.8 × 104 CFU; p = 0.01), and liver (0.6 ± 0.6 × 103 vs. 7.8 ± 1.4 × 103 CFU; p = 0.001). Real-time PCR analysis of lung samples indicated that anti-IL-6 mAb treatment of chronically diabetic Mtb-infected mice was associated with reduced expression of IFN-γ, IL-17, TNF-α, IL-10, and TGF-β (Fig 5D) when compared with that in mice treated with the isotype-matched control mAb. Histological examination of the lungs suggested a similar degree of inflammation in PBS-treated and isotype-matched control antibody-treated mice with chronic T2DM harboring Mtb (Fig 5E). By contrast, anti-IL-6 mAb treatment significantly reduced inflammation in the lungs of Mtb-infected mice with chronic T2DM (Fig 5E).
To confirm our finding that IL-6 levels in the lungs of Mtb-infected T2DM mice increased at 6 months p.i., mice were euthanized and lung sections were examined for IL-6 expression by immunohistochemistry (IHC). As shown in Fig 6A and 6B, the mean histology score (H-score) for IL-6 in Mtb-infected diabetic mice was significantly higher than that in Mtb-infected control mice and uninfected diabetic mice. To determine the cellular source of IL-6 in Mtb-infected T2DM mice, we first examined the leukocyte populations by flow cytometry. As shown in Table 1, the number of CD11c+MHCII+CD103+, CD11c+CD11b+MHCII+, F4/80+CD64+MHCII+, Ly6G+ neutrophils, and lymphocytes in the lungs of Mtb-infected T2DM mice at 1 month post-Mtb infection was significantly higher than that in Mtb-infected non-diabetic mice or uninfected T2DM. A similar increase was observed at 6 months p.i.
We next examined the phenotype of IL-6 producing pulmonary cells at 1 and 6 months p.i. There were no significant differences in the absolute numbers of IL-6-producing Ly6G+ neutrophils, B220+IgM+ B cells, CD3+NK1.1-T cells, or CD3-NK1.1+ NK cells (Fig 6C). However, the absolute number of IL-6-producing CD11c+MHCII+CD103+ and CD11c+CD11b+MHCII+ cells in the lungs of T2DM mice at 1 month after Mtb infection was significantly higher than that in the lungs of uninfected T2DM mice (Fig 6C) or Mtb-infected control mice (Fig 6C). As shown in Fig 6C, a similar increase in IL-6-producing CD11c+ cells were noted in the lungs of T2DM mice at 6 months post-Mtb infection. Although there was an increased frequency of F4/80+CD64+MHCII+IL-6+ cells in the lungs of Mtb-infected T2DM mice compared with those of uninfected T2DM mice at 6 months p.i. (Fig 6C), there was no significant difference between Mtb-infected non-diabetic control mice (Fig 6C). Overall, these results suggest that CD11c+ cells are the major source of IL-6 in Mtb-infected T2DM mice. To further confirm the cellular source of IL-6 at 6 months p.i., we isolated various cell populations from the pooled spleen, lymph node, and lung cell populations of Mtb-infected control and T2DM mice by magnetic sorting and measured IL-6 expression by real-time PCR. We found that CD11c+ cells are the major source of IL-6 (data shown in the Dryad Data Repository; doi:10.5061/dryad.qn42t).
The interaction between NK cells and macrophages is crucial for the initiation and amplification of early immune responses [19]. We found a significant increase in NK and CD11c+ cell numbers in the lungs of Mtb-infected T2DM mice (Table 1). To determine whether NK cells are involved in increased IL-6 production by CD11c+ cells, we first examined lung sections from Mtb-infected T2DM mice by confocal microscopy. Imaging results at 6 months p.i. indicated that more NK cells in Mtb-infected T2DM mice were in close proximity to IL-6-producing CD11c+ cells than in Mtb-infected control mice (Fig 7A and S1 Fig). More importantly, the result indicates that the marked increase in IL-6 production occurs in the region in which both NK cells and CD11c+ cells interact. We further determined whether the NK and CD11c+ cell interaction increases IL-6 production by lung mononuclear cells in Mtb-infected T2DM mice. Six months after Mtb infection, mononuclear cells were isolated from the lungs of T2DM and non-diabetic control mice and some cell populations were depleted of NK cells by magnetic separation. Lung mononuclear cells and NK cell-depleted lung mononuclear cells were cultured with γ-irradiated Mtb H37Rv (γ-Mtb). After 48 h, IL-6 levels in the culture supernatants were measured by ELISA and the phenotype of IL-6-producing cells was identified by flow cytometry. Stimulation with γ-Mtb significantly enhanced IL-6 production by pulmonary mononuclear cells from Mtb-infected T2DM mice (Fig 7B) when compared with those from Mtb-infected control mice (Fig 7B). However, depletion of NK cells from Mtb-infected T2DM pulmonary mononuclear cells led to a significant reduction in IL-6 levels (Fig 7B). Furthermore, we found that the frequency of IL-6+CD11c+MHCII+ and IL-6+CD11b+MHCII+ cells (Fig 7B) increased significantly after culture of Mtb-infected T2DM pulmonary mononuclear cells with γ-Mtb. However, depletion of NK cells from Mtb-infected T2DM pulmonary mononuclear cells resulted in a significant reduction in the frequency of IL-6+CD11c+MHCII+ cells (Fig 7B). By contrast, depletion of NK cells had no effect on the frequency of IL-6+CD11b+MHCII+ cells (Fig 7B). These results further confirm that CD11c+ cells are a major source of IL-6 and that NK cells from Mtb-infected T2DM mice increase IL-6 production by CD11c+MHCII+ cells.
NK cell-activating receptors play an important role in the development of diabetes [20]. We next examined expression of NK cell-activating receptors in Mtb-infected mice by flow cytometry. Lung CD3-NKp46+ NK cells from Mtb-infected T2DM mice expressed higher levels of NKG2D (Fig 7C) and DNAM-1 (Fig 7C) than those from Mtb-infected control mice. We then examined the possible role of these activating receptors in stimulating CD11c+ cells to produce IL-6. At 6 months p.i., lung mononuclear cells from Mtb-infected T2DM mice were cultured with γ-Mtb in the presence of blocking NKG2D or DNAM-1 mAbs or isotype-matched control antibodies. The frequency of IL-6-expressing CD11c+MHCII+ cells (Fig 7D) increased significantly after culture of Mtb-infected T2DM pulmonary mononuclear cells with γ-Mtb in the presence or absence of the isotype-matched control antibodies. Blocking the NKG2D (Fig 7D) or DNAM-1 (Fig 7D) interaction with CD11c+ cells led to a significant reduction in the frequency of IL-6+CD11c+ cells. Similarly, IL-6 levels in the culture supernatants of cells cultured with blocking NKG2D (Fig 7D) or DNAM-1 mAbs (Fig 7D) decreased significantly.
To further confirm the above findings, NK cells and CD11c+ cells were isolated from pooled splenic, lymph node, and lung cells from Mtb-infected control and T2DM mice by magnetic selection. Autologous NK cells and CD11c+ cells were cultured together at a ratio of 1:4 (1 NK and 4 CD11c+) in the presence or absence of γ-Mtb and with or without the isotype control or NKG2D or DNAM-1 blocking antibodies. After 48 h, the culture supernatants were collected and IL-6 levels were measured by ELISA. Culture of Mtb-infected control or Mtb-infected T2DM mouse NK cells with autologous CD11c+ cells in the absence of γ-Mtb did not induce IL-6 production. Culture of Mtb-infected control mouse NK cells with autologous CD11c+ cells in the presence of γ-Mtb resulted in 251.5 ± 65.1 pg/ml IL-6; this increased to 556.9 ± 52.5 pg/ml (p = 0.02) in NK cells and CD11c+ cells from Mtb-infected T2DM mice (Fig 7E). This increase in IL-6 production by CD11c+ cells was inhibited by anti-DNAM-1 and anti-NKG2D blocking antibodies (Fig 7E).
The above result indicated that the interaction between NK and CD11c+ cells increases IL-6 production in Mtb-infected T2DM mice. Therefore, we next determined whether depletion of NK cells with a NK1.1 antibody affected survival, cytokine production, or the bacterial burden in acute T2DM mice. One month after STZ/NA treatment, acutely diabetic mice were infected with50–100 CFU Mtb. At 6 months p.i., mice were treated with an anti-NK1.1 mAb, an isotype-matched control mAb, or PBS (Fig 8A). As shown in Fig 8B, 65% (p<0.05) of Mtb-infected T2DM mice that received the isotype-matched control mAb or PBS died within 2 months. By contrast, all Mtb-infected T2DM mice that received the anti-NK1.1 mAb survived. Anti-NK1.1 mAb treatment also reduced the bacterial burden in the lungs (Fig 8C), spleen (1.5 ± 0.86 × 104 vs. 10.5 ± 0.86 × 104 CFU; p = 0.0003), and liver (1.75 ± 0.25×103 vs. 2.75 ± 0.25 × 103 CFU; p = 0.03) by a marginal, but statistically significant, amount. Real-time PCR analysis of lung samples indicated that anti-NK1.1 mAb treatment of acutely diabetic Mtb-infected mice was associated with significantly lower levels of IL-6, TNF-α, IL-10, and TGF-β (Fig 8D) expression than those observed in mice treated with the isotype-matched control mAb. Histological examination of lung tissue indicated a similar degree of inflammation in PBS-treated and isotype-matched control antibody-treated mice with acute T2DM mice and Mtb (Fig 8E). By contrast, anti-NK1.1 mAb treatment significantly reduced inflammation in the lungs of Mtb-infected mice with acute T2DM (Fig 8E).
To determine the relevance of the above findings with respect to human Mtb infection, we obtained blood from pulmonary tuberculosis patients with or without T2DM and first determined the frequency of pro-inflammatory cytokine-producing T cells by flow cytometry. The frequency of IFN-γ- (1.5-fold, p = 0.026, Fig 9A) and IL-2- (2.1-fold, p = 0.002, Fig 9A) producing cells was significantly higher in the blood of diabetic than in that of non-diabetic pulmonary tuberculosis patients. By contrast, there were no significant differences in the frequency of TNF-α- and IL-17-producing cells between the two groups. We also cultured whole blood in the presence of 10 μg/ml purified protein derivative (PPD). After 18 h, the frequency of IFN-γ-, IL-2-, TNF-α-, and IL-17-producing cells was determined by flow cytometry. As shown in Fig 9A and 9B, PPD significantly induced expression of IFN-γ, IL-2, TNF-α, and IL-17A. The frequency of IFN-γ- (1.7-fold, p = 0.0002; Fig 9A), IL-2- (2.1-fold, p = 0.0001; Fig 9A and 9B), TNF-α- (1.6-fold, p = 0.0005; Fig 9A), and IL-17A- (2-fold, p = 0.0004; Fig 9A) producing cells was significantly higher in diabetic than in non-diabetic pulmonary TB patients. We also examined whether neutralizing the IL-6 receptor affected PPD-induced changes in the frequency of IFN-γ-, IL-2-, TNF-α-, and IL-17A-producing cells in pulmonary TB patients with T2DM. As shown in Fig 9B, the anti-IL-6 antibody significantly reduced the frequency of IFN-γ- (1.7-fold, p = 0.0005), IL-2- (2.2-fold, p = 0.009), TNF-α- (6.6-fold, p = 0.0005), and IL-17A- (3.3-fold, p = 0.0005) producing cells when compared with the isotype-matched control antibody. However, the anti-IL-6 antibody had no effect on the frequency of cytokine-producing cells in healthy volunteers (Fig 9C).
In this study, we investigated the immune response of mice to Mtb infection following the induction of T2DM. Diabetic mice were found to have increased lung bacterial burden and mortality compared to non-diabetic controls. Alveolar macrophages from T2DM mice were more permissive to Mtb growth ex vivo compared to non-diabetic controls, indicating an impairment of innate antimicrobial function. Multiplex cytokine and chemokine data and real-time PCR analysis of Mtb-infected T2DM lungs also demonstrated significantly higher expression of genes encoding pro- and anti-inflammatory cytokines than in lungs from uninfected T2DM and infected non-diabetic control mice. Neutralization of IL-6 increased the survival of all Mtb-infected T2DM mice, reduced the bacterial burden, and reduced cytokine production. We found that CD11c+ cells were the major source of IL-6 in Mtb-infected T2DM mice. IL-6 production by CD11c+ cells was further enhanced by NK cells. We also found that IL-6 enhances inflammatory cytokine production in pulmonary tuberculosis patients with T2DM. Limited information is available about protective immune responses in type 2 diabetic hosts during Mtb infection. Our results suggest that the NK-CD11c+ cell interaction increases IL-6 production, which drives the pathological immune response and reduces survival of Mtb-infected T2DM mice.
Chemically induced type 1 diabetes (T1DM) models are widely used in research, and mice with STZ-induced insulin deficiency develop susceptibility to TB [8]. Approximately 90% of people living with diabetes have T2DM, making it by far the most prevalent form of diabetes in TB patients [21]. It is therefore important to employ T2DM models for mechanistic studies of this dual burden. Different approaches have been used to model T2DM in animals, including the combination of high-fat diet with low to intermediate doses of STZ [22]. Some approaches require relatively long periods of time to exhibit all of the major features of the disease and some fail to replicate persistent hyperglycemia [23]. This is an important limitation since the vascular and renal complications of diabetes only develop after prolonged hyperglycemia, and related mechanisms may drive at least some features of diabetic immunopathy [24]. In the current study, we induced T2DM in mice using STZ and NA, which resulted in sustained hyperglycemia for up to 8 months. After 6 months, T2DM mice showed significantly elevated blood cholesterol and triglyceride levels. These findings suggest that our model can be used to investigate TB defenses in the setting of acute or chronic T2DM.
Experimental Mtb infection has been investigated in other animal models of T2DM. Sugawara et al. [10] reported that GK/Jcl rats, which spontaneously develop T2DM, have a higher bacterial load and more severe immune pathology than non-diabetic Wistar rats at 5–12 weeks after infection with Mtb Kurono. Expression of mRNA encoding several cytokines, including IFN-γ, TNF-β, and IL-1β, was higher in Wistar rats at 1 and 3 weeks p.i., but higher in GK/Jcl rats by 12 weeks p.i. Podell et al. [11] reported increased TB susceptibility in guinea pigs with T2DM induced by low dose STZ plus a high-fat, high-sucrose diet. The guinea pig TB phenotype was characterized by an increased bacterial burden, more severe immune pathology, increased cytokine expression, and increased mortality. The results of TB studies using the rat and guinea pig T2DM models, the mouse T2DM model presented here, and the mouse T1DM model [8] differ in some aspects; however, all show impaired control of Mtb replication, more severe immune pathology, and increased expression of multiple cytokines. TB in diabetic people is associated with increased sputum smear positivity at the time of diagnosis (a surrogate marker for higher bacterial burden), increased radiographic severity of disease, increased mortality (reflecting increased immune pathology), and increased expression of several nominally protective cytokines [8,9,25,26]. These similarities support the relevance of our T2DM mouse model to the interaction between TB and diabetes in people.
We found that C57BL/6 mice with STZ/NA-induced T2DM exhibited increased expression of pro- and anti-inflammatory cytokine genes, including IL-6, after Mtb infection. IL-6 is a pleiotropic cytokine that has both protective and pathogenic roles in diabetes [14]. There are conflicting reports about the role of IL-6 in Mtb infection [15,16]. IL-6 contributes to vaccine-induced protective immunity in mice [17], and IL-6 knockout mice are highly susceptible to Mtb [15]. By contrast, IL-6 produced by macrophages infected with Mtb in vitro selectively inhibits macrophage responses to IFNγ, thereby contributing to the survival of mycobacteria [27]. Correspondingly, IL-6 neutralization increases IFN-γ-mediated killing of intracellular Mtb by inducing autophagy [28]. In human TB patients, IL-6 is implicated in the pathogenesis of the immune reconstitution inflammatory syndrome [29,30]
We found that in vivo neutralization of IL-6 conferred a survival benefit and was associated with a reduced bacterial burden in the lung and pro-inflammatory cytokine expression in Mtb-infected mice with acute or chronic T2DM; however, it did not alter the hyperglycemic status of the mice. We also found that CD4+ cells from TB patients with T2DM produced significantly elevated levels of Th1 and Th17 cytokines, and that this was inhibited by neutralizing IL-6. A variety of hematopoietic and non-hematopoietic cells produce IL-6, and we found that CD11c+ cells were the major source of IL-6 in Mtb-infected T2DM mice. IL-6 production was enhanced by the interaction between NKG2D and DNAM-1 and the corresponding ligands on CD11c+ cells. These results suggest that IL-6 produced during the NK-CD11c+ interaction may be a key factor that drives the damaging immune pathology in diabetic hosts infected by TB. Of note, increased expression of activation markers by unstimulated myeloid cells from diabetic individuals has been documented [31].
NK cells are prominent components of the innate immune system and play a central role in resistance to microbial pathogens. NK cells protect against viruses, parasites, and bacteria, including Mtb, by destroying infected cells and secreting cytokines that shape the adaptive immune response [32–34]. NK cells interact with antigen-presenting cells and T cells, and are involved in one or more stages of immune-mediated attack. Abnormalities in the frequency and activity of NK cells have been described both in animal models and in patients with diabetes. By contrast, depletion of NK cells prevents the development of diabetes [35,36]. In TB patients with T2DM, altered CD8+ and NK cell function leads to enhanced pathology [37].
In conclusion, we found that hyperactive NK cells interact with CD11c+ cells to amplify the IL-6-mediated inflammatory immune response in TB. Our data suggest that NK cell-mediated IL-6 production by CD11c+ cells is responsible for driving hyperinflammation and increased mortality in T2DM mice infected with Mtb. The mechanism that underlies NK cell hyperactivation in T2DM mice remains unknown, but NK cell activation was recently reported to be an upstream event in T2DM pathogenesis [38]. The NK-CD11c+ axis and the IL-6 pathway may be promising new targets for host-directed therapies aimed at reducing the severity of immune pathology, which drives morbidity and mortality in those infected by TB.
Specific pathogen-free female wild-type C57BL/6 mice (4 to 6 weeks old) were purchased from Jackson Laboratory and housed at the animal facility at the University of Texas Health Science Center at Tyler. All animal experiments were approved by the Institutional Animal Care and Use Committee of the University of Texas Health Science Center at Tyler.
Blood was obtained from 20 healthy controls, 20 pulmonary tuberculosis patients, and 20 pulmonary tuberculosis patients with type T2DM. All subjects were HIV-seronegative with culture-proven pulmonary tuberculosis who had received anti-tuberculosis therapy for < 1 week. Acid-fast stains of sputum samples were positive for all patients. Type 2 diabetes patients had HbA1c levels > 6.5% and random blood glucose levels > 200 mg/dl.
All human studies were approved by the Institutional Review Board of the National Institute of Research in Tuberculosis (NCT01154959), Chennai, India, and informed written consent was obtained from all participants. All animal studies were approved by the Institutional Animal Care and Use Committee of the University of Texas Health Science Center at Tyler (Protocol #533). All animal procedures involving the care and use of mice were undertaken in accordance with the guidelines of the NIH/OLAW (Office of Laboratory Animal Welfare).
PE-conjugated anti-IL-6 (eBioscience), FITC-conjugated anti-CD3 (Tonbo Biosciences), PE-conjugated anti-NKp46 (BioLegend), PE-cy7 anti-NKG2D (eBioscience), and APC-anti-DNAM-1 (eBioscience) were used for flow cytometry. Antibodies used for the in vivo neutralization experiments were purchased from BioXcell (mouse anti-IL-6 [MP5-20F3], anti-NK1.1, and isotype controls [rat IgG1 and mouse IgG2a antibodies]). The NKG2D and DNAM-1 blocking antibodies were obtained from eBioscience. STZ and NA were obtained from Sigma Chemicals. Anti-CD11c, anti-IL-6, anti-NK1.1, secondary antibodies (goat anti-hamster IgG-Alexa 568, donkey anti-rat-Alexa 488, and goat anti-rabbit-Alexa 647), and DAPI were obtained from Life Technologies and used for confocal microscopy. γ-irradiated Mtb H37Rv (γ-Mtb) was obtained from BEI Resources. Highly purified mouse recombinant IL-6 with a specific activity of 1 108 units/mg was purchased from BioLegend (Bedford, MA).
T2DM was induced by combined administration of STZ and NA. STZ was dissolved in a 50 mM citric acid buffer and administered (180 mg/kg of body weight) intraperitoneally three times, with an interval of 10 days between doses. NA was dissolved in saline and administered intraperitoneally (60 mg/kg of body weight) 15 min before STZ. Mice were fasted for 16 h before the STZ and NA injections. Blood glucose was measured using a glucometer at weekly intervals for up to 8 months. Mice were considered diabetic if their blood glucose was > 250 mg/dl. Control mouse blood glucose levels were always between 80 and 100 mg/dl.
Serum insulin levels in fasting (16 h) control and diabetic mice were measured using a Mercodia Ultrasensitive Insulin ELISA Kit (Mercodia AB Uppsala, Sweden).
Serum free fatty acids, cholesterol, and triglyceride levels were measured using either a fluorometric or colorimetric assay (Cayman Chemicals, USA), according to the manufacturer’s instructions.
OGTTs were performed in control and diabetic mice after fasting (16 h). A glucose solution (2.0 g/kg) was given orally. Blood glucose concentrations were measured 30 min before and 15, 30, 60, and 120 min after administration.
Murine AMs were isolated from control and T2DM mice by bronchoalveolar lavage at 1 and 6 months post-induction of diabetes. Briefly, mice were euthanized by CO2 asphyxiation. The trachea was then cannulated following a midline neck incision, and the lungs were lavaged five times with 1.0 ml of ice cold PBS. Alveolar cells were separated from the lavage fluid by centrifugation at 1800 RPM for 10 min. Alveolar cells were plated (on plastic) to permit adherence of alveolar macrophages and subsequent removal of non-adherent NKT and T lymphocytes by washing three times with normal PBS. Adherent cells were resuspended in RPMI-1640. Highly purified AMs were used to determine whether reduced growth of Mtb was due to dysfunctional AMs. Alveolar cells were plated in 96-well tissue culture plates at a density of 2 ×105/100 μl/well, incubated for 24 h at 37°C in 5% CO2, and washed three times with antibiotic-free RPMI-1640. Around 98% of the cells expressed CD11c, as determined by flow cytometry. AMs were infected with Mtb H37Rv at a MOI of 1:2.5 (2.5 Mtb to 1 macrophage). This MOI was based on the viability of AMs at different MOIs for up to 7 days p.i. More than 90% of AMs were viable at this MOI. Cells were incubated for 2 h at 37°C in a humidified 5% CO2 atmosphere, washed to remove extracellular bacilli, and cultured in RPMI 1640 containing 10% heat-inactivated human serum. To quantify the intracellular growth of Mtb H37Rv, infected AMs were cultured for 5 days after which the supernatant was aspirated and AMs were lysed. Bacterial suspensions in cell lysates were ultrasonically dispersed, serially diluted, and plated in triplicate on 7H10 agar. The number of colonies was counted after 3 weeks.
Before infecting mice with Mtb H37Rv, bacteria were grown in liquid medium to the mid-log phase and then frozen in aliquots at -70°C. Bacterial counts were determined by plating on 7H10 agar supplemented with oleic albumin dextrose catalase (OADC). For infection, bacterial stocks were diluted in 10 ml of normal saline (to 0.5 ×106 CFU [colony forming units]/ml, 1 ×106 CFU/ml, 2 ×106 CFU/ml, and 4 × 106 CFU/ml) and placed in a nebulizer within an aerosol exposure chamber custom made by the University of Wisconsin. In preliminary studies, groups of three mice were exposed to the aerosol at each concentration for 15 min. After 24 h, mice were euthanized and homogenized lungs were plated on 7H10 agar plates supplemented with OADC. CFUs were counted after 14–22 days of incubation at 37°C. The concentration that deposited ~75–100 bacteria in the lung during aerosol infection was used for further studies.
For some experiments, mice were treated with neutralizing anti-IL-6 antibodies. For the first set of experiments, conducted 1 month after the induction of T2DM, mice were challenged with aerosolized Mtb. At 6 months p.i., mice received 0.3 mg of anti-IL-6 mAb (BioXcell) or isotype-matched control Ab (rat IgG1) intravenously every 4 days for up to 2 months. For the next set of experiments, conducted at 6 months after the induction of T2DM, mice were infected with Mtb H37Rv. On day “0” of infection, mice received 0.3 mg of anti-IL-6 or isotype control Ab intravenously every 4 days for up to 2 months.
One month after the induction of T2DM, mice were infected with aerosolized Mtb. On day “0” of infection, mice received 0.3 mg of anti-NK1.1 mAb or isotype control Ab intravenously every 4 days for up to 1 month. Previously, we used the same anti-NK1.1 (PK136) antibody to deplete 95% of the NK1.1 cells [34].
Lungs from Mtb-infected mice were mechanically homogenized and filtered through a 70 μm cell strainer. Cells were washed twice, and mononuclear cells were isolated from lung single-cell suspensions using a one-step gradient separation method (GE Healthcare), according to the manufacturer’s instructions. Some mononuclear cell populations were depleted of NK cells by positive selection using antibody-labeled magnetic beads (Miltenyi Biotech), according to the manufacturer’s instructions. Lung mononuclear cells and NK cell-depleted mononuclear cells were seeded in 12-well culture plates (1 × 106 cells per well) and cultured with γ-Mtb. After 48 h, cells were harvested and IL-6 expression was measured by intracellular flow cytometry. Supernatants were also collected to measure IL-6 levels by ELISA.
Single-cell suspensions of pooled lung, lymph node, and murine splenocytes were prepared. CD11c+ cells were isolated by positive selection with magnetic beads (Miltenyi Biotec) conjugated to anti-CD11c; positively selected cells comprised > 96% CD11c+ cells, as measured by flow cytometry. NK cells were isolated by negative selection using kits obtained from Miltenyi Biotec. Isolated cells comprised > 97% CD3-NK1.1+ cells, as measured by flow cytometry.
After the mice were euthanized, the lungs were perfused with 5 ml of PBS via the right ventricle. Lungs were mechanically homogenized and passed through a 70 μm cell strainer. The remaining red blood cells were lysed using BD Pharm Lyse (BD Biosciences). Surface staining to identify leukocyte populations was then performed. For IL-6 intracellular staining, cells (106/ml) were suspended in RPMI 1640 containing 10% FBS and brefeldin A (5 μg/ml), placed in 24-well culture plates, and stimulated with LPS (1 μg/ml), and incubated for a further 4 h at 37°C to allow intracellular accumulation of cytokines. Cells were then permeabilized with 0.1% saponin and stained for intracellular IL-6. The cells were washed, resuspended in FACS buffer, and analyzed by flow cytometry using a FACS Calibur flow cytometer.
Whole blood was diluted 1:1 with RPMI-1640 medium and cultured in 12-well plates (0.5 × 106 cells/well) in RPMI 1640 containing penicillin/streptomycin (100 U/100 mg/ml), L-glutamine (2 mM), and HEPES (10 mM) (Invitrogen, Carlsbad, CA) in the presence or absence of a PPD (10 μg/ml) at 37°C in a humidified 5% CO2 atmosphere. In some cases, an IL-6R neutralizing Ab (2.5 μg/ml) was added to the culture. Brefeldin A (10 μg/ml) was added to the cultures 2 h before the termination of the cultures (total culture time, 6 h). Cells were washed, and red blood cells were lysed with lysis buffer. The cells were fixed using Cytofix/Cytoperm buffer (BD Biosciences) and cryopreserved at -80°C. Intracellular staining for IFN-γ, TNF-α, IL-2, and IL-17A was performed as previously described [9].
Mouse multiplex ELISA kits (23-Plex kits, Bio-Rad) were used to measure chemokine and cytokine levels, according to the manufacturer’s instructions.
Total RNA was extracted from lung leukocytes or lung tissue as described previously [39]. Total RNA was reverse transcribed using the Clone AMV First-Strand cDNA synthesis kit (Life Technologies). Real-time PCR was performed using the Quantitect SYBR Green PCR kit (Qiagen) in a sealed 96-well microtiter plate (Applied Biosystems) on a spectrofluorometric thermal cycler (7700 PRISM; Applied Biosystems). PCR reactions were performed in triplicate as follows: 95°C for 10 min, followed by 45 cycles of 95°C for 15 s, 60°C for 30 s, and 72°C for 30 s. All samples were normalized to the amount of β-actin/GAPDH transcript present in each sample. The primers used in the study are listed in Table 2.
Lungs were inflated and fixed in 10% neutral buffered formalin (v/v) for 24 h. Tissue sections were stained with hematoxylin and eosin. A semi-quantitative analysis was performed using a score from 0 (no inflammation) to 4 (severe inflammation) for each of the following criteria: alveolar wall inflammation, alveoli destruction, leukocyte infiltration, and perivascular inflammation. Immunostaining of thin paraffin-fixed lung sections was performed using antibodies against IL-6 according to the manufacturer’s instructions (Novus Biologicals, USA). Unstained sections of formalin-fixed lung tissue from paraffin blocks were first deparaffinized and subjected to antigen retrieval in a citrate buffer at 95°C, as previously described. Endogenous peroxidase activity was blocked by addition of 3% H2O2 in methanol. Slides were incubated in 3% BSA in TBS for 2 min, after which primary antibodies were added at predetermined dilutions in TBS-Tween + 1% BSA (1:100) for 1 h at 25°C. Sections were then washed three times in TBS-T for 15 min. The IL-6 antigen was detected by IHC and the DAB (DAKO) chromogen, as previously described. Lung inflammation [40,41] and immunohistochemical readouts were independently assessed by two investigators as previously described [42]. The H-score was determined according to the method described by Pirker et al. Briefly, the percentage of cells with different staining intensities was determined by visual assessment and assigned a score (1+ for light staining, 2+ for intermediate staining, and 3+ for dark staining) using the ImageJ IHC profiler. The H-score was calculated using the formula 1 × (% of 1 + cells) + 2 × (% of 2 + cells) + 3 × (% of 3 + cells).
Confocal microscopy was performed to colocalize IL-6-producing CD11c+ and NK1.1+ cells in lung sections. Nonspecific binding was blocked with 1% goat serum in PBS for 30 min. The slides were then incubated at 4°C overnight with hamster monoclonal anti-CD11c (Abcam), rabbit polyclonal anti-NK 1.1 (Bioss), and rat monoclonal anti-IL-6 (Novus Bio) antibodies. Subsequently, the slides were washed thoroughly using 1 × PBS. Then, cells were stained with the respective secondary antibodies (goat anti-hamster IgG-Alexa 568, goat anti-rabbit-Alexa 647, or donkey anti-rat-Alexa 488; Life Technologies), washed with PBS, and mounted with Prolong Gold antifade reagent containing DAPI (Life Technologies, USA). The slides were examined and analyzed under a laser-scanning confocal microscope (Zeiss LSM 510 Meta laser-scanning confocal microscope).
The efficacy of in vivo IL-6 neutralization was assessed in a IL-6 bioassay using the 8G2 cell line (an IL-6-dependent murine hybridoma cell line LS132.8G2) instead of the B9 cell line as previously described [43].
Data analyses were performed using GRAPHPAD PRISM (GraphPad Software, Inc., La Jolla, CA). The results are expressed as the mean ± SE. For normally distributed data, comparisons between groups were performed using a paired or unpaired t-test and ANOVA, as appropriate. Statistically significant differences between two clinical groups were analyzed using the non-parametric Mann–Whitney U-test. Data are deposited in the Dryad Data Repository: (doi:10.5061/dryad.qn42t)[44].
|
10.1371/journal.pgen.1001382 | ESR1 Is Co-Expressed with Closely Adjacent Uncharacterised Genes Spanning a Breast Cancer Susceptibility Locus at 6q25.1 | Approximately 80% of human breast carcinomas present as oestrogen receptor α-positive (ER+ve) disease, and ER status is a critical factor in treatment decision-making. Recently, single nucleotide polymorphisms (SNPs) in the region immediately upstream of the ER gene (ESR1) on 6q25.1 have been associated with breast cancer risk. Our investigation of factors associated with the level of expression of ESR1 in ER+ve tumours has revealed unexpected associations between genes in this region and ESR1 expression that are important to consider in studies of the genetic causes of breast cancer risk. RNA from tumour biopsies taken from 104 postmenopausal women before and after 2 weeks treatment with an aromatase (oestrogen synthase) inhibitor was analyzed on Illumina 48K microarrays. Multiple-testing corrected Spearman correlation revealed that three previously uncharacterized open reading frames (ORFs) located immediately upstream of ESR1, C6ORF96, C6ORF97, and C6ORF211 were highly correlated with ESR1 (Rs = 0.67, 0.64, and 0.55 respectively, FDR<1×10−7). Publicly available datasets confirmed this relationship in other groups of ER+ve tumours. DNA copy number changes did not account for the correlations. The correlations were maintained in cultured cells. An ERα antagonist did not affect the ORFs' expression or their correlation with ESR1, suggesting their transcriptional co-activation is not directly mediated by ERα. siRNA inhibition of C6ORF211 suppressed proliferation in MCF7 cells, and C6ORF211 positively correlated with a proliferation metagene in tumours. In contrast, C6ORF97 expression correlated negatively with the metagene and predicted for improved disease-free survival in a tamoxifen-treated published dataset, independently of ESR1. Our observations suggest that some of the biological effects previously attributed to ER could be mediated and/or modified by these co-expressed genes. The co-expression and function of these genes may be important influences on the recently identified relationship between SNPs in this region and breast cancer risk.
| Recent genome-wide analysis has revealed that the way in which genes are arranged on chromosomes and the conformation of these chromosomes are crucial for the regulation of gene expression. Reflecting this arrangement, clusters of genes which are regulated together have been discovered. We have identified a previously unreported transcriptional activity hub spanning ESR1, the gene encoding the important breast cancer biomarker oestrogen receptor. Genetic variants immediately upstream of ESR1 have recently been linked to breast cancer risk. We found that three open reading frames within this region are tightly co-expressed with ESR1. We investigated the function of these genes and discovered that one of these co-expressed genes, C6ORF211, affects proliferation in cultured cells and is correlated with proliferation in breast tumours. Another of the genes, C6ORF97, is negatively correlated with proliferation in breast tumours and predicts for outcome on the anti-oestrogen drug tamoxifen. These findings suggest that the genes could contribute to the phenotype associated with oestrogen-receptor positivity. In addition, they may be involved in the mechanism by which genetic variation in this region of the genome contributes to breast cancer susceptibility.
| Breast cancer is the most common malignancy in women, accounting for more than 400,000 deaths per year worldwide [1]. Approximately 80% of human breast carcinomas present as oestrogen receptor α-positive (ER+ve) disease and ER status is arguably the most clinically important biological factor in all oncology [2]. The major molecular features of breast cancer segregate differentially between ER+ve and ER−ve tumours [3], [4]. Tumours which express ERα have been termed luminal type [3], [5] and are associated with response to antioestrogen therapy and improved survival, although the mechanisms by which oestrogen receptor dictates tumour status are poorly understood.
Recent genome wide studies have identified SNPs around C6ORF97, an open reading frame (ORF) immediately upstream of the gene encoding ER (ESR1) to be associated with increased risk of breast cancer. Zheng et al. found that heterozygosity at rs2046210, a SNP in the region between C6ORF97 and ESR1, increased breast cancer risk by an odds ratio of 1.59 in a Chinese population and that this risk was also present in a European population, albeit to a weaker extent [6]. Easton and colleagues confirmed the risk associated with this SNP and reported an at least partly independent risk associated with a second adjacent SNP (rs3757318) in intron 7 of C6ORF97 [7]. Using ancestry-shift refinement mapping, Stacey et al. closed in on the identification of the pathogenic variant and found that the risk allele of a novel SNP in this region (rs77275268), disrupts a partially methylated CpG sequence within a known CTCF binding site [8]. More recently, two further studies have confirmed an association with the region [9], [10]. Our studies have revealed unexpected relationships in the expression patterns in breast carcinomas between ESR1, C6ORF97 and the two genes immediately upstream (C6ORF211 and C6ORF96 [RMND1]).
Oestrogenic ligands, predominantly oestradiol, are the key mitogens for ER+ve breast cancer. In recent years, high throughput genomic technologies have revealed significant numbers of genes that are expressed in response to oestradiol stimulation in vitro [11]–[13] and downregulated in response to oestrogen deprivation in tumours [14]–[16]. Similarly, the transcriptional targets of ERα have been characterised in detail using genome wide chromatin interaction mapping in MCF7 cells [17], [18]. Key oestrogen responsive genes such as TFF1 and GREB1 have been shown to be highly responsive to oestradiol stimulation in cell culture models through the binding of ERα to their promoters [19], [20]. Additional genes have been found in hierarchical clustering analyses of ER+ve and ER−ve tumours as part of the so-called “luminal epithelial” gene set characterized by the expression of genes typically expressed in the cells that line the ducts of normal mammary glands including GATA3 and FOXA1 [12]. However, the correlates of ESR1 within an exclusively ER+ve group and the inherent heterogeneity within an exclusively ER+ subgroup remain poorly defined.
Modern, non-steroidal aromatase inhibitors (AIs) are widely used, effective treatments for ER+ve breast cancer [21], [22] and are also excellent pharmacological probes for oestrogen-dependent processes in vivo because of their specificity and highly effective suppression of oestrogen synthesis. In this study, we found that the expression of genes in the region immediately upstream of ESR1 associate strongly with ESR1 expression in ER+ve primary breast cancers before and after AI treatment and uncover evidence that these associations might impact upon the biological and clinical importance of ERα.
To investigate correlates of ESR1, expression profiles were derived from pairs of 14-guage core cut biopsies before and after 2 weeks' treatment with 1 mg/d anastrozole, an AI, from 104 patients with ER+ve primary breast cancer [23]. Genes whose expression correlated with expression of ESR1 levels pre-treatment were identified (Spearman corrected for multiple testing at false discovery rate <1×10−7, Table 1 pre-treatment). The mRNA species most highly correlated with ESR1 were chromosome 6 ORF 97 (C6ORF97, Rs = 0.67) (Figure 1a), followed by C6ORF211. Other notable inclusions amongst the top 20 most correlated genes included well-established ER-associated genes such as FOXA1, MYB and GATA3, plus C6ORF96, also known as RMND1 (Required for Meiotic Nuclear Division 1 homolog). The mean pre-treatment expression of the three ORFs was highly correlated with ESR1 (Rs = 0.70, Figure 1b). After 2 weeks' AI treatment, the top three genes correlating with ESR1 were C6ORF96, C6ORF97 and C6ORF211 (Rs>0.7 for all, Table 1 two weeks post-treatment). These three ORFs are all located less than 0.5 MB upstream of the ESR1 start site on the q arm of chromosome 6 (Figure 1e). The expression of other genes located within a 50 MB region surrounding ESR1 were not correlated with ESR1 expression (Rs<0.25) (Table S1).
The correlation was present in all of five published microarray data sets of ER+ve breast cancer in which the C6orfs were included on the array (Table 2). The expression of the three ORFs was lower in ER−ve than ER+ve tumours in the Wang dataset [24] (p = 0.002). No significant correlation was found in the ER−ve subgroup of this dataset. This may be a characteristic of ER−ve tumours or, alternatively, the measurement error associated with low levels of ESR1 transcript could preclude detection of a significant correlation in microarray data.
Amplification of the ESR1 locus has been reported inconsistently [25], [26]. To determine whether the ESR1/C6ORFs correlation may be the result of underlying genomic co-amplification or deletion events, copy number (CN) status of ESR1 and the C6orfs was examined using array CGH analysis (resolution 40–60 kb) [27] on DNA from the 44 tumour samples from which adequate further tissue was available. One tumour was shown to be amplified and eight showed gains at ESR1, C6ORF96, C6ORF97 and C6ORF211, while four showed losses at all four loci. One was measured as having loss of C6ORF96, C6ORF211 and part of C6ORF97. While there was some correlation between CN and transcription of the four genes (Figure S1), CN alterations did not explain the correlation between ESR1 and the C6orfs. In fact, when samples with identified CN changes were removed from the dataset, the correlation between ESR1 and mean C6orf expression levels strengthened rather than weakened (Rs = 0.83) (Figure 1c), suggesting that transcriptional co-regulation rather than genomic changes is more likely to underlie ESR1/C6ORF co-expression.
To assess whether the correlation in ESR1/C6ORF expression seen in pre-treatment biopsies is reflected in a concordant change in expression of these genes upon treatment, the relationship between the magnitude of change of each of these genes was investigated. Change in expression of ESR1 induced by aromatase inhibitor treatment over 2 weeks was strongly correlated with change in the C6orfs (Rs = 0.70) (Figure 1d). Given that this short duration of treatment, which has no measurable impact on cellularity or tumour size, is unlikely to facilitate DNA copy number changes throughout the sample this supports the probability that the co-regulation of these genes is at a transcriptional level.
To determine whether the ESR1/C6ORF correlations were maintained in vitro, transcript levels of ERα and the three C6orfs were measured in oestrogen-deprived MCF7 cells and lapatinib-treated BT-474 cells over a 48- and 96-hour period, respectively. These treatments are both known to have significant effects on the expression of ESR1. Lapatinib has been shown to increase ERα in BT-474 cells [28], [29], potentially via loss of Akt and de-repression of FOXO3a. This provides a useful model for manipulation to test the correlation between ESR1 and the C6orfs in vitro. Conversely, absence of oestradiol leads to a short-term reduction in ER expression [30]. Expression of all four genes followed a similar time-course of expression and was highly correlated (Figure 2a and 2b).
ICI 182,780 (ICI) is a steroidal pure anti-oestrogen which causes ERα expression to be suppressed and downregulated [31], [32]. Treatment of MCF7 cells with ICI did not affect ORF expression or their correlation with ESR1 (Figure 2c). To confirm that the observed correlation was not being influenced by RNA transcribed prior to the addition of ICI, we also measured newly synthesised nascent RNA using PCR amplicons designed to cross an exon/intron boundary [33]. This analysis revealed that nascent transcripts for ESR1 and the C6orfs remained correlated in both the presence and absence of ICI. The observation that transcription of the genes remains strongly correlated in the presence of ICI suggests that transcriptional regulation by ERα is not the main driver of the ESR1/C6ORF co-expression.
The effect of reducing expression of each C6orf on cell proliferation was determined by transfecting siRNA SMARTPOOLs directed against each ORF into MCF7 cells. In cells grown in both E2-containing media and without E2, all three siRNAs reduced transcript levels of their target ORF to <30% of levels in cells transfected with the control non-targeting siRNA pool. Levels of ESR1, and the non-targeted ORFs were unaffected by the SMARTpool's (Figure S2) while ESR1-SMARTpool siRNA led to a reduction in levels of all three C6orfs (Figure S3). Immunoblotting with a polyclonal antibody raised against a polypeptide of the predicted product of C6ORF211 showed an 86% reduction at the protein level (Figure S4). Cells transfected with C6ORF211 siRNA showed a mean 36% reduction in cell number (p<0.0001) over four separate repeat experiments (Figure 3A). C6ORF211 knockdown had no effect on oestrogen-dependent proliferation (Figure 3B). Deconvolution of the SMARTPOOL showed that the four constituent siRNAs had a reproducible anti-proliferative effect when compared with scrambled control siRNA (Figure S5). No consistent alteration in proliferation was observed in cells transfected with siRNAs directed against C6ORF96 or C6ORF97 (Figure 3A).
To determine whether the association between C6ORF211 expression and proliferation seen in cultured cells is reflected in tumours, the relationship between C6ORF211 expression and a metagene composed of known proliferation-associated genes [34] was investigated. In baseline biopsies, levels of C6ORF211 but not ESR1 correlated significantly with proliferation (C6ORF211, Rs = 0.23, p = 0.04; ESR1, Rs = −0.01, p = ns) (Figure 4a), suggesting that C6ORF211 is more strongly associated with proliferation than ESR1. Correlations were also observed with a number of well-known proliferation-associated genes (Table S2). The relationship with proliferation was validated in data from a set of 354 ER+ve tumours [35] (Rs = 0.18, p = 0.0008) (Figure 4b) and the 209 ER+ve tumours from the Wang dataset [24] (Rs = 0.21, p = 0.004). Consistent with the findings in our own data, ESR1 was not significantly correlated with the proliferation metagene in either of the publicly available datasets (Loi, Rs = −0.03, p = ns; Wang, Rs = 0.02, p = ns). In contrast, C6ORF97 showed an independent, reproducible negative correlation with proliferation, in our dataset (Rs = −0.19, p = 0.05) and in the Loi (Rs = −0.22, p<0.0001) (Figure 4c) and ER+ve Wang datasets (Rs = −0.24, p = 0.0007).
To determine whether the relationship of the ORFs with proliferation is related to clinical outcome, recurrence free survival (RFS) in tamoxifen-treated patients was investigated for association with C6ORF97 and C6ORF211 expression. Despite the fact that in the Loi dataset ESR1 was not predictive of a significant difference in survival over 5 years [36], the lowest quartile of C6ORF97 was associated with significantly higher risk of recurrence (HR = 3.1, p = 0.0014) (Figure 4d). A similar trend was observed in untreated ER+ve tumours from the Wang dataset [24], although this was not significant (HR = 1.6, p = 0.16) (Figure S6a). C6ORF211 was not significantly associated with RFS (Figure S6b and S6c).
Our observation of a previously unreported transcriptional activity hub in the ESR1/C6ORF region of 6q25.1 has implications for recently identified associations between SNPs in the ESR1 region and breast cancer risk, as well as broader implications for the biological and clinical importance of ERα in established breast cancer. A number of SNPs, including rs3757318 within intron 7 of C6ORF97 [7], have been associated with breast cancer risk but the causative variant and mechanism remain undefined [6]–[10]. In an attempt to identify the pathogenic variant, Stacey and colleagues recently reported that GG homozygotes at rs9397435, located immediately downstream of C6ORF97, may express higher mean levels of ESR1 and that the rs9397435 [G] allele conferred significant risk of both hormone receptor positive and hormone receptor negative breast cancer in European and Taiwanese patients [8]. The association of a SNP in this region with ER expression is consistent with findings from our own group which have revealed that the variant genotype of SNP rs2046210 is associated with increased ERα expression as measured by immunohistochemistry [37]. The findings reported in this paper suggest that, due to their high degree of correlation with ESR1, levels of C6ORF97, C6ORF96 and C6ORF211 are also likely to correlate with the rs2046210 and rs9397435 genotype. Consequently, these genes may be involved in the pathogenesis of the variant SNPs and could explain the apparent anomaly noted by Stacey and colleagues in that the SNPs predispose to both hormone receptor positive and negative disease.
To date, analysis of ESR1 co-expressed genes has focussed on genes which are also downstream targets of the oestradiol-activated transcription factor activity of ERα such as FOXA1, TFF1 and GATA3. High throughput technologies have identified numerous classical and novel ERα-dependent targets of oestradiol [11], [17]. This association with the expression of ORFs has, however, not been reported other than by ourselves in abstract form [38].
The transcriptional correlation between ESR1 and these ORFs is highly statistically significant in our dataset, and in all of the publicly available datasets we examined. In our own patient cohort, we showed that two weeks' treatment with anastrozole induces a concomitant change in ESR1 and the C6orfs and a yet stronger correlation in their expression. Genomic amplification does not account for the correlations. This suggests that transcriptional co-regulation rather than major genomic rearrangement is likely to underlie their co-expression. To our knowledge, a transcriptional activity hub surrounding a major cancer related gene has not previously been identified.
The observation that the four transcripts remain correlated over a short timecourse in MCF7 and BT474 cells further supports the idea that the co-regulation of these genes is likely to occur at a transcriptional level. Given that ERα can autoregulate its own transcription by binding to an oestrogen responsive element (ERE) in its promoter [17], [39], the possibility that ERα could co-regulate itself and the C6orfs provides an attractive potential explanation for the correlation. We tested this hypothesis by treating MCF7 cells with the ERα antagonist ICI in the absence of E2. Our finding that the nascent transcripts of ESR1 and the three C6orfs remain correlated in the presence of ICI (Figure 2c) suggests that this co-regulation is not dependent on ERα transcriptional activation.
Regulation of the steady-state level of ERα in breast cancer cells is a complex phenomenon that includes transcriptional and post-transcriptional mechanisms [40]–[42]. C6ORF96 is transcribed off the opposite DNA strand to ESR1 (Figure 1e), therefore excluding the possibility that ESR1 and the ORFs are transcribed as a single polycistronic mRNA. Recent genome-wide mapping experiments have revealed the importance of chromatin organisation for gene expression [18], [43] suggesting that 3-D chromatin arrangement could represent a potential explanation for C6ORF/ESR1 co-expression. However, analysis of the data produced by Fullwood and colleagues [18] shows that C6ORF96, C6ORF97 and C6ORF211 are not encompassed by an ERα-bound long-range chromatin loop. Nevertheless, it remains possible that a loop driven by an alternative transcription factor could explain the transcriptional activity in this area.
At the nucleotide level, all three ORFs show some homology with ESR1, suggesting they may have arisen from gene duplication events [44]. C6ORF97 encodes a 715 amino acid coiled-coil domain-containing protein that is conserved across 11 species [45] while C6ORF211 is a member of the UPF0364 protein family of unknown function and is also conserved across multiple species [45]. Confocal analysis revealed that the protein encoded by C6ORF211 was expressed mainly in the cytoplasm and did not co-localize with ER (Figure S7). In a proteomic screen it has been found to interact with SAP18, a Sin3A-associated cell growth inhibiting protein [46].
This reported interaction with a growth inhibitory protein could explain our observation that knockdown of C6ORF211 induces suppression of proliferation in cultured cells. This association is mirrored in tumours, where a proliferation metagene correlates significantly with C6ORF211. Conversely, C6ORF97 expression correlates negatively with expression of the proliferation metagene and high C6ORF97 predicts for improved disease-free survival in a tamoxifen-treated published dataset, independently of ESR1 (Figure 4d). As high ESR1 has previously been shown to be associated with improved outcome on endocrine therapy [47], this raises the possibility that, given the observed correlation of C6ORF97 with ESR1, some of this association with outcome could be attributable to C6ORF97.
The high degree of correlation between ESR1 and the C6orfs has significant potential implications for our interpretation of ER levels and therapy of ER+ve breast cancers. As a transducer of mitogenic oestrogen signalling, disruption of ER represents a key target of therapies for ER+ve breast cancer, including tamoxifen and fulvestrant. Our data shows that C6ORF211 and C6ORF97 may contribute to the proliferative phenotype of ER+ve tumours, yet these proteins are unlikely to be affected by therapies targeted directly at ERα. Consequently, these proteins may represent potential targets for synergistic therapies in patients with high levels of C6orf expression or targets for breast cancer prevention. In addition, along with further research these relationships could shed light on recent associations between breast cancer risk and SNPs in the region.
Core-cut tumor biopsies (14-gauge) were obtained from 112 postmenopausal women with stage I to IIIB ER+ early breast cancer before and after two-weeks' anastrozole treatment in a neoadjuvant trial [23]. This study received approval from an institutional review board at each site and was conducted in accordance with the 1964 Declaration of Helsinki [48] and International Conference on Harmonization/Good Clinical Practice guidelines. Written informed consent was obtained from each patient before participation. Tissue was stored in RNAlater at −20°C. Two 4 µm sections from the core were stained with hematoxylin and eosin to confirm the presence of cancerous tissue and the histopathology and six 8 µm sections were retained for microarray CGH analysis (see below). Total RNA was extracted using RNeasy Mini kits (Qiagen, Sussex, UK). RNA quality was checked using an Agilent Bioanalyser (Santa Clara, CA, USA): samples with RNA integrity values of less than 5 were excluded from further analysis. ER status and Ki67 values by immunohistochemistry were already available [23].
RNA amplification, labelling and hybridization on HumanWG-6 v2 Expression BeadChips were performed according to the manufacturer's instructions (http://www.illumina.com) at a single Illumina BeadStation facility. Tumor RNA of sufficient quality and quantity was available to generate expression data from 104 pre-treatment biopsies. Data was extracted using BeadStudio software and normalized with variance-stabilizing transformation (VST) and Robust Spline Normalisation method (RSN) in the Lumi package [49]. Probes that were not detected in any samples (detection p value >1%) were discarded from further analysis.
Multiple correlation analysis was performed in BRB-Array Tools (http://linus.nci.nih.gov/BRB-ArrayTools.html). A statistical significance level for each gene for testing the hypothesis that the Spearman's correlation between expression of ESR1 and other genes was zero was calculated and p-values were then used in a multivariate permutation test [50] from which false discovery rates were computed. Other statistical analyses were performed in SPSS for Windows (SPSS Inc., Chicago, IL), S-PLUS (TIBCO Software Inc., Palo Alto, CA) and Graphpad Prism (Graphpad Software Inc., La Jolla, CA).
Multivariable analysis was performed in a forward stepwise fashion, the most significant additional variable (satisfying p<0.05) being added at each stage. Cases with missing values for any of the variables in the model were excluded from analysis.
For analysis of the breast cancer datasets from public resources the publicly available normalised gene expression data and clinical data were retrieved from Gene Expression Omnibus (http://www.ncbi.nlm.nih.gov/geo/) (‘Wang’ dataset [24], n = 286;GEO, accession number GSE2034) or obtained from the authors (‘Loi’ dataset [35], n = 354 tamoxifen-treated tumours composed of GEO, accession numbers GSE9195, GSE6532 and GSE2990; combined normalised dataset received courtesy of Dr Christos Sotiriou). Correlations between ESR1 and the C6orfs in the ‘Miller’ [51] (n = 251), ‘TransBig’ (n = 198) [52] and Huang [53] (n = 23 cell lines) were calculated using the correlation analysis tool in Oncomine (http://www.oncomine.org).
Data from the 72 genes comprising the proliferation metagene was retrieved from tumours from the Wang and Loi datasets and proliferation metagene scores were calculated as described previously [54]. Spearman correlation between the proliferation metagene and ESR1 and the C6orfs was calculated in Graphpad Prism. Survival analysis was carried out in these datasets using the quartiled expression of the C6orfs and the endpoints of recurrence free survival or time to relapse, according to the original publication.
Five 8 µm sections from frozen core biopsies were mounted onto Superfrost glass slides, stained with nuclear fast red, and microdissected with a sterile needle under a stereomicroscope to obtain a percentage of tumor cells >75% as described previously [55]. Genomic DNA was extracted as described previously [55]. The concentration of the DNA was measured with Picogreen according to the manufacturer's instructions (Invitrogen).
The 32K bacterial artificial chromosome (BAC) re-array collection (CHORI) tiling path aCGH platform used for this study was constructed in the Breakthrough Breast Cancer Research Centre [55]. DNA labelling, array hybridisations, image acquisition and filtering were performed as described in Natrajan et al. [56]. Data were smoothed using the circular binary segmentation (cbs) algorithm [27]. A categorical analysis was applied to the BACs after classifying them as representing gain, loss or no-change according to their smoothed Log2 ratio values as defined [56].
MCF7 cells were routinely maintained in phenol red free RPMI1640 (Invitrogen, Paisley, UK) supplemented with 10% foetal bovine serum and oestradiol (1 nM). Cells were passaged weekly and medium replenished every 48–72 hours. In the case of BT474, cell monolayers were cultured in phenol red containing medium supplemented with 10% foetal bovine serum. Cell lines were shown to be free of mycoplasma by routine testing.
Total RNA from treated MCF7 and BT-474 cells was extracted using the RNeasy Mini Kit (Qiagen) according to the manufacturer's instructions. All RNA quantification was performed using the Agilent 2100 Bioanalyzer with RNA Nano LabChip Kits (Agilent Technologies, Wokingham, Berkshire, UK). RNA was reverse transcribed using SuperScript III (Invitrogen), and random primers. Twenty nanograms of resulting cDNA of each sample was analyzed in triplicates by qRT-PCR using the ABI Perkin-Elmer Prism 7900HT Sequence detection system (Applied Biosystems). Taqman gene expression assays (Applied Biosystems) were used to quantitate processed transcripts of ESR1 (Hs01046818_m1), C6ORF96 (Hs00215537_m1), C6ORF97 (Hs01563344_m1), C6ORF211 (Hs00226188_m1), which were normalized to two housekeeping genes, FKBP15 (Hs00391480_m1) and TBP (Hs00427620_m1). These housekeepers were selected from a previously published list of appropriate reference genes for breast cancer [57]. Custom assays using primers designed to span intron-exon boundaries were used to measure nascent RNA (Table S3). Gene expression was quantified using a standard curve generated from serial dilutions of reference cDNA from a pooled breast cancer cell line RNA.
Cell monolayers were washed with cold PBS twice and collected by scraping. Cell pellets were lysed in extraction buffer, resolved by SDS-PAGE and transferred to nitrocellulose membranes as described previously [30]. Membranes were blocked and probed with a polyclonal antibody directed against the predicted peptide (amino acids 368–382) of C6orf211 (Eurogentec, Southampton, UK) and anti β-actin (Sigma-Aldrich, Poole, UK) using the methods described previously [58]. Quantification of immunoblots was performed using the NIH ImageJ software, and immunoblots were normalized to actin.
Cells were grown on glass coverslips in standard growth medium. Cells were fixed and incubated in the presence of primary antibodies as described previously [58]. Coverslips were washed with PBS and cells were incubated in the presence of appropriate Alexa Fluor 555 (red) or Alexa Fluor 488 (green)-labeled secondary antibodies (Molecular Probes, Invitrogen, Paisley, UK) diluted 1∶1000 for 1 hr. Cells were washed in PBS and nuclei (DNA) were counterstained with 4,6-diamidino-2-phenylindole (DAPI; Invitrogen) diluted 1∶10000. Coverslips were mounted onto glass slides using Vectashield mounting medium (Vector Laboratories, Peterborough, UK). Images were collected sequentially in three channels on a Zeiss LSM710 (Carl Zeiss Ltd, Welwyn Garden City, UK) laser scanning confocal microscope at the same magnification (×63 oil immersion objective).
Cell lines were depleted of steroids for 3 days by culturing in DCC-medium [59], seeded into 12-well plates at a density of 1×104 cells/well for MCF7 and 4×104 cells per well for BT474, monolayers were allowed to acclimatize for 24 h before treatment with drug combinations indicated for 6 d with daily changes. Cell number was determined using a Z1 Coulter Counter (Beckman Coulter). Results were confirmed in a minimum of three independent experiments, and each experiment was performed in triplicate.
Wt-MCF7 cells were stripped of steroid for 3 days as described above. Cells were subsequently seeded into 12 well plates at a density of 1×105 cells/well. After 24 hours monolayers were treated with vehicle (0.01% v/v ethanol), oestradiol (1 nM) or ICI182780 (10 nM) for the time intervals indicated. RNA was extracted using RNeasy Mini kit (Qiagen) and subjected to qRT-PCR as described.
Wt-MCF7 cells were stripped of steroid for 24 hours in DCC-medium. Stripped cells were subsequently seeded into 12 well plates at a density of 2×104 cells/well for proliferation assays or 1×105 cells/well for RNA expression analysis. After 24 hours monolayers were transfected with 100 nM of either siRNA against C6ORF96, C6ORF97, C6ORF211 or control siRNA using DharmaFECT 1 reagent (Dharmacon, Thermo Fisher Scientific, UK). Medium was then replenished the following day and cells were allowed to acclimatise for a further 24 hours. After 24 hours samples were taken for RNA expression analysis. For analysis of oestrogen-dependent proliferation, the monolayers were treated with increasing concentrations of oestradiol (0.01, 0.1 or 1 nM) 48 hours post transfection. The remaining plates were treated daily with the treatments indicated for 6 days before carrying out cell counts as described above.
|
10.1371/journal.pntd.0007178 | Spontaneous healing of Mycobacterium ulcerans disease in Australian patients | Mycobacterium ulcerans causes necrotising infections of skin and soft tissue mediated by the polyketide exotoxin mycolactone that causes cell apoptosis and immune suppression. It has been postulated that infection can be eradicated before the development of clinical lesions but spontaneous resolution of clinical lesions has been rarely described.
We report a case series of five Australian patients who achieved healing of small M. ulcerans lesions without antibiotics or surgery. The median age of patients was 47 years (IQR 30–68 years) and all patients had small ulcerative lesions (median size 144mm2, IQR 121-324mm2). The median duration of symptoms prior to diagnosis was 90 days (IQR 90–100 days) and the median time to heal from diagnosis without treatment was 68 days (IQR 63–105 days). No patients recurred after a median follow-up of 16.6 months (IQR 16.6–17.9 months) from the development of symptoms and no patients suffered long-term disability from the disease.
We have shown that healing without specific treatment can occur for small ulcerated M. ulcerans lesions suggesting that in selected cases a robust immune response alone can cure lesions. Further research is required to determine what lesion and host factors are associated with spontaneous healing, and whether observation alone is an effective and safe form of management for selected small M. ulcerans lesions.
| Mycobacterium ulcerans causes a destructive infection of skin and soft tissue known as Buruli ulcer that when severe can lead to serious long-term deformity and disability. It is currently not well documented whether people with Mycobacterium ulcerans disease can cure themselves without treatment. In our study we describe five people with small ulcers who cured their disease without specific medical or surgical treatment. This suggests that a proportion of people can develop an immune response sufficient enough to eradicate the disease without the help of medical intervention. This is an important step, as recognition of this possibility provides important further insights into the human immune response against the disease. It also opens the possibility to further studies that may determine characteristics of the organism and hosts that favour spontaneous healing of lesions. This knowledge may in turn improve efforts to prevent and control the disease which are currently lacking.
| Mycobacterium ulcerans causes a necrotising infection of skin and soft tissue known as Buruli ulcer (BU). If untreated it usually progresses, can result in major tissue destruction and be complicated by bone or joint infection.[1] In severe cases it may require plastic and reconstructive surgery and result in long-term disability.[2] The pathogenesis of M. ulcerans is mediated by a plasmid produced polyketide exotoxin called mycolactone which causes tissue destruction by inducing cell apoptosis.[3] It also allows persisting infection to develop by inhibiting dendritic cell function and secondarily T-cell activation,[4,5] as well as reducing the function of monocytes and macrophages by inhibiting cytokine production including tissue necrosis factors and gamma interferon.[6,7]
It has been postulated that infection can be eradicated before the development of clinical lesions,[8] and a partial protective effect of BCG in humans has been reported.[9] If true, these observations suggest that the hosts immune response can be protective against the development of BU, thought to be mediated via a protective T-helper-1 (TH1) cell mediated immune response.[10] Spontaneous resolution without medical or surgical treatment of clinical lesions in humans has been rarely reported.[11–13] Furthermore, in one of these studies involving five lesions from Africa the lesions were not bacteriologically confirmed to be M. ulcerans,[12] and in another study of a single lesion from Australia the lesion was surgically excised.[11] Recently Marion et al reported a case from Benin where a small nodular M. ulcerans lesion healed without medical or surgical intervention, as well as a small group of patients with active large M. ulcerans lesions who had separate scars suggestive of previously spontaneously healed large M. ulcerans lesions.[13] It is also unknown how often spontaneous resolution occurs and the factors associated with it. Therefore treatment is recommended for all M. ulcerans lesions.[14,15] The recommended first-line treatment is combination antibiotics for 8 weeks which is highly effective.[16,17] Wide surgical excision without antibiotics can be performed, with cure rates of greater than 90% if reserved for selected lesions with no risk factors for recurrence.[14,18] In a study from Africa, local heat application without antibiotics achieved high initial wound healing rates, but 18% of patients developed a recurrent lesion within 2 years.[19]
The endemic region of Victoria, Australia, is facing a worsening epidemic of M. ulcerans disease, with control efforts hampered by the limited understanding of transmission mechanisms to humans as well as the risk and mechanisms of disease development following exposure and infection.[20] Identifying that some patients can heal their disease without treatment, and the study of the factors that allowed them to do so, may provide insights that could aid the improved control of M. ulcerans disease. In this paper we report a case series of five Australian patients who achieved healing of their confirmed M. ulcerans lesions without recommended antibiotic regimens or surgery.
This was an observational study of routinely collected data from a clinical cohort of M. ulcerans patients managed at Barwon Health as previously described.[21] All patients were from the M. ulcerans endemic region of the Mornington and Bellarine Peninsulas in Victoria, Australia.[22] They were all diagnosed in 2017 on the basis of a positive IS2404 PCR for M. ulcerans.[23] The size of the lesion was determined by measuring with a ruler the diameter of induration in millimetres and calculating the surface area in millimetres squared. Patients were followed up on a 2 to 4 week basis until wound healing was achieved, and then at the end of the study period. Data on the type and frequency of wound dressings was not collected, although due to the small size of lesions, wound dressings were frequently not administered.
The five cases of M. ulcerans disease all occurred in adults as single small ulcerative lesions ranging from 16 to 858 mm2 in size (median size 144mm2, IQR 121-324mm2). (Table 1). The median age of patients was 47 years (IQR 30–68 years) and there were 3 males and two females. No patients were known to be immune suppressed or have diabetes. HIV testing was not performed. The median duration of symptoms prior to diagnosis was 90 days (IQR 90–100 days). In one case an acid-fast bacilli (AFB) stain and culture for M. ulcerans were also positive.
All patients gave informed oral consent to be managed with observation only and ethics approval for the study was provided by the Barwon Health Ethics Committee. All data were analysed anonymously.
Patient # 2 had an incisional biopsy but no other specific treatment. No other patients received recommended antibiotics or surgical treatment due to patient choice—in all 5 cases due to a reluctance to risk the toxicity of antibiotics or to undergo surgery in view of the small size of their M. ulcerans lesion. No patients were given heat treatment. The median time to heal from diagnosis was 68 days (IQR 63–105 days). (Figs 1 and 2) No patients had recurred after a median follow-up of 16.6 months (IQR 16.6–17.9 months) from the development of symptoms. No patients suffered long-term disability from the disease.
One patient in the cohort, a 37 year old male diagnosed in 2017 by IS2404 PCR after 120 days of symptoms, was initially managed with observation alone but was changed to active treatment after 49 days of observation following an increase in size of his lesion from 154 mm2 to 340 mm2. He was subsequently cured with 4 weeks of rifampicin and clarithromycin antibiotic treatment combined with a surgical curette and did not suffer any long-term disability.
This case series demonstrates that there are a proportion of patients with confirmed small ulcerative M. ulcerans lesions that spontaneously heal without specific antibiotic or surgical treatment. In our case series it is likely that all patients have been cured of their disease as they were followed for at least 14 months from the development of symptoms without evidence of relapse. We have previously demonstrated in Australian patients that disease relapse is rare more than 12 months following diagnosis and treatment.[24]
This suggests that in selected patients, the development of host immunity following the development of clinical disease may be effective in curing lesions. This presumably results from the host’s immune protection overcoming the immune suppressive effects of the mycolactone. This may relate to the development of a more robust TH1 immune response. The importance of the TH1 immune response in combating M. ulcerans disease is suggested by the fact that the expression level of gamma-interferon is inversely correlated with the severity of M. ulcerans lesions,[25] and gamma-interferon knockout mice developed more severe M. ulcerans disease with a greater numbers of organisms.[10]
It is notable that all our patients had symptoms for at least 84 days prior to diagnosis yet the lesions had remained small (<900mm2). This suggests lesions that have not progressed significantly in the first three months are exhibiting a degree of immune control that may allow spontaneous healing to occur. In addition, small lesions may have a lower number of organisms with less mycolactone production to inhibit the immune system, favouring the host’s immune response against the organism’s persistence. Observed factors in our cases series that may favour spontaneous resolution include small lesion size after at least approximately 90 days of symptoms, the lack of associated co-morbidities such as diabetes or malignancy that may impair the host’s immune response, and ulcerative lesions which allow the discharge of necrotic material that may contain live organisms and mycolactone. Our study is limited by the lack of further immunological testing of the host and biological testing of isolates and therefore we suggest further research be performed to examine host and pathogen factors associated with spontaneous healing of M. ulcerans disease. This will hopefully further enhance the understanding of human immune function against the organism which may in turn allow improved treatment of the disease. Furthermore, it may provide insights that allow the development of interventions that prevent disease post exposure, such as vaccination, an area for which the current lack of knowledge hampers disease control efforts.[20]
It is important to acknowledge that we have not performed a controlled trial comparing observation alone to active treatment of small ulcerative M. ulcerans lesions and therefore we cannot make conclusions about the safety and effectiveness of this approach as a mode of management. It is also important to understand that all patients had very small lesions that were not at risk of severe complications or disability without specific treatment—for larger lesions immediate antibiotic treatment is important to achieve optimal outcomes, and although spontaneous healing may be possible in severe lesions over a lengthy period, in most patients irreversible physical impairment occurs as a consequence.[13] Nevertheless, the recognition that some small lesions can resolve spontaneously suggests that further studies could be performed to determine the true prevalence of spontaneously healing small M. ulcerans lesions and whether there is the potential for treating clinicians to safely and effectively employ close observation for similar small lesions, rather than immediate antibiotic or surgical treatment. Additionally, it would be important to determine what lesion and host factors favour this approach. Observation alone has the advantage of avoiding the potential toxicity of antibiotic treatment which results in serious adverse effects in more than 20% of treated Australian patients.[26] Although surgery alone can be an effective option for small lesions[18] this usually involves a financial cost and is not always easily accessible. Importantly, observation alone as a mode of management for small lesions would need to be evaluated in settings with lower resources and more isolated populations where close monitoring may be less feasible, increasing the risk of undetected disease progression.
In conclusion, healing without specific treatment can occur for some small ulcerated M. ulcerans lesions in Australian patients suggesting that in selected cases a robust immune response alone can cure lesions. Further research is required to determine what lesion and host factors are associated with spontaneous healing, and whether observation alone is an effective and safe form of management for selected small M. ulcerans lesions.
|
10.1371/journal.pcbi.0030108 | A Mass Conserved Reaction–Diffusion System Captures Properties of Cell Polarity | Cell polarity is a general cellular process that can be seen in various cell types such as migrating neutrophils and Dictyostelium cells. The Rho small GTP(guanosine 5′-tri phosphate)ases have been shown to regulate cell polarity; however, its mechanism of emergence has yet to be clarified. We first developed a reaction–diffusion model of the Rho GTPases, which exhibits switch-like reversible response to a gradient of extracellular signals, exclusive accumulation of Cdc42 and Rac, or RhoA at the maximal or minimal intensity of the signal, respectively, and tracking of changes of a signal gradient by the polarized peak. The previous cell polarity models proposed by Subramanian and Narang show similar behaviors to our Rho GTPase model, despite the difference in molecular networks. This led us to compare these models, and we found that these models commonly share instability and a mass conservation of components. Based on these common properties, we developed conceptual models of a mass conserved reaction–diffusion system with diffusion–driven instability. These conceptual models retained similar behaviors of cell polarity in the Rho GTPase model. Using these models, we numerically and analytically found that multiple polarized peaks are unstable, resulting in a single stable peak (uniqueness of axis), and that sensitivity toward changes of a signal gradient is specifically restricted at the polarized peak (localized sensitivity). Although molecular networks may differ from one cell type to another, the behaviors of cell polarity in migrating cells seem similar, suggesting that there should be a fundamental principle. Thus, we propose that a mass conserved reaction–diffusion system with diffusion-driven instability is one of such principles of cell polarity.
| Eukaryotic cells such as neutrophils and Dictyostelium cells respond to temporal and spatial gradients of extracellular signals with directional movements. In a migrating cell, specific molecular events take place at the front and back edges. The spatially distinctive molecular accumulation inside cells is known as cell polarity. Despite numerous experimental and theoretical studies, its mechanism of emergence has yet to be clarified. We first developed a mathematical model of the Rho small GTP(guanosine 5′-tri phosphate)ases that cooperatively regulate cell polarity, and showed that the model generates specific spatial accumulation of the molecules. Based on our Rho GTPases model and other models, we further established a conceptual model, a mass conserved reaction–diffusion system, and showed that diffusion-driven instability and a mass conservation of molecules that have active and inactive states are sufficient for polarity formation. We numerically and analytically found that molecular accumulations at multiple sites are unstable, resulting in a single stable front–back axis, and that sensitivity toward changes of a signal gradient is specifically restricted at the front of a polarized cell. We propose that a mass conserved reaction–diffusion system is one of the fundamental principles of cell polarity.
| Eukaryotic cells such as neutrophils and Dictyostelium cells respond to temporal and spatial gradients of extracellular signals with directional movements [1–6]. This process, known as chemotaxis, is a fundamental cellular process [5,7–9]. In a migrating cell, specific molecular events take place at the front and back edges [1,2,5,10]. The spatially distinctive molecular accumulation inside cells is known as cell polarity. The front–back polarity usually has one axis, and this uniqueness is an important property because a migrating cell with two fronts could not move effectively [11]. Another behavior of the front–back polarity is higher sensitivity of the front to a gradient of extracellular signals [10,12]. This would also be important because the direction of movement should be controlled at the front edge.
Many molecules that are involved in chemotaxis in mammalian cells have been identified [4,5]. Some molecules, including phosphoinositide 3-kinase (PI3K), phosphatidylinositol 3,4,5-triphosphate (PIP3), Cdc42, Rac, and F-actin, are specifically localized at the front, whereas others, including phosphatase and tensin homologue deleted on Chromosome 10 (PTEN) and RhoA, are at the back of migrating cells [1,4,10,13–15]. The Rho family of small GTP(guanosine 5′-tri phosphate)ases in particular play a central role in chemotaxis and in establishing cell polarity [15–17]. However, the mechanism of generating spatial accumulation of the Rho GTPases in cell polarity has yet to be clarified.
Many mathematical models that account for gradient sensing and signal amplification in cell polarity have been proposed [12]. The local excitation and global inhibition model has been proposed to explain spatial gradient sensing [6,18]. Some models involve positive feedback loops for amplified accumulation of signaling molecules [19–23]. A reaction–diffusion model that includes local self-enhancement and long-range antagonistic effects has been proposed for directional sensitivity [24]. Most of the reported models of cell polarity, which involve the detailed parameters such as concentrations or rate constants, have been constructed with many parameters and equations. Although these detailed models are at least partially successful in reproducing experimental observations in cell polarity, the theoretical essence underlying cell polarity has not been explicitly demonstrated; thus, a simple conceptual model that can be used for analytical study is needed to extract common principles in cell polarity. Although the reported models consist of distinct molecular species or networks, it should be especially emphasized that many of them are able to exhibit similar behaviors of cell polarity regardless of their different frameworks. This fact indicates that a common principle should underlie the models, and a conceptual model is suitable for extracting common principles in cell polarity.
Because the Rho small GTPases are key regulators for cell polarity [16,17], we first developed a reaction–diffusion model of the Rho GTPases on the basis of an earlier model [25] to examine the spatial properties of the Rho GTPases. We found that the interaction of the Rho GTPases per se can generate specific spatial accumulation of the Rho GTPases, and that our model shows important behaviors of cell polarity. We also found that our model exhibits behaviors similar to the model by Narang and Subramanian [22,23], which is based on the molecular networks that are different from ours. This suggests that common principles should underlie both models. We found that a mass conservation of components and diffusion-driven instability are commonly conserved in the Narang and Subramanian models and in our model. Based on these common properties, we established conceptual models of a mass conserved reaction–diffusion system, and found that such properties can account for the critical behaviors of cell polarity. These results strongly suggest that a mass conservation of components with diffusion-driven instability is one of the fundamental principles of cell polarity.
We developed a reaction–diffusion model of the Rho GTPases (Rac, Cdc42, RhoA) on the basis of the earlier model of the Rho GTPase [25], which explains the temporal behaviors, to examine whether the interaction of the Rho GTPases can generate the spatial behaviors in the cell polarity of migrating cells. The Rho GTPases exhibit guanine nucleotide–binding activity and function as molecular switches, cycling between an inactive GDP(guanosine 5′-bis phosphate)-bound form and an active GTP-bound form. The Rho GTPases in active forms are located in the plasma membrane, and those in inactive forms are in the cytosol (Figure 1A) [26]. It is likely that molecules in the cytosol have larger diffusivity than those in the plasma membrane. According to some studies, Cdc42 activates Rac [27–29], and RhoA has mutual inhibitory interactions with Cdc42 and Rac [29–33]. In addition, Rac plays a dominant role in a positive feedback loop, which involves PI3K, PIP3, and F-actin [13,34–36]. Based on these experimental findings, we developed a diagram of the Rho GTPases interaction (Figure 1B). We assume that molecules of Rac, Cdc42, and RhoA are activated by guanine nucleotide exchange factors (GEFs; kai) and are inactivated by GTPase-activating proteins (GAPs; kii), and that interactions between molecules (kij) are additive to GEFs or GAPs. Some molecule–molecule interactions are stimulation dependent. Activations of molecules by the stimulation (ksi) are also assumed to be additive to GEFs. As in many previous models [18–23], we describe the spatial kinetics of molecules by simple diffusion equations. A recent study in which the diffusion coefficients of the Rho GTPases in the plasma membrane are determined [37] may support this assumption. The model of the interaction of the Rho GTPases is as follows (see also Materials and Methods):
where Rac, Cdc, and Rho with suffixes m and c denote the concentrations of Rac, Cdc42, and RhoA in the active state (membrane) and inactive state (cytosol), respectively. The numerical suffixes represent the following: 1, Rac; 2, Cdc42; and 3, RhoA. Dmi and Dci denote the diffusion coefficients of molecules in the active state and inactive state, respectively (Dmi < Dci). The position-dependent parameter, S, denotes the intensity of stimulation. Because the parameters have not been fully obtained experimentally, we set parameters to reproduce the behaviors of cell polarity (Figure 2), and further analyzed the generality of such behaviors in detail with conceptual models (see below).
The behaviors in the Rho GTPases model, such as switch-like reversible accumulation, uniqueness of axis, and sensing of the stimulation gradient by the polarized peak, were similar to those observed in the models of Subramanian and Narang [22,23]. Despite the differences of the molecular species and networks among the models, the similarity in behaviors among them raises the possibility that a common principle could underlie them. Therefore, we examined whether common properties can be seen.These models belong to reaction–diffusion systems with a periodic boundary condition and exhibit switch-like response, implying that instability is important for accumulation of the components. In addition, these models involve components whose masses are conserved. The total amount of phosphoinositides between the plasma membrane and the endoplasmic reticulum is conserved in Narang and Subramanian's model [22,23], and the total amounts of the Rho GTPase between the membrane (e.g., Rhom) and cytosol (e.g., Rhoc) are conserved in our Rho GTPases model. Based on these common properties, we derived a new concept (i.e., mass conserved reaction–diffusion system with diffusion-driven instability) and hypothesized that this system is a fundamental principle of cell polarity. We therefore developed a simple conceptual model with two components (u and v), which belong to a mass conserved reaction–diffusion system with instability, and examined whether the model can cause the behaviors of cell polarity to emerge sufficiently.
where a1 and a2 are parameters of the model. The position-dependent parameter, S, is intensity of stimulation (see Materials and Methods). The stability of the homogenous solution in this model depends on the value of S (see Materials and Methods).
We found that the conceptual model exhibits behaviors similar to the Rho GTPases model, such as switch-like reversible accumulation (Figure 3A–3C), uniqueness of axis (Figure 3D–3F), and sensing of the stimulation gradient by the polarized peak (Figure 3G–3I). This finding indicates that the conceptual model retains the essential behaviors in the Rho GTPases model. We further used this conceptual model to examine in detail two behaviors of cell polarity: uniqueness of axis and sensitivity of the polarized peak.
To better understand the results of the numerical simulations, we investigated the following model (see Equations 1a–1c), by analytical approximations:
where Du = αDv. Here, a1, a2, and α are parameters of the models. This model belongs to the mass conserved reaction–diffusion system, and is more advantageous for analytical examination. The homogenous solution was unstable regardless of the values of parameters in this model (see Materials and Methods), so this model did not show reversible accumulation. However, the model still retained the important properties such as uniqueness of axis and localization of sensitivity, so we can use this model to analytically examine whether these behaviors can emerge from a mass conserved reaction–diffusion system with instability.
In the following sections, we show that: (1) the model has one-peak stationary states, regardless of the system size (if not too small); (2) multiple-peak stationary states are unstable; and (3) the polarized peak moves depending on the gradient of the parameter value and the sensitivity is localized. Finally, (4), we verified our analyses by comparing analytical results with the values obtained by numerical simulations.
In this study, we used a mathematical model to clarify the role of the interaction between the Rho GTPases in cell polarity and developed a conceptual model of cell polarity to glean a theoretical understanding of the unique behaviors of cell polarity.
The Rho GTPases regulate the remodeling of the actin cytoskeleton via actin polymerization, depolymerization, and myosin activity [15,17,26], which ultimately establishes cell polarity. Then, what regulates the spatial activity of the Rho GTPases? The model proposed by Sakumura et al. indicates that the interaction of the Rho GTPases can regulate their own temporal activities [25]. We demonstrated that the interaction of the Rho GTPases can regulate their own spatial activities. In reality, the interaction of the Rho GTPases can provide more complicated temporal and spatial regulation of their activities. Further study, including the determination of kinetic parameters of the interaction, is necessary to develop a more realistic model of the Rho GTPases.
The activator–inhibitor model for pattern formation [38] and the local excitation and global inhibition model for directional sensing in chemotaxis [6] belong to conceptual models, rather than to detailed biological models. Such conceptual models with reduction of equations and parameters make analysis simpler and clearer. We identified a mass conserved reaction–diffusion system with instability as common properties between the cell polarity models. The model belonging to this system can sufficiently reproduce the important behaviors of cell polarity, such as uniqueness of axis and localization of sensitivity, and enabled us to theoretically understand such behaviors, which are difficult to examine without the models.
When interleukin-8, a chemoattractant, is applied simultaneously from two directions at a 45° angle, normal neutrophils choose one direction for migration instead of responding to both sources [11]. Neutrophils with multiple leading edges are rarely observed under normal conditions [39]. When HL60 cells are transfected with a dominant-negative Rho construct or treated with Rho-kinase inhibitors, many cells exhibit the multiple pseudopods, where, in some cells, protrusions gradually withdraw, leaving a single, prominent pseudopod [10]. In addition, inhibitions of PI3Ks cause HL-60 cells to form multiple pseudopods, which are weak and transient [39]. These results suggest the instability of multiple leading edges, which may make the front of a migrating cell single and stable. Chemotactic cells must have only one front–back axis because multiple fronts would prevent fine migration. Subramanian and Narang investigated the response of their model to two almost identical stimulations [23] and showed that only one of the two peaks that arise persists, which agrees with our results. Here we show that uniqueness of axis emerges from instability of multiple peak solutions in a mass conserved reaction–diffusion system (Figures 4C and 5).
In neutrophils [40], HL-60 cells [10], and Dictyostelium cells [41], the polarized cells respond to changes in direction of a gradient by performing U-turns rather than by simply reversing polarity. In addition, polarized migrating cells move forward without responding to the chemoattractant source near their rears [12]. These experimental findings indicate that the sensitivity to chemoattractants is localized at the leading edge of polarized cells. The localized sensitivity focuses the activity of the actin cytoskeleton at the leading edge, resulting in faster movement toward a chemoattractant source [12]. Few mathematical theories, however, have been proposed to explain the localization of sensitivity. Here we show that localization of sensitivity depends on the specific localization of a sensing window at the polarized peak in a mass conserved reaction–diffusion system (Figure 4D and 4E and Figure 6). It should be added that many other systems can also exhibit a localized sensing window.
Consider a molecule that satisfies the following conditions: (1) the molecule (X) has two states (Xm and Xc); (2) the total amount of X is conserved; and (3) the diffusion coefficient of Xc is larger than that of Xm. Two states of this molecule can be treated as components of a mass conserved system. Some kinds of small GTPases, such as those of the Rho family, have two forms, active and inactive forms; the Rho GTPases in the active forms are located in the membrane, and those in the inactive forms are in the cytosol [26]. Some enzymes involved in the cell polarity of chemotactic cells, such as PI3K and PTEN, are also reported to show a relationship between their activity and membrane binding [42–46]. Molecules in the cytosol may well diffuse faster than those in the plasma membrane. Thus, these molecules can be considered as components of mass conserved systems.
Chemotactic cells, such as Dictyostelium cells and neutrophils, polarize within a few minutes (30 s to 3 min) after they are exposed to chemoattractants [6,10,47]. Because the time scale of cell polarity is likely to be much shorter than that of gene expression and protein synthesis, we can assume that the masses of molecules are constant during the polarization of chemotaxis.
We numerically and analytically show that multiple-peak solutions are unstable. To facilitate an understanding of the physics of this instability, we attempt to give an intuitive physical explanation of the behavior of molecules in the case where there exist two peaks (Figure 8), just as in Figure 4C. We simplify the situation as follows. (1) There are two spaces (each space has one peak). (2) The molecules have two forms, u and v, which have small and large diffusivities, respectively. (3) No molecule is generated or degraded in the spaces. (4) The molecules move between spaces, mainly in v-form, depending on the concentration gradient of v-form molecules. (5) The u-form molecules convert v-form molecules into u-form, and this positive feedback is so strong that infusion of molecules into the space causes a decrease in v-form molecules. Here, consider that a few molecules move from one space (S1) to the other (S2). According to (5), because of the nature of the positive feedback, the number of v-form molecules in S2 declines as the total number of molecules in S2 increases. In turn, according to (4), the declining of the number of v-form molecules in S2 successively facilitates further transfer of v-form molecules from S1 to S2, resulting in the further increase of the total number of molecules in S2. This flux is never disrupted by the generation or degradation of molecules because of the mass conservation. Therefore, two peaks in such a system are unstable. The condition (5) seems critical for instability, and we analyzed such conditions mathematically elsewhere [48].
One of the most extensively studied reaction–diffusion models is the Turing model, in which robust spatial patterns, such as stripes or spots, emerge via a diffusion-driven instability [38,49]. An ordinary Turing pattern in 1-D space is stripes with an intrinsic scale length [50]. Mass conserved models also generate multiple peaks from the homogenous state during the early phase, which is explained by diffusion-driven instability. However, they exhibit characteristic transitions after the initial peaks arise: most peaks become smaller and eventually disappear, and only one peak persists (Figure 4A and 4B). Why is the behavior of the mass conserved system so different from that of ordinary Turing models? Consider a reaction–diffusion system with vast size (L → ∞) and interval I [x1, x2] within the system, where x1 and x2 are arbitrary but far apart. Can we predict what will happen to interval I? For an ordinary Turing model, the linearization analysis around the homogenous solution gives us sufficient information [50]. The mass conserved system is more complex, however, because the behavior differs between the case where the components flow into interval I versus the case where they flow out, and we cannot predict which will occur. This difference in predictability seems to be fundamentally linked to the different behavior and the specificity of the mass conserved system. Mass conserved models have multiple stationary states that are spatially homogenous or periodic, including a one-peak state and multiple-peak states. We show that the multiple-peak stationary states are unstable, resulting in a single stable peak.
In some reaction–diffusion systems, such as the activator–inhibitor model (or substrate–depleted model), high diffusivity of inhibitor (or substrate) make multiple peaks unstable, resulting in a single stable peak [51]. An intuitive explanation for this instability is that an inhibitor, which is generated at the peak, rapidly diffuses throughout the cell. In the cell where the inhibitor can be generated and degraded, the spread of the inhibitor requires the large diffusivity to overwhelm the inhibitor degradation. On the other hand, in a mass conserved reaction–diffusion system, where no molecule is generated or degraded, a peak takes up molecules from its surroundings to grow. That is, the growing peak inhibits the system not by spread of inhibitor but by deprivation of molecules. In this case, large diffusivity is not required because there is no competitor to overcome, such as generation of molecules, and the inhibition can eventually spread throughout the cell. Indeed, Equation 16b clearly indicates that any Dv can make multiple peaks unstable (μ > 0), at least in Model II.
It may be counterintuitive that any mass conserved system finally exhibits a one-peak pattern. For Model II, the final steady state was a one-peak solution regardless of the system size, even when L was infinitely large. But for Model I, the final steady state had two peaks when we set L = 80 (unpublished data). Some mass conserved models probably have a maximum size to have a unique peak. However, this maximum size is independent of the linearization analysis. The conditions for the uniqueness of concentration peak will be elucidated in future analyses.
Because properties observed in simple models are expected to be conserved in more detailed models, we assume that movement of molecules follows a simple diffusion equation in our conceptual model. Active transport systems, such as actomyosin- and microtubule-based active transports, regulate cell polarity in various cellular processes [52]. Such active transports are likely to follow the formation of intracellular asymmetry, which takes place under the resting condition where the cell polarity is yet to be generated. Under such conditions, the diffusion of the Rho small GTPases has been measured and shown to be approximated by an apparent simple diffusion, if viewed on the order of seconds or tens of seconds [37]. Because our concern in this study is the earlier asymmetry formation rather than the completion of the polarity, which includes active transport, we here assumed the simple diffusion of the Rho GTPases. However, we will readily incorporate the detailed mechanism of the transport system of the Rho–GTPases in a future model.
In this study, we focused on the stationary state, but not on the transient state, which involves adaptation in response to a transient signal [44,53–55]. Such properties, as well as high dimensionality and multiple components, should be incorporated into a future model.
Although our model is rather simple, it shows the important properties of cell polarity such as switch-like reversible response, uniqueness of axis, and localization of sensitivity. We further demonstrated that the instability of multiple-peak solutions and the specific localization of a sensing window at the polarized peak in a mass conserved reaction–diffusion system are responsible for uniqueness of axis and localization of sensitivity, respectively. One remarkable feature of a mass conserved reaction–diffusion system compared with other models is that a mass conserved reaction–diffusion system does not require strict assumptions for diffusion coefficients, such as smaller and much larger diffusivities of excitatory and inhibitory molecules, respectively, in the local excitation and global inhibition models [6,18], or an extraordinary large diffusivity of the inhibitor in the Gierer–Meinhardt model [51]. Since the Rho GTPases, PI3K, or PTEN have thus far not been demonstrated to involve such ad hoc assumptions of diffusivity, a mass conserved reaction–diffusion system is more likely to explain cell polarity where these molecules are involved, and to be adapted to a wide range of cell polarity. Taking into consideration that the Rho GTPases system satisfies conditions of a mass conserved reaction–diffusion system, it is likely that this system is one of the fundamental principles of cell polarity.
We considered a one-dimensional circular system with circumference L. The position is represented by x (−L/2 ≤ x ≤ L/2). We applied the periodic boundary condition, which is used in some models that explain cell polarity [23,24]. We used explicit difference methods to perform simulations. The difference intervals for calculations are shown in the following text.
The parameter values were set as follows: L = 10, Dmi = 0.04, Dci = 3 (i = 1, 2, 3), ks1= 1, ks2 = 1, ks3 = 1, ka1 = 0.2, ka2 = 0.2, ka3 = 0.2, ki1 = 0.4, ki2 = 0.2, ki3 = 0.2, k11 = 4, k12 = 3, k13 = 5, k23 = 6, k31 = 4, and k32 = 2. The difference intervals for calculations were taken to be Δt = 0.01 and Δx = 0.33. We set Xm(x) = 0.3, Xc(x) = 0.7 (X = Rac, Cdc, Rho), unless specified.
The parameter values were set as follows: a1 = 2.5, a2 = 0.7, Du = 0.01, Dv = 1, and L = 10. The difference intervals for calculations were taken to be Δt = 0.01 and Δx = 0.2. We set u = 1 and v = 1, unless specified.
The parameter values were set as follows: a1 = 0.5, a2 = 2.2, Du = 0.1, Dv = 1 or 2,
N̄ = 2. The difference intervals for calculations were taken to be Δt = 0.005 and Δx = 0.2.
Approximation of a one-peak solution. The computation was performed by setting L = 10 and Dv = 1 and taking the initial state as u = 1 and v = 1 with small perturbation (±0.01). The final profile (t = 200) is shown in the left panel of Figure 7A. The solid line indicates the profile of N, and the dashed line indicates P.
Instability of a two-peak solution. We examined the instability of two-peak solutions. First, we obtained a stable one-peak pattern in Model II (Equation 3) by taking the size to be L/2, where L = 20, 30, 40, and taking the initial state as u = 1 and v = 1 with small perturbation (±0.01). Because we applied the periodic boundary condition to this system, we could set the center of the concentration peak at x = 0 by translation. Next, by duplicating and coupling this profile (L/2), we obtained a new profile (L) with two peaks. We used this profile (L) with small perturbations (±0.01) as the initial state of the following simulation. All trials (Dv = 1, 2 and L = 20, 30, 40) showed instabilities of two-peak profiles, and we obtained the growth rate, μsml, from the change in peak height.
Movement of a polarized peak in response to the parameter gradient. We examined the response to the parameter gradient. First, we obtained a stable one-peak pattern in Model II (Equation 3) by setting L = 10 and taking the initial state as u =1 and v = 1. We set the center of the concentration peak at x = 0 by translation. Next, we substituted
(x) = a2{1 + (ɛ/2)sin[2π(x/L)]} for a2 in Equation 3. All trials (Dv = 1, 2 and ɛ = 0.02, 0.04, 0.06) showed movement of the polarized peaks, and we obtained the velocities, vsml, from the results.
In the homogenous stationary state, the Jacobian matrix for the reaction terms is given by
where fu and fv denotes the partial derivatives of f by u and v, respectively, at a homogenous stationary state of Equations 1a and Equations 1b. We obtained a condition for instability of the homogenous solution: (Dufv − Dvfu)/(DuDv) > 0. For example, the homogenous solution is stable with S = 0.2 in Model I, whereas unstable with S = 1, under the following conditions: a1 = 2.5, a2 = 0.7, Du = 0.01, Dv = 1, u + v = 2. The homogenous solution in Model II is always unstable.
Through the stability analysis using J, the range of wave numbers (kh) that have positive eigenvalues is obtained as 0 < kh < [(Dufv − Dvfu)/(DuDv)]1/2, and the wave number that has the largest eigenvalue and grows most rapidly from the homogenous state,
, is obtained as follows:
For Model I (Equation 2), we obtain
= 1.32 under the following conditions: a1 = 2.5, a2 = 0.7, Du = 0.01, Dv = 1, S = 1, u + v = 2.
Equation 6b has the same formulation as classical Newton mechanics. We define V(Ne) as
and Equation 6b implies
where E is a constant value, corresponding to period and total mass of Ne(x). The period λ and the average mass
N̄ = (1/λ)
Ne(x)dx satisfy the following equations:
where Nmin and Nmax are minimum and maximum levels of Ne(x), respectively, and are derived from V(Nmin) = V(Nmax) = E (0 < Nmin <
N̄ < Nmax). Equation 26a and Equation 26b give the relationship among Pe, λ, and
N̄, where
N̄ is straightforwardly derived from the initial condition of (u, v). For Model II (Equation 3),
and E can range between E* < E < 0 for Ne(x) to be a periodic solution. Here E* = −{[(Dv − Du)/DuDv][(
)/6
]}. As E becomes smaller (E → E*), the period λ converges to λmin, which is the shortest wavelength in the periodic solutions. As E becomes larger (E → 0), the period λ diverges. The solution of Ne(x) at E = 0 corresponds to the separatrix of Equation 6b, indicating an infinite period (λ → ∞). The explicit form of Ne(x) for E = 0 can be obtained by
which has the sole peak at x = xp and decays to zero as x → ±∞. For a sufficiently large system, Equation 28 is a good approximation of the solution for −L/2 < x < L/2. Equations 28 and 7 lead to Equations 8a and 8b, and 9a–9c.
If there is nontrivial (nμ(x), pμ(x)) that satisfies Equations 11a and 11b for μ with a positive real part, the solution is unstable. Note that (nμ0, pμ0) = (n0sech2(bx), − n0Pe/N0) satisfies Equations 11a and 11b for μ = 0 under periodic boundary conditions. Here n0 is an arbitrary factor, originated from the linearity of equations, and we can set n0 = 1. For μ with an absolute value near zero, we can obtain (nμ, pμ) by the expansion from (nμ0, pμ0) with regard to μ. To do this, we take nμ = nμ0 + μnμ1 + … and pμ = pμ0 + μpμ1 + … . In the first order of the expansion, (nμ1, pμ1) obeys the following equations:
Thus, pμ is immediately derived from Equation 29a as:
where C1 and C2 are integral constants. We can obtain nμ by solving by solving Equation 29b with substitution of pμ1. C2 is determined by the mathematical condition that nμ1 should be orthogonal to nμ0. In practice, C2 has little influence on (nμ, pu), and we set C2 = 0 in the analysis.
The linearized approximations of Equations 19a and Equations 19b are given as follows:
where hN(z) = ∂f*(Ne(z), Pe,a2)/∂N, hP(z) = ∂f*(Ne(z), Pe,a2)/∂P, ha(z) = ∂f*(Ne(z), Pe,a2)/∂
. Because t is no longer a variable, we set t = 0 without loss of generality and replace z with x in the following analysis. Equation 31a under the periodic boundary condition leads to pε as follows:
where C3 is an integral constant. By substituting, Equation 32 into Equation 31b, we obtain the following:
where Gn(x) and Ga(x) are defined by:
By solving Equation 33, we obtain nɛ:
where W1(x) and W2(x) are defined by
Considering the periodic boundary condition for nɛ(x), we obtain the following equation for sufficiently large L (solvable condition):
This leads to the velocity of the peak:
where Z is given as follows:
|
10.1371/journal.pntd.0004981 | Dengue Outbreak in Mombasa City, Kenya, 2013–2014: Entomologic Investigations | Dengue outbreaks were first reported in East Africa in the late 1970s to early 1980s including the 1982 outbreak on the Kenyan coast. In 2011, dengue outbreaks occurred in Mandera in northern Kenya and subsequently in Mombasa city along the Kenyan coast in 2013–2014. Following laboratory confirmation of dengue fever cases, an entomologic investigation was conducted to establish the mosquito species, and densities, causing the outbreak. Affected parts of the city were identified with the help of public health officials. Adult Ae. aegypti mosquitoes were collected using various tools, processed and screened for dengue virus (DENV) by cell culture and RT-PCR. All containers in every accessible house and compound within affected suburbs were inspected for immatures. A total of 2,065 Ae. aegypti adults were collected and 192 houses and 1,676 containers inspected. An overall house index of 22%, container index, 31.0% (indoor = 19; outdoor = 43) and Breteau index, 270.1, were observed, suggesting that the risk of dengue transmission was high. Overall, jerry cans were the most productive containers (18%), followed by drums (17%), buckets (16%), tires (14%) and tanks (10%). However, each site had specific most-productive container-types such as tanks (17%) in Kizingo; Drums in Nyali (30%) and Changamwe (33%), plastic basins (35%) in Nyali-B and plastic buckets (81%) in Ganjoni. We recommend that for effective control of the dengue vector in Mombasa city, all container types would be targeted. Measures would include proper covering of water storage containers and eliminating discarded containers outdoors through a public participatory environmental clean-up exercise. Providing reliable piped water to all households would minimize the need for water storage and reduce aquatic habitats. Isolation of DENV from male Ae. aegypti mosquitoes is a first observation in Kenya and provides further evidence that transovarial transmission may have a role in DENV circulation and/or maintenance in the environment.
| The first dengue outbreak in Kenya was reported in 1982 in the coastal region. This was followed almost 30 years later by the 2011 dengue outbreak in Mandera, northern Kenya and subsequently in Mombasa city in the coastal region (2013–2014). An entomologic investigation was conducted to establish the density of mosquito species causing the outbreak. Affected parts of Mombasa city were identified with the help of public health officials. Adult mosquitoes were collected using various tools, processed and screened for dengue virus. All indoor and outdoor containers in every accessible house and compound within affected suburbs were inspected for Ae. aegypti immatures. Over 2,000 adult Ae. aegypti mosquitoes were collected and 192 houses and 1,676 containers inspected for Ae. aegypti immatures. Although jerry cans (18%) and drums (17%) were the most productive, there was site-specificity in container type productivity. Therefore all containers would be targeted for effective control of the dengue vector, including proper covering of water storage containers and eliminating all discarded outdoor containers through a public environmental clean-up exercise. However, providing reliable piped water to all households in Mombasa city would be a long-term solution to reduce the risk of dengue transmission.
| Dengue virus (DENV) is a member of the genus flavivirus (family Flaviviridae) that is transmitted principally by Aedes aegypti mosquitoes in an Ae. aegypti-human cycle [1], sometimes resulting in epidemics. Although the presence of other Stegomyia spp. including Ae. simpsoni complex, Ae. africanus and Ae. vittatus in disease endemic areas in Kenya in sympatric/allopatric manner with Ae. aegypti have been documented [2], their role in DENV transmission remains unknown. Ae. aegypti is well distributed in the tropical and subtropical regions and readily develops in water held in artificial, often man-made, containers in and around human habitations [1], hence it is well adapted to domestic and urban settings [3]. The vectors live so close to humans on whom they preferentially feed, and DENV transmission can occur even when Ae. aegypti population densities are low [4,5]. The other known vector of DENV, Ae. albopictus, is not as associated with humans or their habitats as Ae. aegypti, and is responsible for dengue transmission mainly in Asia [6]. Whereas Ae. albopictus has recently been documented in parts of Central and West Africa including Equatorial Guinea, Cameroon, Gabon and Mozambique [7,8,9], surveillance conducted from 2007–2011 did not detect occurrence of Ae. albopictus in the coastal sites [2].
DENV causes dengue fever (DF), an acute mosquito-borne viral infection. DF is presently the world’s most important re-emerging arboviral disease with over 50% of the world’s population at risk of the disease and 50% residing in dengue endemic countries [10]. Approximately 3.6 billion people are currently at risk of dengue infection in over 100 countries of Asia, Americas and Africa [11]. It has been estimated that 390 million dengue infections occur worldwide annually [12].
The epidemiology and public health effect of dengue in Africa is poorly understood, although the vectors of DENV are widely distributed [13]. Dengue diagnosis is likely confounded by other diseases such as malaria and lack of laboratory diagnostic capability [14,15]. For example in regions endemic for malaria, 70% of febrile illnesses are treated as presumptive malaria or designated as having fever of unknown origin, hence the potential for misdiagnosing dengue as malaria. The distribution of dengue vectors and several other factors including rapid population growth, unplanned urbanization, and increased international travel increase the risk of dengue transmission [16]. Indeed, over the past 5 decades, dengue cases have been reported in many countries in sub-Saharan Africa [10] including European travelers returning from Tanzania, Zanzibar, the Comoros, Benin, Cape Verde, Gunea Bisau and Senegal [17–20] and the 2011–2013 outbreaks in Angola and Kenya [21,22]. This apparent emergence of DENV in most of Africa might be due to increased awareness of the disease, availability of better diagnostic tests, and improved access to specialized laboratory facilities [23].
Although Kokernot et al suggest that dengue existed in Africa as far back as 1926 [24], the first outbreak in eastern Africa was in Comoros in 1948 and later in 1983 and 1984 [25]. Between 1977 and 1979, a major outbreak caused by dengue 2 was reported in the Seychelles Islands affecting >75% of the population [26]. The Seychelles outbreak was followed by the first outbreak of DF, caused by dengue 2 virus, in Kenya along the coast in 1982 [27]. In 2004, DENV IgG antibodies were detected among humans in Malindi [28] suggesting continued circulation of this virus on the coast of Kenya. In 2007, a dengue 2 outbreak was reported in Gabon [29]. Also during this year, DENV antibodies were detected in 7 of 8 of the previous administrative provinces of Kenya (all except Nairobi) [30]. More recently in 2014, a dengue outbreak occurred in the United Republic of Tanzania [31], while between November 2011 and February 2014, an outbreak involving three DENV serotypes (1, 2 and 3) occurred in Mandera in northern Kenya [32] and in Mombasa city located on the coast of Kenya, where 58% of the suspected hospital cases (n = 267) were positive for dengue infection by RT-PCR [22]. Based on this data, we initiated entomologic surveillance activities to establish the mosquito species associated with the dengue cases, determine the densities of immature stages of the mosquitoes, identify the most productive container types in areas with ongoing DENV transmission and estimate the density of adult Ae. aegypti mosquitoes inside and around houses in areas with dengue cases. Data generated would lead to recommendations on control measures aimed at reducing the Ae. aegypti population densities [33] to stop further transmission.
Because dengue infection rates in Ae. aegypti are typically low [5] to base a surveillance and risk assessment program on entomological infection rates (EIR), this entomologic investigation was based largely on larval indices (i.e. container index (CI), house index (HI) and Breteau index (BI)). The Pan American Health Organization (PAHO) and World Health Organization (WHO) have described threshold levels for dengue transmission as low HI<0.1%, medium HI 0.1%–5% and high HI>5%. However, there is weak association between these indices and DENV transmission [34,35], hence they are limited to indicating vector presence or absence [36]. Because these threshold indices also differ from place to place [37], recent studies have recommended an area-specific re-evaluation of the utility of larval indices [38].
Entomological “outbreak” investigations were launched as a result of detected dengue transmission in humans in Mombasa city. Sampling locations were selected purposely based on the occurrence of laboratory-confirmed dengue cases. Within the locations, specific sites and households were selected randomly. Due to limited resources, investigations were conducted for a short period of time in only 7 out of the 9 affected locations.
Entomological sampling was conducted from 21 to 28 April 2013 and 28 November to 2 December 2013, while the 2014 sampling was from 4 to 15 March. Mosquitoes were collected indoors and outdoors, as larvae and adults, on a daily basis for the duration of each visit. One of the challenges of indoor sampling was the extreme difficulty in accessing some of the residences, especially in more affluent areas such as Nyali and Kizingo as the residents insisted on preserving their privacy.
Several sampling tools were employed to capture adult Ae. aegypti. Ten BG-Sentinel traps (Biogent), the current gold standard for adult Ae. aegypti surveillance, were set outside houses and monitored daily for three consecutive days in each site. Although resting boxes (RB) are not usual surveillance tools for Ae. aegypti, these devices were tested in Mombasa to determine their efficacy for possible future use. A total of six RBs were placed outside the same houses where BG-traps were deployed in Kizingo and Nyali and also monitored for three consecutive days. Electromechanical aspirators, which included backpack/Prokopack (BP/PP) aspirators, were used to collect indoor resting adult mosquitoes. The time spent at each house varied depending on the size and number of rooms. Additionally, 10 CO2-baited CDC light traps (LT), (John W. Hock Company, Gainesville, FL, U.S.A.) were hung at least two meters from the ground either immediately outside the houses or along the edges of the compound. Each trap was baited with 0.5 kg of dry ice held in igloos next to the traps [41] and left on site from dusk to dawn.
Mosquitoes were retrieved from the traps early every morning (and evening in the case of BG-Sentinel traps) and transported to a temporary site laboratory in Mombasa where they were knocked down using triethylamine (TEA). Collection of mosquitoes from indoors was conducted during the day using BP/PP aspirators. All collected mosquitoes were sorted, morphologically identified to species using keys [42–45] and pooled (≤ 25 mosquitoes per pool) by sex, species, collection method and date. All identification was done on ice packs to preserve the virus for isolation work in cell culture. Identified mosquitoes were preserved in 1.5-ml cryogenic vials and transported in liquid nitrogen to the biosafety level 2 Arbovirus and Viral Hemorrhagic Fever (VHF) Laboratory at KEMRI for analysis.
All water-holding containers found indoors (inside every accessible house) and outdoors (outside the houses and within the peridomestic environment) including some natural habitats such as tree holes and plant leaf axils were inspected using flashlights where necessary. Samples from each positive container were collected using ladles and pipettes or, in the case of jerry cans, the water was poured through a sieve onto a white basin and the larvae or pupae then picked from the sieve using Pasture pipettes. The samples were linked by geo-coding using a GPS to the premises where they were collected. Live immature mosquitoes sampled from each water container type were reared to adults and identified to species as for adult collections. Indoor and outdoor containers were then scored separately as either being wet negative (with no Ae. aegypti immatures) and wet positive (with at least one immature Ae. aegypti), were then scored separately.
The mosquito indices were calculated as House Index (HI)—the percentage of houses positive with immature mosquitoes, Container Index (CI)—the percentage of water holding containers in which mosquito breeding is occurring and Breteau Index (BI)—the number of positive containers per 100 houses [46]. The following formulas were used to determine these indices:
HI=Number of houses with immature mosquitoesNumber of inspected houses×100
CI=Number of containers with immature mosquitoesNumber of wet containersx100
BI=Number of containers with immature mosquitoesNumber of inspected housesx100
Mosquito pools were homogenized in a biosafety level 2 laboratory at KEMRI’s Centre for Virus Research using 4.5-mm diameter copper beads (BB-caliber airgun shot) in 1 ml of Minimum Essential Medium Eagle (MEM), with Earle’s salts and reduced NaHCO3 (Sigma-Aldrich, St. Louis, MO) supplemented with 15% heat-inactivated fetal bovine serum (FBS; Sigma-Aldrich), 2% L-glutamine (Sigma-Aldrich), and 2% antibiotic/antimycotic solution (Sigma-Aldrich) with 10,000 U penicillin, 10 mg streptomycin, and 25 μg amphotericin B per milliliter. The homogenates were clarified by centrifugation at 12000 rpm (Eppendorf centrifuge 5417R) for 10 min at 4°C and the supernatants transferred into 1.5-ml cryogenic vials. Each mosquito pool supernatant (50 μl) was inoculated in a single well of a 24-well culture plate containing a confluent monolayers of Vero cells (CCL81) grown in MEM, which was supplemented with 10% FBS and 2% L-Gulatamine and 2% antibiotic/antimycotic solution. The inoculated cultures were incubated for 45 min to allow for virus adsorption, and each sample maintained in MEM supplemented with 2% FBS and 2% antibiotic/antimycotic solution. The cultures were incubated at 37°C in 5% CO2 and monitored daily, through day 14, for cytopathic effects (CPE) as an indication of virus infection. The samples were also inoculated in C6/36 Aedes albopictus cells) grown in Dulbecco’s Modified Eagle’s Medium (DMEM) (Sigma-Aldrich) and incubated at 28°C.
Total RNA was isolated from the supernatant of each Ae. aegypti mosquito pool and culture exhibiting CPE by the Trizol-LS-Chloroform method [47]. Extracted RNA was reverse transcribed to cDNA using SuperScriptIII reverse transcriptase (Invitrogen, Carlsbad, CA) and random hexamers followed by RT-PCR using AmpliTag Gold PCR Master Mix (Applied Biosystems) [48]. The cDNA was tested using a panel of general (alphavirus and flavivirus) and consensus primers for DENV [49–51]. A positive control cDNA and a no-template negative control were included during the setting up of all PCR reactions. Amplification products were resolved in 1.5% agarose gel in Tris-Borate-EDTA buffer stained with ethidium bromide.
An overall total of 1,676 containers were inspected indoors and outdoors. From these, jerry cans were the most abundant, 704 (42%), followed by tires, 242 (14%), plastic buckets, 228 (14%) and drums, 169 (10%). However, tires had the highest percentage of Ae. aegypti larvae/ pupae positivity, 165 (68%) by container type among the most sampled containers, followed by drums, 71 (42%), plastic buckets, 64 (28%) and jerry cans, 106 (15%).
A total of 827 containers were sampled indoors and 158 of them found positive for Ae. aegypti immatures, giving an indoor CI of 19. Indoor containers were also less diverse (7 container types) and although jerry cans were the most abundant (61%) only 12% of them were positive while 39% each of drums and plastic basins were positive. No immatures were sampled in clay pots and plastic bottles (Table 1).
A total of 849 containers were sampled outdoors and 362 of them found positive for Ae. aegypti immatures, representing an outdoor CI of 43%. Outdoor containers were also more diverse (35 container types). Tires were the most abundant, 242 (29%) and most positive by container type (68%). These were followed by jerry cans, 196 (23%) of which only 22% were positive (Table 2).
An overall total of 192 houses was sampled (between 3 and 70 per site) of which 42 were positive for Ae. aegypti immatures, representing a HI of 22%. A total of 1,676 containers was also inspected indoors and outdoors and 520 (31%) were positive, with an overall CI of 31% and BI of 270.1. All these indices exceeded the WHO-documented thresholds for risk of dengue outbreak/transmission: all indices >1, HI > 1% and BI > 5, suggesting that all the areas sampled were at risk of dengue transmission (Table 3).
From the overall 520 positive containers, 2,510 Ae. aegypti immatures emerged into adult mosquitoes, of which 76% were from large containers: jerry cans, 451 (18%), drums, 431 (17%), buckets, 404 (16%), tires, 359 (14%) and large water tanks, 253 (10%). Although jerry cans were the most productive containers overall, each site had specific container types that were most productive. For instance, tanks and drums were the most and second most productive containers in Kizingo, with 17% and 16%, respectively. Drums were the most productive in Nyali (30%) and Changamwe (33%), plastic basins (35%) in Nyali-B, tires (82%) in Bamburi, plastic buckets (30%) and jerry cans (30%) in Tudor while plastic buckets (81%) were the most productive in Ganjoni (Table 4).
Of the 2,510 Ae. aegypti immatures collected over the entire sampling period, 995 (40%) were from indoors and 1,515 (60%) from outdoors. Kizingo recorded the highest number (1,148) especially outdoors while Nyali (77) and Bamburi (68) recorded the least (Table 5)
A total of 5,461 adult mosquitoes of diverse species were collected indoors and outdoors by a combination of methods. The majority of mosquitoes collected were Cx. pipiens, 2,979 (55%) followed by Ae. aegypti, 2,065 (38%), (Table 6).
Only 78 (4%) adult Ae. aegypti mosquitoes were collected indoors while the majority, 1,987 (96%), were from outdoors. Overall, most of the collections were from Mwembe Tayari, 521 (25%), Kizingo, 317 (15%) and Ganjoni, 312 (15%), while Bamburi recorded the least, 21 (1%). The BGS traps, which were used in all the sampling periods, collected the highest number, 1,460 (73%) followed by CDC light traps, 347 (18%) while resting boxes, used only once during the April 2013 collection, yielded the least, 11 (1%).
Out of 273 pools of Ae. aeypti sampled as immatures and reared to adults and those sampled as adults, identified and processed, one DENV-2 was isolated in Vero cells, and confirmed by RT-PCR [49], from a pool of 2 male mosquitoes collected as adults, representing a minimum infection rate (MIR) of 0.2. No DENV was detected in pools of Ae. aegypti homogenates directly by RT-PCR.
The April-June 2013 and March-June 2014 dengue outbreaks coincided with the long rain seasons along the coast of Kenya. These rains may have resulted in increased aquatic habitats for Ae. aegypti breeding [52], thus increasing the vector population density and the risk of dengue transmission. However drought also promotes vector abundance through increased storage of water in which Ae. aegypti mosquitoes breed [53]. For example the isolated outbreak that was reported in Nyali-B occurred at a time of diminished rainfall reported to be less than 50 mm per month by the Kenya Meteorological Department. Nyali-B, a government institution with dormitories housing approximately 150 people, had no piped water at the time of the outbreak and the residents were storing water in many open container types indoors. Outdoor water storage containers were comprised mainly of large water tanks that were difficult to sample from hence the observed low frequency of Ae. aegypti immatures collected outdoors relative to indoors. The water storage practices resulted in high CI, HI and BI of 37%, 38% and 164.3, respectively, for the Nyali-B dormitories. Overall, the CI of 31% (indoor, 19%; outdoor, 43%), HI (21.9) and BI (270.1) observed for Mombasa in general were also high and well above the WHO-documented thresholds, suggesting that most of the areas sampled in Mombasa city were at risk of dengue transmission. However these threshold levels are controversial since transmission can still occur even when the indices are safely low or fail to occur even when they are high [54,55]. The thresholds also differ from place to place [37] and are affected by human serotype-specific herd immunity and ambient temperature [56]. Hence pupal indices have been recommended instead as the most appropriate for assessing DENV transmission risk especially since there is also no correlation between larval indices and actual pupae that emerge to contribute to adult population [36,57].
Out of all mosquito species collected, only Ae. aegypti is known to transmit DENV in urban areas. Ae. vittatus and Ae. simpsoni both of which co-exist with Ae. aegypti [2] while the Culex spp. have not been associated with DENV. The high number of Ae. aegypti caught implies that the dengue vector is well established in Mombasa and the risk of DF, chikungunya and yellow fever transmission is high in the absence of effective vector control. Adult Ae. aegypti mosquitoes collected indoors were fewer than those collected outdoors, probably reflecting different capture efforts and techniques. However the large number of immatures collected indoors suggests that the coastal Ae. aegypti population breeds indoors just as successfully as outdoors, subject to availability of aquatic habitats, but is mostly an outdoor resting mosquito. Previous studies in Rabai, also a coastal town, demonstrated differential domesticity of Ae. aegypti [59]
Kizingo, one of the most affluent regions, recorded the highest number of Ae. aegypti immature collection in April 2013 and this is attributed to the heavy construction work that was on-going at the time of the outbreak. The many containers which were serving as water reservoirs for the construction work may have provided favorable aquatic habitats for the mosquitoes. Therefore, construction sites should be monitored closely as important sources of Ae. aegypti mosquitoes and especially targeted for vector control to reduce the risk of dengue transmission. Dengue cases were also reported in areas which had low vector densities. This is in agreement with previous observation that low vector density does not always result in lower levels of dengue transmission because a single infected mosquito can transmit the virus to many people given its day biting, anthropophilic, interrupted and multiple-biting behavior [60].
This study has demonstrated further the efficacy of BGS traps which collected 73% of all adult Ae. aegypti followed by the CO2-baited CDC light traps (18%) and BP/PP aspiration (9%) respectively. While BGS traps were only used outdoors during all sampling periods, the BP/PP devices were used mostly to collect resting mosquitoes indoors in only the houses that we were permitted to enter, and only twice out of the three sampling occasions, a situation that may have influenced the catch. Previous studies have also found the BG-Sentinel traps comparatively more effective than other tools [61]. The relatively large number of Ae. aegypti mosquitoes collected by the CO2-baited CDC light traps suggests that this tool can significantly complement other tools in the surveillance of Ae. aegypti at the coastal region of Kenya. While it is difficult to understand how a day-feeding mosquito species can be attracted to CO2 at night, it is possible that Ae. aegypti starts to seek blood meals very early in the morning, perhaps before the traps are collected or they feed beyond sundown allowing some to be attracted to light traps.
In general, this study demonstrated a relationship between the number of containers, level of positivity and adult productivity. For example, the most sampled containers were jerry cans, discarded tires plastic buckets and drums, an observation similar to that made in previous studies in Malindi district, also in coastal Kenya, where these same containers were the highest in positivity and the most productive, although not in the same order [62]. However, container productivity varied by site depending on the type of containers commonly used for water storage regardless of the social status.
Establishment of a disease in an area depends on a number of factors including its maintenance mechanism including appropriate competent vector, availability of amplifying hosts and favorable climatic conditions. Isolation of DENV-2 from a pool of male Ae. aegypti mosquitoes collected as adults during this study marks the first such case in Kenya and provides further evidence that DENV may have maintenance mechanisms that include vertical transmission by mosquitoes [63,64]. This may have epidemiological significance with regard to the maintenance of DENV in nature in conditions adversarial to the virus. The coastal region in Kenya is usually characterized by high temperatures, throughout the year, that favor the proliferation of DENV and subsequent transmission by Ae. aegypti [65,66]. This likely explains the widespread occurrence of dengue cases across the city this time [22].
Considering these factors, dengue fever will likely be a recurrent problem at this coastal city going forward. We recommend that organized vector surveillance and control programs against Ae. aegypti mosquitoes be instituted in Mombasa, in particular, and in Kenyan in general, where currently vector control activities focus on malaria vectors only, as in other parts of Africa [67]. Vector control should involve public participation to focus on routine clean-up campaigns to reduce mosquito-producing containers, a basic step to prevent and/or control dengue outbreaks. This activity should target all container types with the potential to hold water, since this study has demonstrated that the dengue vector can successfully breed in a wide range of container types, and also construction sites for targeted source reduction and maximum adult reductions [54,68]. Community participation through sensitization, public awareness about the disease and the best practices of preserving water and disposal of tires and containers would be key in reducing Ae. aegypti densities. Also providing a reliable supply of piped water in every household would reduce the need for water storage containers that also act as aquatic habitats for dengue vectors. However, success of these efforts will require legislation and proper inter-agency (health and environment) coordination and funding, with the support of the national and county governments. In addition, training of vector control personnel on Ae. aegypti biology, surveillance and control based on the WHO guidelines [10] should be prioritized.
|
10.1371/journal.ppat.1006256 | Integrated systems biology analysis of KSHV latent infection reveals viral induction and reliance on peroxisome mediated lipid metabolism | Kaposi’s Sarcoma associated Herpesvirus (KSHV), an oncogenic, human gamma-herpesvirus, is the etiological agent of Kaposi’s Sarcoma the most common tumor of AIDS patients world-wide. KSHV is predominantly latent in the main KS tumor cell, the spindle cell, a cell of endothelial origin. KSHV modulates numerous host cell-signaling pathways to activate endothelial cells including major metabolic pathways involved in lipid metabolism. To identify the underlying cellular mechanisms of KSHV alteration of host signaling and endothelial cell activation, we identified changes in the host proteome, phosphoproteome and transcriptome landscape following KSHV infection of endothelial cells. A Steiner forest algorithm was used to integrate the global data sets and, together with transcriptome based predicted transcription factor activity, cellular networks altered by latent KSHV were predicted. Several interesting pathways were identified, including peroxisome biogenesis. To validate the predictions, we showed that KSHV latent infection increases the number of peroxisomes per cell. Additionally, proteins involved in peroxisomal lipid metabolism of very long chain fatty acids, including ABCD3 and ACOX1, are required for the survival of latently infected cells. In summary, novel cellular pathways altered during herpesvirus latency that could not be predicted by a single systems biology platform, were identified by integrated proteomics and transcriptomics data analysis and when correlated with our metabolomics data revealed that peroxisome lipid metabolism is essential for KSHV latent infection of endothelial cells.
| Kaposi’s Sarcoma herpesvirus (KSHV) is the etiologic agent of Kaposi’s Sarcoma, the most common tumor of AIDS patients. KSHV modulates host cell signaling and metabolism to maintain a life-long latent infection. To unravel the underlying cellular mechanisms modulated by KSHV, we used multiple global systems biology platforms to identify and integrate changes in both cellular protein expression and transcription following KSHV infection of endothelial cells, the relevant cell type for KS tumors. The analysis identified several interesting pathways including peroxisome biogenesis. Peroxisomes are small cytoplasmic organelles involved in redox reactions and lipid metabolism. KSHV latent infection increases the number of peroxisomes per cell and proteins involved in peroxisomal lipid metabolism are required for the survival of latently infected cells. In summary, through integration of multiple global systems biology analyses we were able to identify novel pathways that could not be predicted by one platform alone and found that lipid metabolism in a small cytoplasmic organelle is necessary for the survival of latent infection with a herpesvirus.
| Viruses have evolved functions to reprogram the proteomic landscape of their host and modulate cellular signaling pathways to adjust the regulation of cellular machinery. These cellular alterations support the survival of infected cells to allow replication and spread of the virus. Many viruses rewire host cell signaling pathways to activate the host cell and to enable lytic replication, and in the case of the herpesviruses, to support long-term latent infection [1, 2]. During latency, herpesviruses are known to modulate host cell signaling pathways that lead to inhibition of apoptosis, subversion of the host immune response, and alteration in host carbon and lipid metabolism among many other pathways. Importantly, alteration of these pathways by some oncogenic gamma-herpesviruses may influence tumor formation given the optimal cellular milieu [3, 4].
Kaposi’s Sarcoma Associated Herpesvirus (KSHV), a human gamma-herpesvirus, is the etiological agent of Kaposi Sarcoma and two B-cell lymphoproliferative diseases, Primary Effusion Lymphoma (PEL) and Multicentric Castleman Disease (MCD) [5–7]. KS is the most common AIDS-associated malignancy worldwide and among the most common tumors overall in Sub-Saharan Africa [8]. KSHV is found in the main KS tumor cells, the spindle cells, which are cells of endothelial origin [9, 10]. In the KS spindle cells, KSHV is predominantly in the latent state (>90%) where only a handful of the more than 90 annotated viral genes are expressed as well as a number of viral microRNAs [11, 12]. A limited number of spindle cells (< 5%) express markers of lytic replication as well [13]. While there are limited animal models for the disease, there are well-established mammalian cell culture systems that recapitulate the latent and lytic infection rates seen in KS tumors [14–17]. We and others have successfully used these cell culture models to demonstrate that KSHV promotes angiogenesis, modulates carbon utilization and alters lipid profiles in KSHV latently infected endothelial cells [18–21]. Our previous work showed that latent KSHV infection leads to profound changes in central carbon metabolism and fatty acid (FA) synthesis and that both are required for the survival of latently infected cells indicating the importance of altered metabolism and lipid homeostasis to latent infection [19, 22]. Many of these cellular changes induced by KSHV are similar to phenotypes that commonly occur in cancer cells [3].
Several of the signaling pathways modulated by KSHV infection have been studied through traditional approaches of identifying individual host proteins or pathways predicted to play a role in the phenotype investigated. Here we are applying a more comprehensive approach where the global response of cell host in response to KSHV infection during latency at the protein and transcript levels are evaluated. Systems biology approaches can be utilized to identify important cellular networks on a cell-wide scale. In particular, advancement of recent mass spectrometry-based techniques using affinity-based phosphopeptide enrichment coupled with chemical labeling and high-resolution chromatography have been adapted to query changes in protein phosphorylation [23–25]. In addition, the assembly of large-scale, high-quality, protein-protein interaction databases provide an extensive and detailed context for interpreting proteome changes [26]. To evaluate gene expression profiles, next generation sequencing technology provides comprehensive analysis of the presence and quantity of the transcriptome. The use of transcriptomics data to predict transcription factor (TF) activity as a function of changes in mRNA provides an effective tool to link proteomics to transcriptomics data [27]. The integration of these two different data types has been successfully demonstrated in several biological systems, including glioblastoma [28], breast cancer [29], epithelial-mesenchymal transition [30], yeast salt stress response [31] and influenza virus infection [32]. These studies have provided insights and a comprehensive view of cellular networks from stimuli to gene expression/suppression.
We performed a systems-level data integration approach to identify global changes in cellular networks that are important for KSHV latent infection. To dissect cellular changes and examine the signal transduction from upstream signaling to downstream targets induced by KSHV infection, we first conducted a mass spectrometry-based proteomics and phosphoproteomics analysis, including both tyrosine and serine/threonine phosphoproteomics. We also evaluated gene expression profiles following KSHV infection using high throughput sequencing to generate global cellular transcriptomics data. Virally induced changes in both the proteome and the transcriptome were integrated using an inference algorithm to predict TF activation. A comprehensive protein-protein interaction network was used to identify predicted cellular pathways subverted by KSHV. This integrated systems biology approach identified multiple pathways altered by KSHV infection including peroxisome metabolism.
Peroxisomes have been identified as a nexus of lipid metabolism and signaling [33, 34]. While we have previously shown that FA synthesis is required for the survival of endothelial cells latently infected with KSHV, how these downstream FAs are utilized and why they are necessary have not been determined [19]. Peroxisomes have been studied in the context of infection with RNA viruses including influenza [35–40]. Interestingly, infection with influenza virus led to an increase in peroxisomes while infection with flaviviruses led to a significant decrease in peroxisomes metabolism. Our results show that KSHV latent infection of endothelial cells leads to increased numbers of peroxisomes. One major function of peroxisomes is to metabolize very long chain fatty acids (VLCFAs). Peroxisomal defects have been associated with several clinical disorders, including. Zellweger syndrome, a disease characterized by abnormal peroxisome lipid metabolism presenting with deficiency of ACOX1 function, D-bifunctional protein (D-BP) and X-linked adrenoleukodystrophy (X-ALD) [41]. Lipidomics analysis in fibroblasts cells from these patients present with abnormal lipid profiles specifically high levels of VLCFs and low levels of DHA indicating abnormal function of VLCFs breakdown and DHA synthesis [42]. ABCD3 is a peroxisomal lipid transporter of VLCFAs involved in transporting 24:6n3, the precursor of DHA [34]. After 24:6n3 is transported into the peroxisome; it gets further metabolized by ACOX1, a peroxisomal enzyme. ACOX1 synthesizes DHA by partial β-oxidation of 24:6n3 [43, 44]. In the current studies, transient knockdown of ACOX1 and ABCD3 led to cell death in the KSHV latently infected endothelial cells but not the mock-infected control. Overall, these findings validate our integrated global approach and strongly suggest that KSHV modulates peroxisomal lipid metabolism for the increased maintenance of latently infected cells.
To quantify global signaling events modulated by KSHV, we used quantitative phosphoproteomics and proteomics to compare mock and KSHV infected endothelial cells (Fig 1A). Tert-immortalized microvascular endothelial cells (TIME) [16] were mock or KSHV infected and harvested at 48 hours- post-infection (hpi), when latency has been established. Three biological replicates were performed with separate infections performed on different days (Fig 1A). Latent infection, in greater than 90% of the cells was confirmed by immunofluorescence (IFA) assays. This approach identifies the presence of a latent protein (ORF73) and the absence of ORF59, a protein marker of lytic infection. ORF59 stained positive in less than 2% of the infected cells in all experiments. To quantify differential peptide expression levels between mock and KSHV infected cells, each sample was chemically labeled with isobaric tags for relative and absolute quantification (iTRAQ) [45] (Fig 1A). Peptide quantification was normalized prior to labeling using quantitative fluorimetric peptide assay to ensure that similar amounts of peptides were labeled across all samples. Labeled peptides from each biological sample were pooled (Fig 1A). Peptides were, then separated, sequenced and analyzed using one or two-dimensional (1D or 2-D) HPLC tandem high-resolution mass spectrometry (LC- MS/MS) for phosphotyrosine enrichment and total phosphroteome and proteome, respectively (Fig 1A).
LC-MS/MS analyses were conducted in three stages. First, low-abundance phosphotyrosine (pT) containing peptides were identified and quantified after enrichment by immunoprecipitation (IP) (Fig 1A). The IP flow-through was then used to enrich and quantify peptides containing phosphorylated serine (pS), threonine (pT) and the remaining tyrosine residues using immobilized metal affinity chromatography (IMAC). Finally, the IMAC flow-through was used to quantify total protein levels (Fig 1A). Upon data acquisition and analysis, we confirmed there was not a statistically significant difference between the mean relative abundance of peptides across the samples, indicating that the sample labeling was equally effective in each case (S1A–S1D Fig). A total of 2304 unique proteins from the proteome and 1038 unique phospho-proteins from the phosphoproteome runs were analyzed that includes phospho-tyrosine/threonine and serine. Activation of a phosphorylated residue within the same protein can vary; therefore, we analyzed individual peptides. From the phosphoproteome, we analyzed 1644 unique phospho-peptides, including 175 unique phosphotyrosine peptides that comprised 75 unique phosphotyrosine proteins (Fig 1B). The protein and peptide population distribution for the proteome and phosphoproteome, respectively, were plotted based on the sum of relative peptide intensities from the iTRAQ reporter ions from mock and KSHV infected samples versus log10 of the ratios/fold change of KSHV over mock (Fig 1D and 1E and S2 Table). Of the 1644 unique phospho-peptides identified, 192 were differentially phosphorylated, of which half were upregulated and half were down regulated (paired t-test p < .05). Phosphorylated signal transducer and activator of transcription 3 (STAT3) is the top hit of the tyrosine-phosphorylated residue from the phosphoproteome analysis (Fig 1D). Our lab has previously shown that KSHV induces persistent activation of phospho-STAT3 during latency validating the phosphoproteomic results [46]. From the upregulated hits including both phosphoproteome and proteome, there are several proteins involved in metabolism, immunity, insulin resistance, endocytosis, NFk-B signaling and others, providing potentially interesting targets for future study. From the proteome analysis, we measured 2304 unique proteins among which 289 were altered by KSHV latent infection; 164 were upregulated and 125 downregulated. This corresponds to viral induced changes in 13% of the proteins detected and 12% of the phosphorylated residues, including unique phosphotyrosine (Fig 1B). Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway analysis of the phosphoproteins and proteins measured and altered during KSHV infection identified several pathways consisting of more than 4 proteins annotated [47] (Fig 1C). These pathways include metabolic processes involved in carbon and lipid metabolism as well as hypoxia inducible factor (HIF) signaling, both of which had been previously associated with KSHV latent infection, providing internal positive controls for our proteomic data [18, 22, 48]. Gene Ontology analysis also provided similar results (S2B Fig).
To identify changes at the transcriptional level, high throughput cDNA sequencing from mRNA was performed to identify genes expression differences between mock and KSHV infected endothelial cells at 48 hpi. Three separate mock and KSHV infections of TIME cells performed on different days were analyzed by high throughput sequencing of cDNA. Expression of 12,375 cellular genes in all replicates were identified. Of the genes measured, 985 cellular genes were significantly upregulated following KSHV infection of endothelial cells and 1,134 were significantly downregulated at a 1% FDR using a method based on the negative binomial distribution (Fig 2A and S3A Fig).
The transcriptomic data was used to predict the activities of transcription factors based on binding motifs in the promoter regions of transcripts that are activated or repressed following latent KSHV infection. The enrichment of a motif in the promoters of genes whose expression is significantly altered implicates the motif’s associated TF as a possible regulator (Fig 2B). The software FIMO identified putative binding sites by motif presence [49], and two-sided Wilcoxon rank-sum tests [50] quantified and assigned p-values to the enrichment of those binding sites in promoters of genes significantly changed in expression after KSHV infection (Fig 2B).
FIMO was used to scan for the locations of 426 TF binding motifs, from a curated database of position-weight matrices compiled and derived from multiple experimental types, in 1000bp regions upstream of annotated transcription start sites [27]. The enrichment scores of the 261 motifs whose corresponding TF’s mRNAs were reliably detected in the RNA-seq data and the mRNA’s fold-change in expression after KSHV infection (Fig 2C). A positive enrichment z-score indicates that the motif’s putative target genes increase in expression on average, and a negative enrichment score indicates that the motif’s putative target genes decrease in expression on average. Wilcoxon rank-sum tests assigned statistical significance to the motif enrichments and found that five motifs were significantly enriched at a 5% FDR (p-value < 0.001) and twenty-four more were enriched at a less stringent cutoff of p-value < 0.05 (Fig 2C).
The motif of four of these TFs, interferon regulatory factors 1, 2, 7 (IRF1, IRF2, IRF7) and STAT2, are enriched only in upregulated promoters. However, because the motifs for these factors are similar (S3C Fig), it is not clear which of the TFs or TF complexes are actually relevant from just the motifs. The mRNA of IRF1 exhibits a 3.3-fold increase in gene expression (p-value < 0.001) as measured by our transcriptomics data, which may imply that IRF1 is a more relevant player. A motif associated with the transcriptional repressor zinc finger protein 148 (ZNF148), was significantly enriched in downregulated promoters (S3C Fig), suggesting that ZNF148’s repressor activity increased post-infection. The repressor E2F5’s motif was also significantly enriched in downregulated promoters (p-value = 0.0038). It has been shown that E2F5 is inhibited by retinoblastoma protein 1, which is directly inhibited by the KSHV protein, LANA [51]. These data support that motif enrichment scores can successfully denote prediction of transcription factor activity for use in the analyses below.
To build a comprehensive network model that describes the host response to viral infection, we used the Prize-Collecting Steiner forest algorithm to integrate the proteomic and TF motif analyses [52]. This algorithm parsimoniously identifies the protein-protein interaction most likely to be relevant for connecting the relevant factors identified in the two types of analyses. In addition, it identifies Steiner nodes, which are proteins that were not implicated in the proteomic or TF analyses but form crucial connections between other important proteins identified as altered by KSHV in the global data sets generated. The proteomic data was integrated with the TF enrichment scores rather than the differentially expressed genes from the RNAseq data because TF transcript levels do not necessarily reflect regulatory activity [53]. When combining gene expression data with other types of protein scores for pathway reconstruction, it is therefore preferable to use inferred TF activities [28, 29, 32, 54]. The complete predicted Steiner forest network is large, connecting hundreds of proteins that respond to KSHV infection and the TFs inferred to regulate the transcriptional changes (S4 Fig). Randomization analysis shows that the selected proteins and interactions are specific to KSHV infection and do not reflect biases in the protein-protein interaction network or Steiner forest algorithm (S7A and S7B Fig).
To focus on specific biological functions, we assessed the overlap between the proteins in our Steiner forest network and KEGG pathways. We required a minimum overlap of 5 proteins and used the Benjamini and Hochberg multiple hypothesis test correction (FDR < 10%). The significantly enriched pathways included pathways involved in phagocytosis, immune response and several metabolic processes among others (Fig 3A and S1 Table for complete list). Our lab has shown that metabolism is altered during latent KSHV infection, including carbon and lipid metabolism, which supports our integrated network analysis [19, 22, 48]. There are several interesting pathways that are predicted to be altered during KSHV latent infection (S1 Table). From this analysis, we decided to follow up on proteins that clustered together, particularly those involving peroxisome metabolic lipid signaling. We have previously identified that lipid metabolism is required during latency [19], but how these metabolites are further utilized during KSHV latent infection still unknown. Therefore, we chose to further analyze activation of peroxisomes by KSHV. Peroxisome related proteins identified in the subnetwork including SCP2, PRDX5, ACSL3, MLYCD, AGPS, EHHADH, PEX19 were upregulated following KSHV infection and two Steiner nodes PEX12 and PEX5 were predicted by the algorithm to be activated by KSHV (Fig 3C). The proteins in this cluster are involved in lipid metabolism (SCP2, ACSL3, MLYCD, AGPS, EHHADH) and peroxisome organelle biogenesis and transport (PEX19, PEX12 and PEX5). Therefore, this sub- network predicts that KSHV induces peroxisome pathways involved in lipid metabolism and biogenesis. In addition, IRF3 is an interferon inducible gene activator that was predicted in the TF analysis to have increased transcriptional activity. It is not annotated as a peroxisome pathway protein, but the Steiner forest algorithm includes IRF3 in our peroxisome subnetwork due to its predicted relationship with PEX19 and previous evidence of a direct IRF3-PEX19 interaction [55] cataloged in the iRefIndex database. This observation suggests that peroxisomes might also play a role in immune signaling during latency, which might be an important regulatory control point of KSHV infection.
In addition, we incorporated a protein-protein interaction database from a study that mapped global interactions between KSHV genes and host proteins using viral gene pulldowns [56]. Since our study is mainly focused on latency, we included only the KSHV latent viral proteins and host proteins hits with our predictive Steiner forest network. From the database, we found that two viral genes have been shown to interact with proteins associated with peroxisome biogenesis. The latent KSHV proteins that are predicted to be associated with peroxisome biogenesis are shown as purple diamonds in Fig 3C. The KSHV protein-protein interaction database used total protein pull downs and therefore does not demonstrate direct interactions, rather it shows associations with the identified protein. All the proteins from the KSHV major latent locus were included in the Steiner forest analysis shown in figure S5 Fig. The utilization of KSHV protein-protein interaction dataset with the other protein-protein interaction databases advances our predictions of pathways that could be important to KSHV pathogenesis.
To validate the prediction that peroxisome pathways involved in lipid metabolism and likely peroxisome biogenesis are increased during latent infection, we examined the protein levels of ABCD3, an ATP Binding Cassette Subfamily D Member 3, a lipid transporter specific to peroxisomes and a common marker to study peroxisomes, using flow cytometry at 48 and 96 hpi (Fig 4A–4H). Staining with an antibody to ABCD3 showed a significant increase in fluorescent staining in the KSHV infected cells compared to mock infected cells in three different infections at 48 and 96 hpi in TIME cells and 96 hpi in primary human dermal microvascular endothelial cells (hDMVECs) and lymphatic endothelial cells (LECs) (Fig 4E–4H), indicating that during latent infection KSHV significantly upregulates ABCD3 protein expression. In addition, we evaluated MLYCD and PEX19 and a non-clustered peroxisome protein PEX3 levels, in TIMECs, hDMVECs and LECs. PEX3 is a PEX19 docking factor required for PEX19 to deliver proteins into the peroxisome matrix [57]. Staining with MLYCD and PEX3 antibody showed a significant increase of the protein levels in KSHV infected TIME cells, primary hDMVECs and primary LECs compared to mock infected at 96 hpi (S6 Fig) while PEX19 was significantly upregulated in TIME cells (S6 Fig).
We next evaluated the peroxisome organelle number using confocal imaging analysis of mock and KSHV infected TIME cells stained with antibody to ABCD3. Peroxisome size ranges approximately between .4–1 uM. We evaluated 3D particles using z-stacks imaging and then quantified the particle number as a proxy of peroxisome organelle per cell with a minimum threshold of .5 uM. There is an approximately 50% increase in the number of peroxisomes per cell in the KSHV infected endothelial cells as compared to mock infected cells (Fig 4I and 4J). Representative pictures of the mock and KSHV infected cells stained with an antibody to ABCD3 are shown in Fig 4I. Combined, these observations support the prediction from the Steiner forest analysis that KSHV promotes peroxisome biogenesis.
To determine that the increase of peroxisome numbers per cell was not a cellular response to infection but rather induced by virally encoded genes, cells infected with UV irradiated KSHV were stained with ABCD3 antibody and measured by flow cytometry. UV irradiated virus can bind and enter cells but does not express viral genes. Flow cytometry analysis showed no increase in the expression of ABCD3 following infection with UV irradiated virus (Fig 5A and 5B). Therefore, the increase of ABCD3 in latently infected cells requires KSHV gene expression and it is not a cellular response to virus entry into the cell.
The KSHV latent locus is comprised of LANA, vCyclin, vFlip, Kaposins and 12 microRNA loci. To assess the role of the KSHV latent locus in increasing the number of peroxisomes, we evaluated whether the KSHV latency associated region (KLAR) is sufficient to induce the increase of ABCD3 protein expression levels. The KLAR locus (a kind gift from Dr. Rolf Renne) was cloned into a helper-dependent gutted adenovirus vector that does not express any adenovirus genes. Cells were infected with a control gutted adenovirus (Ad) only expressing GFP (AdGFP) and the gutted adenovirus expressing KLAR (AdKLAR) and stained with ABCD3 antibody. Infection rates for AdGFP and AdKLAR were 59% and 97% respectively as determined by expression of GFP or LANA. To adjust for the differences in the infection rates, we gated on the GFP positive cells from the AdGFP infected cells and then compared to the AdKLAR cells. Cells infected with the AdKLAR expressing gutted adenovirus exhibited increased ABCD3 protein expression compared to mock infected cells and AdGFP (Fig 5C and 5D). Therefore, the latency genes are sufficient to induce ABCD3 protein expression levels.
Peroxisomes are involved in lipid signaling and metabolism. Our previous metabolomics screen indicated that several lipid metabolites are altered by KSHV during latent infection of endothelial cells, including two metabolites generated in the peroxisome, dihydroxyacetone phosphate (DHA-P) and docosahexaenoate (DHA; 22:6n3) [19] and metabolites upstream of DHA are also upregulated as indicated in red numbers (Fig 6A). DHA is synthesized from 24:6n3 by Acyl-CoA Oxidase 1 (ACOX1) an enzyme mainly expressed in the peroxisome and it is involved in the first step of peroxisomal β-oxidation [43] (Fig 6A). To determine if ACOX1 is necessary during KSHV latent infection, small interfering RNA (siRNA) was used to knockdown its gene expression (Fig 6B). Loss of ACOX1 did not alter the cellular proliferation of uninfected cells or the KSHV infection rates but resulted in a significant increase in cell death of the KSHV infected cells compared to controls at 96 hpi (Fig 6C–6E). As ACOX1 is the main enzyme involved in metabolizing DHA, these observations suggest that DHA might be required during infection.
The precursor of DHA, 24:6n3 is transported into the peroxisome by the lipid transporter ABCD3 [34, 58]. Therefore, we evaluated if ABCD3 is required during latency by transiently silencing its gene expression. Similarly, to ACOX1, loss of ABCD3 did not alter cellular proliferation of uninfected cells or KSHV infection rates but resulted in a significant increase in cell death of the KSHV infected cells compared to controls at 96 hpi (Fig 6C–6E). Therefore, both ACOX1 and the ABCD3 transporter are required for the survival of endothelial cells latently infected with KSHV. In parallel, we treated cells with a pan-caspase inhibitor, QVD, to test whether apoptosis was the main cell death mechanism. KSHV-siABCD3 and KSHV-siACOX1 cells treated with QVD showed a 3-fold decrease in cell death, indicating that apoptosis was the main cell death mechanism (Fig 6D). Therefore, this data strongly indicates that peroxisomal proteins involved in lipid metabolism are required for the survival of endothelial cells latently infected with KSHV.
We integrated transcriptomics, proteomics and metabolomics analyses, to provide a comprehensive view of cell signaling in an oncogenic virus infection in human endothelial cells, the cell type likely to be most relevant to KS tumor cells (Fig 7). From quantitative measurements of the phosphoproteome and proteome analysis of endothelial cells latently infected with KSHV, we found that latent infection alters the levels of at least 289 proteins, approximately 13% of the proteome quantified, as well as 192 altered phosphorylation sites, approximately 12% of the phosphosites quantified in this study. Previous studies using mass spectrometry based proteomics and KSHV, was done with targeted proteomics to identify protein-protein interactions specific to single viral proteins, LANA and K5 for example, using immunoprecipitation and 2D-gel mass spectrometry [59–69]. Our dataset is the first that we are aware of, that analyzes the global response to latent KSHV infection with both phosphoprotein and proteomic studies in endothelial cells. The list of phosphosites altered by KSHV infection may provide deeper insights into cell signaling activation following KSHV infection of endothelial cells and should serve as a useful dataset for future studies. From transcriptomics analysis, we found that KSHV infection leads to alterations in approximately 17% of the host cellular transcriptome. This dataset was generated using next generation sequencing providing more comprehensive gene expression profiles in endothelial cells latently infected with KSHV than previously published. Transcriptomic analysis of KSHV infection has been done in KS tumors and in PEL cells, but only older microarray technology for endothelial cells has been previously done [70–78]. The activity of several TFs was predicted to be activated or repressed by latent infection as identified from transcription factor motifs found in the promoters of host genes that were up or down regulated following KSHV infection of endothelial cells. These TFs serve as a link to map protein-protein interactions, connecting upstream signaling to downstream gene-expression targets. The goal of this integrated systems biology approach was to identify novel pathways that could not be predicted by one platform alone. Various functional networks, including phagosomes, endocytosis and multiple metabolic pathways including peroxisome biogenesis were identified by the Steiner forest analysis providing a rich data set for future studies. We chose to further analyze peroxisome biogenesis as peroxisomes are involved in several pathways likely to be important for KSHV pathogenesis including redox control and the breakdown of very long chain fatty acids. A sub-network cluster of peroxisomal proteins predicted to be activated by the Steiner forest analysis is shown in Fig 3C. The presence of this sub- network implies increased peroxisome activity in KSHV latently infected endothelial cells. The integrated analysis is further substantiated by a significant increase in the number of peroxisomes per cell during KSHV latency, induced specifically by KSHV latent gene expression as opposed to a cellular response to a viral infection. Upregulation of peroxisomes was further validated by identifying the upregulation of several peroxisomal proteins in TIME cells, primary dermal microvascular endothelial cells as well as in primary lymphatic dermal microvascular endothelial cells, the cell type that most closely resembles KS spindle cells [78].
We previously found that KSHV latent infection dramatically alters the lipid profile of endothelial cells [19]. In the KSHV infected cells, there was a significant increase in most of the LCFAs measured. We also found that FAs synthesis was necessary for the survival of endothelial cells latently infected with KSHV [19]. In our metabolomic screen we also noted that DHA and its precursors, as well as DHA-P, were increased following KSHV infection during latency [19] (Fig 6A). DHA is an important metabolite involved in anti-inflammatory responses and cellular development and is mainly produced in the peroxisome by partial β-oxidation [43, 44, 79]. Knockdown of ACOX1, the enzyme that produces DHA, results in a significant increase in the death rate of latently infected endothelial cells but not their mock infected counterparts. Furthermore, the peroxisome-specific lipid transporter ABCD3, which transports VLCFAs including a precursor of DHA, is also essential for KSHV-infected endothelial cell survival. Therefore, lipid metabolism in the peroxisome is essential for the survival of endothelial cells latently infected with KSHV.
Peroxisomes appear to play a role in the response to lytic viral infection for several viruses. This organelle serves as a signaling platform for antiviral response against unrelated non-enveloped and lipid-enveloped RNA and DNA viruses including Reovirus, Sendai virus, Dengue virus and Influenza virus in infected mouse embryonic fibroblast cells [36–39]. Interestingly, one study established that Influenza virus modulates and requires peroxisomal ether lipid metabolism for efficient virion replication in A549 epithelial cells [37]. These observations underscore complex and sometimes paradoxical cellular changes that involve peroxisomes during viral infection; while Influenza requires peroxisome metabolism for virion production, peroxisomes also play an important role in the immune response triggered by infection. The interplay between peroxisome responses and viral infection may depend on the cellular environment and virus type. Our study elucidates a novel mechanism by which a latent herpesvirus infection manipulates peroxisomal lipid signaling required for survival in a long-term infection.
KSHV is known to activate COX-2/PGE2/EP vector during de novo infection mediating an underlying pro-inflammatory state conducive to long-term latency [21]. COX-2 converts Arachidonic Acid (20:4n6, or AA) to PGE2, which then regulates autocrine and paracrine signaling. Our previously published metabolomics screen demonstrates that KSHV latent infection upregulates precursors of the COX-2/PGE2/EP signaling, such as AA indicating that signaling upstream of the COX-2/PGE2/EP pathway is active during latency. Furthermore, upregulation of DHA-P and DHA as shown by our metabolomics screen indicates, that the peroxisomes are enzymatically active and producing these metabolites [33]. Therefore, the peroxisome represents a crossroads of lipid signaling and bridges the gap between upstream essential fatty acids (AA, EPA and DPA) and how they are metabolized downstream (DHAP, DHA) during latent KSHV infection (Fig 6A).
AA is a pro-inflammatory metabolite and DHA has been associated with anti- inflammatory responses [80]; however, both are upregulated during latency. The KS tumor environment is characterized by a chronic inflammatory state [11]. Therefore, we hypothesize that KSHV commandeers control of cellular metabolic pathways to fine-tune a higher level of chronic inflammation by altering homeostatic mechanisms and maintaining a shifted equilibrium in this new inflammatory state required for the maintenance of latency. Further work is required to elucidate whether the primary role of peroxisomes is to regulate lipid signaling and inflammation, if peroxisomal ether lipid metabolism is required or if peroxisomes are also involved in regulating H2O2, which often occurs in parallel. It has been shown that in primary endothelial cells during exogenous stress, inflammatory cytokine expression is downregulated by using DHA as anti-inflammatory treatment [81]. It would be interesting to determine if altering DHA synthesis influences inflammatory signaling proteins in endothelial cells latently infected with KSHV.
Currently, pharmacological approaches that target herpesvirus infection focus on lytic replication and there are no treatments specific for latent infections. Since KS tumors primarily exhibit latent infection, this study elucidates critical control point mechanisms in the latent phase offering an understanding of KSHV viral pathogenesis and provides potential novel and combinatorial molecular therapeutic targets through large scale identification of pathways activated by KSHV latent infection of endothelial cells.
QVD-OPH (SMBiochemicals) was dissolved in DMSO and used at a final concentration of 20 μM. YOYO-1 and SytoGreen were purchased from Thermofisher scientific. Tert-immortalized microvascular endothelial (TIME) cells were obtained from the McMahon lab and previously described in Venetsanakos, et. al. [82], human dermal microvascular endothelial cells (hDMVECs) and lymphatic endothelial cells (LECs) (LONZA Walkersville, MD) were maintained as monolayer cultures in EBM-2 media (LONZA Walkersville, MD) supplemented with a bullet kit containing 5% FBS, vascular endothelial growth factor, basic fibroblast growth factor, insulin-like growth factor 3, epidermal growth, and hydrocortisone. KSHV for phosphoproteomic, proteomic and transcriptomic experiments was purified from BCBL-1 cells, a primary effusion lymphoma cell line that maintains wild type KSHV, as described previously [22]. For most of the subsequent studies, KSHV was isolated from iSLK cells containing a recombinant KSHV made from KSHV-Bac16 containing the GFP gene as described previously [83]. For all experiments KSHV was titered and used to infect TIME cells as previously described [46]. Infections were performed in serum-free EBM-2 media and subsequently overlaid with complete EBM-2 media. Infection rates were assessed for each experiment by immunofluorescence and only experiments where greater than 90% of the cells expressed LANA, a latent marker, and less than 2% of the cells expressed ORF59, a lytic marker, were used as previously described [46].
To express the KSHV latent genes in the absence of other viral gene expression, The 12.6 kbp KSHV latency associated region (KLAR) containing the native LANA promoter, LANA, vCyc, vFLIP, all 12 miRNA loci and the kaposins through the native polyadenylation signal downstream of the kaposins, was obtained from the Renne lab. The helper dependent Adenovirus contains the adenovirus packaging signal but no adenovirus genes and was purchased from MicroBix. To create AdKLAR and AdGFP, the KSHV KLAR region or GFP was cloned into a shuttle vector (pBShuttle) flanked by adenoviral sequences. The KLAR/adenovirus expression cassette was then excised from this plasmid and electroporated into BJ5183 cells (Stratagene) along with pC4Hsu helper adenovirus vector (Microbix Biosystems) to allow for homologous recombination. The resulting plasmid (AdKLAR or AdGFP) was transfected into 293Cre cells, which stably express a Cre recombinase enzyme, selectable with puromycin. Cells were passaged in the presence of helper adenovirus (HD14; Microbix), which contains the adenovirus coding regions and allows to produce AdKLAR adenovirus. The Helper Adenovirus contains a modified packaging sequence flanked by loxP sites; therefore, the helper adenovirus is not packaged due to an excision of the packaging sequences. After expansion of the adenovirus, cells were collected, pelleted, and freeze-thawed three times using liquid nitrogen and 37 C water bath. Cell debris was spun out at 2000rpm and the cell-free supernatant was collected. The cleared lysate was layered onto a continuous 15% to 40% CsCl gradient and centrifuged for 2–3 hours at 35,000g using a SW41Ti rotor (Beckman Coulter, Inc., Fullerton, CA). The mature virus band was collected and purified in a second CsCl density gradient. The virus band was collected, dialyzed against three changes of A195 buffer. Infections were performed in serum-free EBM-2 medium supplemented with 1μg/mL poly-L-lysine for 1 hour, after which the medium was replaced with complete EGM-2 media. Infection rates were assessed for each experiment by immunofluorescence for LANA, a latent marker, and GFP for AdGFP expression.
Mock-, KSHV-infected cells were washed with PBS and removed by trypsinization, fixed, with 4% paraformaldehyde for 30 min on ice and processed for flow cytometry. Cells were permeabilized and blocked with .1% triton and 1% NGS. The ABCD3 transporter was detected with the PMP70 (ABCD3) antibody from thermoscientific product# PA 1–650, MLYCD Proteintech Group (15265-1-AP), PEX3 Novus a Biotechne brand (NBP1-86210) and PEX19 Abcam (ab 137072). After staining with the primary antibody for 1 hour, the cells were reacted with a secondary Alexa Fluor 594 and Alexa Fluor 488 both anti-rabbit (ThermoFischer P#A11072) antibodies. Samples were analyzed by FLOWJO, flow cytometry analysis software.
TIME cells were seeded on a 4-well glass chamber and were processed for confocal microscopy by fixing in paraformaldehyde (4% in 1XPBS) at 37°C. Samples were permeabilized with Triton X-100 (0.5% in 1XPBS). Incubations with primary antibodies diluted (1:1,000) in blocking buffer (3% bovine serum albumin [BSA] and 1XPBS) were carried out at room temperature (RT) for 30 minutes. Samples were then incubated with secondary antibodies (Alexa Fluor 488 anti-rabbit) in blocking buffer for 25 min at RT. Prior to mounting; samples were incubated with DAPI for 5 min at RT coverslips were mounted on microscope slides. Confocal images were acquired using Zeiss LSM 510 Meta confocal microscope Olympus.
2-5um Z-stacks were acquired using a Zeiss 510 META confocal microscope equipped with a 63X / 1.4 NA Oil DIC objective. The exported images were then processed using Imaris 7.2.3 software (Bitplane) for peroxisome quantification and ImageJ was use for figure images. Cytoplasmic peroxisomes were quantified based on voxels graphics. The data were then analyzed using student’s t-tests.
siRNAs specifically targeting ACOX1 (pre-validated) were purchased from Santa Cruz Biotechnology (Cat. Sc-94104). A negative-control siRNA (siSCRB) and ABCD3 (pre-validated) were designed and synthesized by Ambion. TIME cells were transfected with siRNA using the Amaxa Nucleofector Kit by Lonza per the manufacturer’s protocol. At 24 hour post transfection, cells were Mock- or KSHV-infected. At 96 hpi cell death was measured using Trypan blue assay and cell were quantified using TC20 cell counter from BioRad. In parallel, cell death fluorescent images were acquired using the IncuCyte from Essen Bioscience using YOYO-1 or SytoGreen (both probes from Thermofisher scientific).
Approximately 5 million cells were lysed in 2 mL of 8M Urea. Protein concentration was determined by the BCA assay (Pierce). Samples were reduced with 5 mM dithiothreitol at 56 C for 1 hour, and then alkylated with 15 mM iodoacetamide for 1 hour at RT in the dark. Samples were diluted 4-fold with 100 mM Ammonium Acetate, pH 8.9, and digested with Sequencing Grade Modified Trypsin (Promega) at a ratio of 1:100 (trypsin to total protein), overnight at RT. Following digestion, peptides were desalted and concentrated using Sep-Pak Plus C18 cartridges (Waters, cat. no. WAT020515) per the manufacturer’s recommendations. Samples were then dried by vacuum centrifugation, lyophilized, and stored at -80 C until further processing.
Phosphorylated samples were labeled with 8-plex iTRAQ reagents (AB Sciex). Lyophilized peptides derived from approximately 1 million cells were resuspended in 30 uL of dissolution buffer (0.5 M N(Et)3HCO3 pH 8.5–9). iTRAQ labels were resuspended in 70 uL of isopropanol and added to the peptide mixture. Samples were incubated at RT for 2 hours, combined, and dried overnight by vacuum centrifugation. The following day, samples were desalted and concentrated using Sep-Pak Vac 1cc (50mg) cartridges (Waters, cat. no. WAT054955) according to the manufacturer’s recommendations. Samples were then dried by vacuum centrifugation, lyophilized, and stored at -80 C until further processing.
Approximately 100 uL of packed Ni beads (Ni-NTA Superflow beads, Qiagen) were washed three times in water and stripped with 100 mM EDTA pH ~8.9 for 30 min. Beads were then washed three times with water and once with 80% ACN in 0.1% trifluoroacetic acid. Lyophilized iTRAQ samples were resuspended in 1.5 mL of ACN in 0.2% TFA and incubated with prepared beads for 1 hour at RT. Beads were then washed three times with 80% ACN in 0.1% TFA, and phosphopeptides were eluted from beads with 2 incubations in 75 uL of 1.4% Ammonia. Samples were then vacuum centrifuged down to ~20 uL. 2 uL of 200 mM ammonium formate pH 10 was added and samples were directly analyzed by mass spectrometry.
Peptide samples were loaded onto a first-dimension trap column (Waters Xbridge, C18, 10 uM particle size, 100 Å pore size, 4 cm packing length 150 uM column inner diameter). Online peptide separation coupled to MS/MS was performed with a 2D-nanoLC system (nanoAcquity UPLC system, Waters) and a Velos-Pro/Orbitrap-Elite hybrid mass spectrometer (ThermoFisher Scientific). Six discrete elutions were performed at 1.5 uL/min with 5mM ammonium formate pH 10 using increasing concentrations of ACN (1%, 3%, 6%, 15%, 25% and 44%) and diluted with 6 uL/min 0.1% formic acid (FA) prior to loading onto a second dimension trap column (Dr. Maisch ReproSil-Pur, C18, 5 uM particle size, 120 Å pore size, 4 cm packing length 150 uM column inner diameter) connected to an analytical column (Orochem Reliasil, C18, 3 uM particle size, 90 Å pore size, 20–25 cm packing length 50 uM column inner diameter) with an incorporated electrospray emitter. Peptide separation was achieved using a gradient from 3 to 80% (V/V) of ACN in 0.1% FA over 115 minutes at a flow rate of 200 nL/min. The mass spectrometer was operated in data-dependent mode using a Top 10 method. Full MS scans (m/z 300–2000) were acquired in the Orbitrap analyzer (resolution = 120,000), followed by high energy collision induced dissociation (HCD) MS/MS (fm/z 100–2000, resolution = 15,000) at a normalized collision energy of 35%.
MS data files were searched using the COMET [24, 84] algorithm and the output was imported into the Trans-Proteomic Pipeline [85] with the following parameters: variable oxidation of methionine, variable phosphorylation of Serine, Threonine, or Tyrosine, up to 4 variable modifications per peptide, fixed oxidation of Cysteine, and fixed iTRAQ labeling of Lysines and the N-terminus, maximum charge of 7. Peptide false discovery rate (FDR) was set to 5% for phosphorylation analysis. Peptide quantification based on the iTRAQ labels was determined using the LIBRA software embedded in the Trans-Proteomic Pipeline. Phosphopeptides were normalized to an internal control peptide (VNQIGpTLSESIK) from the enolase digest containing phosphorylated peptides. For each biological replicate, 2 technical replicates were run. A total of 12 fractions for the phosphoproteome, 12 fractions for the proteome and 2 fractions for phosphotyrosine- enriched runs were analyzed. 3,579 peptide spectra profiles were analyzed for the proteome, 4982 for the phosphoproteome including the phosphotyrosine residues. From the phosphotyrosine enrichment 1053 total spectra were analyzed. There was approximately 70% overlap of proteins identified from both technical runs and approximately 50% overlap in the phosphoproteome (S2A Fig).
Each peptide included in the analysis was identified in a minimum of 2 spectra, and each protein included in the analysis was identified by a minimum of two unique peptides. To deconvolve complex overlapping spectra profiles, we used the Hardklör algorithm [86, 87] in conjunction with the MassSpecUtil tool to merge spectra for iTRAQ analysis quantification. This was done to separate any possible overlapping isotopic envelope and providing better peptide identification. Comet search algorithm was used to identify the peptide spectra with a 5% FDR.
To assess significantly changing protein phosphorylation and abundance, peptide-spectrum matches that did not have an intensity of at least 10 in all channels were removed. All channels were median normalized, and the intensities were summed over all peptides of the same protein for each condition. The KSHV-specific effects were assessed by computing the log2(K/M) fold change for the means of the KSHV- (K) and mock (M)-infected biological replicates. A paired t- test was used to calculate the significance of the changes. Each technical replicate was analyzed independently.
TIME cells were Mock- or KSHV-infected with virus isolated from BCBL-1 cells, as described above, and incubated for 48 hours. Total mRNA was isolated from TIME cells using the NucleoSpin RNA kit (Machery-Nagle, Bethlehem, PA). mRNA was further concentrated and purified using the RNA Clean and Concentrator kit (Zymo Research, Irvine, CA). Purified mRNA samples were processed at the Benaroya Research Institute Genomics core facility and sequenced using an Illumina HiSeq 2500. Image analysis and base calling were performed using RTA v1.17 software (Illumina). Reads were aligned to the Ensembl's GRCh37 release 70 reference genome using TopHat v2.08b and Bowtie 1.0.0 [88, 89]. Counts for each gene were generated using htseq-count v0.5.3p9. The data have been deposited in NCBI's Gene Expression Omnibus [90] and are accessible through GEO Series accession number GSE84237.
DNA sequences 1000bp upstream of annotated transcription start sites were downloaded from the Genome Reference Consortium at http://hgdownload.cse.ucsc.edu/goldenpath/hg19/bigZips/upstream1000.fa.gz on May 29, 2015. Position-frequency matrices of 426 motifs were taken from a database of consensus motifs compiled from a variety of experimental techniques downloaded from http://meme-suite.org/meme-software/Databases/motifs/motif_databases.12.11.tgz. For each motif, FIMO software was used to find the top 1000 instances by p-value of each motif in the sequences. A motif was considered to flank a gene if an instance of the motif exists within 1000bp upstream of the gene’s transcription start site.
For each TF with at least 50 reads in all three mock-infected replicates or all three KSHV- infected replicates, a two-tailed Wilcoxon rank-sum test compared the change in expression post- infection of the genes flanked by of its motif locations to the change in expression of genes that are not. For the test, the genes were ranked by the significance and direction of their change in expression as analyzed by DESeq, where the highest-ranking genes were associated with low DESeq p-values and increased expression post-infection, the lowest-ranking genes were associated with low DESeq p-values and decreased expression, and the intermediate genes had high p-values. The Wilcoxon rank-sum tests compared the sum of the ranks of binding-site- flanked genes to a null normal model.
We conducted network analysis with the Prize-Collecting Steiner Forest (PCSF) algorithm from the Omics Integrator package, which uses msgsteiner for optimization [91]. PCSF identifies a sparse sub-network that connects the proteins highlighted by mass spectrometry with the TFs identified by the motif-based analysis. It assigns positive scores (prizes) to the proteins and TFs that reflect how relevant they are to KSHV infection and edge costs to protein-protein interactions that represent how trustworthy they are, with reliable edges receiving lower costs. The sub-network maximizes the cumulative prizes of the included proteins while minimizing the edges costs. It can include Steiner nodes, which are proteins that were not assigned prizes but form critical links between other proteins. We created an endothelial-specific weighted interaction network by integrating our RNA-seq data with the iRefIndex PPI network [26].
The PCSF algorithm requires protein prizes that quantify how relevant they are to the biological process of interest and PPI edge costs that describe how reliable they are. To calculate protein prizes, we scaled the p-values from the proteomic and phosphoproteomic technical replicates into the range [0, 1] by computing -log10(p-value), subtracting the minimum value across all proteins, and dividing by the maximum value across all proteins. Each proteomic and phosphoproteomic replicate was scaled independently. For the TF prizes, we transformed the Wilcoxon rank-sum q-values as -log10(q-value) but did not rescale them. Visualizing the histograms of the proteomic and TF scores revealed that the proteomic scores would dominate the TF scores if the TF scores were rescaled. For each protein, we then selected the maximum score from the two proteomic technical replicates, two phosphoproteomic technical replicates, and TF motif scores, which produced prizes for 3080 unique proteins.
We used the iRefIndex (version 13.0) PPI network [26]. The iRefIndex database aggregates PPI from multiple primary interaction databases such as BioGRID [92], DIP [93], HPRD [94], and IntAct [95]. All edges represent direct, experimentally-detected physical interactions between two proteins as opposed to predicted PPI or other types of functional relationships. We calculated edge costs between 0 and 1 based on the interaction metadata such as the interaction type, experimental assay, and number of supporting publications as in Ceol et al. [96]. PPIs identified using reliable, low-throughput experiments (for example, co- crystallization) are assigned much lower costs than interactions detected in large-scale screens. Interactions reported in multiple publications similarly receive lower costs. Low cost edges are more likely to be selected by PCSF so the PCSF subnetwork preferentially includes trustworthy protein-protein relationships. We created an endothelial-specific network by removing all unexpressed genes from the iRefIndex PPI network, which initially contains interactions from many types of human cells and tissues. Originally, the network contained 175854 interactions among 15404 proteins. After filtering genes that were not expressed at 50 counts or greater in all six of the RNA-seq replicates, the endothelial-specific network contained 121059 interactions among 9489 genes. To select PCSF parameters, we performed a grid search testing all combinations of β from 0.25 to 5.0 with a step size of 0.25, μ from 0 to 1.0 with a step size of 0.005, and ω from 0.5 to 3.0 with a step size of 0.5. Under some parameter combinations, the hub node ubiquitin C (UBC) was directly connected to a large portion of genes in the network so we discarded all networks containing UBC. The parameters β = 4.75, μ = 0.02, and ω = 2.5 produced the largest network without UBC, and we used these parameters for further analysis. We ran PCSF 1000 times with these parameters and additionally set r = 0.01 to add random noise to the edge costs. The multiple executions of PCSF with random noise were used to identify parallel paths between proteins in the different runs. Our final network was the union of the 532 of these 1000 PCSF networks that did not contain UBC. The union network contained 1253 interactions among 734 proteins, of which 44 were Steiner nodes. For visualization, we calculated edge frequency as the fraction of the 532 PCSF networks that contain a particular edge. The edge thickness in the network figures reflects edge frequency in the collection of PCSF networks, not the original interaction confidence score in the PPI network.
To assess whether the PCSF subnetworks are specific to KSHV infection, we ran PCSF 1000 times with the same parameters and randomized protein prizes. Each random run reassigned the observed KSHV protein prizes to random proteins in the PPI network. We removed the random PCSF outputs that contained the hub node UBC and computed node and edge frequencies with the 131 remaining forests.
We supplemented the PCSF union network with KSHV-human protein-protein interactions obtained from VirHostNet 2.0 [56, 97] (downloaded January 7, 2016). We considered only latency-related KSHV genes as defined by Davis et al.: K1 (K1_HHV8P), K2 (VIL6_HHV8P), K12A (K12_HHV8P), K12B, K12C, ORF71 (VFLIP_HHV8P), ORF72 (VCYCL_HHV8P), and ORF73 (ORF73_HHV8P). We queried VirHostNet and the Davis et al. interactions for all host-virus PPI between these latency-related KSHV genes and any human protein in our PCSF network. We then added all the relevant KSHV-human interactions to our network figures and did not use the PCSF algorithm to filter these edges.
We used WebGestalt [98] to identify KEGG pathways [47] that are enriched for proteins in our PCSF network. Because the predicted network can only contain proteins from the initial iRefIndex interaction network, we used all proteins in the protein-protein interaction network as the reference set. We required a minimum overlap of 5 proteins and used the Benjamini and Hochberg multiple hypothesis test correction (FDR < 10%). To visualize the network regions that are relevant to the enriched KEGG pathways we used Cytoscape [99] to select all proteins in the enriched pathway and their directly connected neighbors in the PCSF network. We then included all KSHV- human PPI that involve the KEGG pathway members and their PCSF neighbors.
|
10.1371/journal.pbio.1002082 | Estimating the Life Course of Influenza A(H3N2) Antibody Responses from Cross-Sectional Data | The immunity of a host population against specific influenza A strains can influence a number of important biological processes, from the emergence of new virus strains to the effectiveness of vaccination programmes. However, the development of an individual’s long-lived antibody response to influenza A over the course of a lifetime remains poorly understood. Accurately describing this immunological process requires a fundamental understanding of how the mechanisms of boosting and cross-reactivity respond to repeated infections. Establishing the contribution of such mechanisms to antibody titres remains challenging because the aggregate effect of immune responses over a lifetime are rarely observed directly. To uncover the aggregate effect of multiple influenza infections, we developed a mechanistic model capturing both past infections and subsequent antibody responses. We estimated parameters of the model using cross-sectional antibody titres to nine different strains spanning 40 years of circulation of influenza A(H3N2) in southern China. We found that “antigenic seniority” and quickly decaying cross-reactivity were important components of the immune response, suggesting that the order in which individuals were infected with influenza strains shaped observed neutralisation titres to a particular virus. We also obtained estimates of the frequency and age distribution of influenza infection, which indicate that although infections became less frequent as individuals progressed through childhood and young adulthood, they occurred at similar rates for individuals above age 30 y. By establishing what are likely to be important mechanisms driving epochal trends in population immunity, we also identified key directions for future studies. In particular, our results highlight the need for longitudinal samples that are tested against multiple historical strains. This could lead to a better understanding of how, over the course of a lifetime, fast, transient antibody dynamics combine with the longer-term immune responses considered here.
| Host immunity against seasonal influenza viruses influences the emergence of new virus strains, the size and severity of “flu” epidemics, and the effectiveness of vaccination programmes. However, the specific factors that shape the immune response of a single human to a particular strain are little understood because individual infections and the development of immunity over a lifetime in that person are rarely observed directly. To determine the aggregate effect of a lifetime of influenza infections on host immunity, we developed a mathematical model that captures the specific strains with which an individual has been infected and for the corresponding antibody response, the relative contribution of boosting, cross-reactivity, and antigenic seniority to its neutralising ability. Combining the model with data from a survey in southern China that examined antibody levels against nine different influenza strains from 1968 to 2009, we revealed key components of the immune response to influenza virus infection, and obtained estimates of the frequency of influenza infection and the ages at which infection occurred. Our results suggest that “antigenic seniority”, whereby strains encountered earlier in life gain more “senior” positions in the immune response, and short-lived cross-reactivity between different strains are important components of the immune response and, therefore, could shape the evolution and emergence of influenza viruses.
| The immunity of a host population against specific influenza A strains can influence a number of important biological processes. It can affect the emergence of new virus strains, and hence shape the evolution of the disease [1,2]. It can also influence the size and severity of a pandemic [3–6], and the effectiveness of vaccination programmes [7].
There are two main ways to measure the adaptive immune response against influenza viruses [8]. In microneutralisation assays, a mixture of virus and diluted serum is used to infect cell cultures; the titre is the highest dilution for which virus infection is blocked. Microneutralisation titres therefore measure the overall neutralising antibody response. Such a response can include several components. Some antibodies are specific to antigenic sites on the globular head of the haemagglutinin (HA) surface protein. These sites are highly variable: the HA undergoes frequent mutation, enabling the virus to escape existing antibody responses [9]. There is also evidence that antibodies target conserved epitopes on the stalk of the HA protein or the neuraminidase (NA) surface protein [10–12]. Alternatively, haemagglutination inhibition (HAI) assays measure the extent to which antibodies inhibit binding of the HA protein to red blood cells. Whereas microneutralisation titres likely capture more of the total antibody response, the HAI assay is a more sensitive measure of antibodies that are specific for antigenic sites on the head of the HA protein [13].
The ability of human sera to neutralise current and historical influenza strains exhibits substantial variation between individuals and with age [10,13–15]. These patterns are likely to be influenced by a number of factors. First, neutralisation titres to a particular strain depend on the immune response following exposure to that virus: after infection or vaccination, the immune response to a particular virus can be boosted [5]. Although the initial response may decay to a lower level after a short period of time [16,17], there is evidence that the subsequent level of response can persist for several decades [18].
Observed titres can also depend on the order and number of influenza infections. Francis [14] coined the term “original antigenic sin” to describe the phenomenon by which HAI titres to the first influenza infection of a lifetime were apparently higher than titres to other strains. Upon subsequent infection, it has been suggested that the original response can be enhanced [14,19–21] and the antibody response to the new strain reduced [22–26]; for original antigenic sin to occur, there is evidence that the original and new strain must be antigenically related [26,27]. Recent work has refined the original antigenic sin hypothesis, proposing that serological patterns should be described in terms of “antigenic seniority” [13,15]. As with original antigenic sin, the primary infection gains the most “senior” position in the immune response, but—as a key refinement to the original sin hypothesis—the hierarchy of responses continues with each subsequent infection, as each strain takes a less senior position in the response.
As well as boosting and original antigenic sin/antigenic seniority, observed serological responses can also depend on cross-reactivity between strains and temporal waning of responses. Even if hosts have not been exposed to a given strain, they can have a raised titre against the virus if the test strain is similar to those already encountered [5,18]. Establishing the contribution of different mechanisms to neutralisation titres remains challenging, however, because the aggregate effect of immune responses over a lifetime are rarely observed directly [13,24]. Moreover, observed titres not only depend on the relationship between infection and immune response: they are also influenced by the specific strains a host has been infected with. It has been suggested that certain age groups are infected more often than others [28,29], but the true frequency of influenza infection cannot be easily measured [30].
To explore the effects of past infections and subsequent immune responses on observed microneutralisation titres, we fitted a mechanistic model of within-host serological dynamics to data from a cross-sectional survey based in southern China [31]. In the study, 151 individuals’ sera were tested against a panel of nine different influenza A(H3N2) strains isolated between 1968 and 2008. Six of these strains corresponded to representative viruses from every second “antigenic cluster” that appeared between 1968 and 2003; in total there were 11 such clusters of antigenically similar strains during this period [32]. The other three test viruses were strains that circulated in southern China between 2003 and 2008.
We used the mechanistic model to assess the relative contribution of boosting, cross-reactivity, and antigenic seniority to observed neutralisation titres, and estimated key immunological parameters. We also estimated which specific strains each individual had been infected with, and hence assembled infection histories for each individual in the study population. This made it possible to calculate the frequency of infection for influenza A(H3N2) in the host population, and to establish how the infection rate varied with age.
Study protocols and instruments were approved by the following institutional review boards: Johns Hopkins Bloomberg School of Public Health, University of Liverpool, University of Hong Kong, Guangzhou No. 12 Hospital, and Shantou University. Written informed consent was obtained from all participants over 12 y of age, and verbal assent was obtained from participants 12 y of age or younger. Written permission of a legally authorised representative was obtained for all participants under the age of 18 y.
Participants were recruited from five study locations, with 20 households randomly selected in each location (further details given in Lessler et al. [31]). Participants’ sera were tested against nine representative influenza A(H3N2) strains using a virus neutralisation assay. The strains included six vaccine strains: A/Hong Kong/1/1968, A/Victoria/3/1975, A/Bangkok/1/1979, A/Beijing/353/1989, A/Wuhan/359/1995, and A/Fujian/411/2002. Three strains that circulated in southern China in the years preceding the study were also tested: A/Shantou/90/2003, A/Shantou/806/2005, and A/Shantou/904/2008. Titres were measured using serial 2-fold dilutions from 1:10 to 1:1,280 in duplicate. In our analysis, we represented the results in terms of log neutralisation titres. A log titre of c corresponded to a dilution of 10 × 2(c−1). Hence, there were nine possible log titres: the lowest was 0, which corresponded to a dilution <1:10; the highest was 8, which corresponded to a dilution of 1:1,280.
We took an “epochal” view of infection [32,33], with individuals either infected or not during each antigenic epoch; we assumed there were 14 such epochs between 1968 and 2008. We modelled serological titres by assuming that the mean neutralisation titre to a specific strain depended on both individual infection history and a combination of serological mechanisms. We considered four specific mechanisms: boosting from infection with the test strain, cross-reactivity from antigenically similar strains, boosting of earlier responses as a result of subsequent infection, and suppression of subsequent responses as a result of prior immunity. The final two mechanisms have been suggested as potential explanations for observed patterns of antigenic seniority [13,15].
We included the four main mechanisms in the model as follows. Suppose an individual has an infection history that consists of a set of strains X (note that we do not distinguish between live infection and vaccination in the model). We assumed that if an individual had been infected with only one strain, they would exhibit a fixed log titre against that strain, controlled by a single parameter, μ. In the absence of antigenic seniority or cross-reactivity, the individual would therefore have titre equal to μ for every strain in their infection history, and zero for all other strains (Fig. 1A). However, if the individual had been infected with more than one strain, titres against earlier strains could be higher than those against later strains as a result of antigenic seniority.
Two mechanisms have been proposed to explain observed patterns of antigenic seniority: previous responses might be boosted as a result of subsequent infections, or subsequent responses might be reduced as a result of previous immunity [15]. To evaluate the contribution of these two mechanisms, we specified the model so that—depending on parameter values—both, one, or neither mechanism could contribute to measured titres. During the fitting process, model outputs could therefore be compared to observed serological data to establish which mechanism(s) were most plausible given the data.
First, we assumed each infection could boost titres against strains encountered previously by a parameter τ1 (Fig. 1B). Hence the titre μ was scaled by a factor s1(X, j) = (1 + τ1)|X|−Nj where Nj is the number of the strain in the infection history (i.e., the first strain is 1, the second is 2, etc.) and |X| is the total number of infections. If τ1 = 0, then there was no boosting as a result of subsequent infection. This mechanism, in which observed titres to a particular strain depended on the number of subsequent infections, was also proposed by Miller et al. [13] following a longitudinal study of influenza A infections.
Second, we assumed prior immunity could reduce observed titres against strains encountered later in life (Fig. 1C). The titre against a particular strain would therefore be scaled by a factor s2(X,j)=e−τ2(Nj−1). Here, observed titres to each strain depended on how many infections had occurred previously. When τ2 = 0, prior infections did not lead to reduced responses against later strains. When τ2 was large, the formulation was equivalent to a model of original antigenic sin, in which immunity from the primary infection suppressed all subsequent responses [28,34].
Finally, we incorporated cross-reactivity by assuming that mean titre against a specific strain was equal to the sum of cross-reactive responses to all strains in an individual’s infection history. We assumed that the contribution made by each strain depended on the temporal distance between the strain in the infection history and the test strain (Fig. 1D). The level of cross-reaction between a test strain j and infecting strain m was given by d(j, m) = e−σ|tm−tj|, where |tm − tj| was the number of years between strains j and m, and σ was a parameter to be fitted. If σ was large, it was equivalent to having no cross-reactivity between strains.
To combine the four mechanisms in the model, we assumed that the log titre individual i has against a strain j was Poisson distributed with the following mean:
λij=μ∑m∈Xd(j,m)s1(X,m)s2(X,m)
(1)
As d(j, m), s1(X, m), or s2(X, m) could equal 1 for certain parameter values, the model was capable of omitting certain mechanisms if necessary.
We also accounted for potential observation error by assuming that there was a uniform probability of observing a titre different to the true one. Hence, the likelihood of observing titre cj against test strain j was equal to the sum over all possible true titres:
L(cj)
=
μ
∑kP(
true titre isk)×P
(observecj|true titre isk)
(2)
We estimated model parameters using Markov chain Monte Carlo (details in S1 Text; dataset and model outputs in S1 Data).
As a sensitivity analysis, we also included the waning of antibody responses in our model. This was achieved by modifying Equation 1:
λ
ij
=μ∑m∈X
d(j,m)s1(X,m)s2(X,m)
e−wtm
(3)
where w was a waning parameter that we fixed. The formulation meant that waning reduced titres to strains in the infection history by a factor e−w per year. If w = 0, then we recovered the model given by Equation 1.
The model captured the observed age distribution of titres for each of the nine test strains. Fig. 2 shows that when splines were fitted to the data (red line) and the model (blue line), there was a similar pattern with age. For time periods that were well represented by test strains in the data, such as 2003–2008, the model captured both the average pattern of titres with age and the variability in titre levels for strains in that period (Fig. 2G–I). When test strains circulated further away in time from neighbouring data points, the estimates did not match the magnitude of titre in the serological data as closely. For example, the model overestimated titre levels against A/Victoria/1975 (Fig. 2B), which was two antigenic clusters from neighbouring test strains. However, even for time periods that are less well represented in the test strains, such as 1968–1989, the model captured the correct average trend for titre levels. While both the model and data generally exhibited more variation in titre levels for more recent strains (S1 Fig.), across all strains, 87% of model estimates were within two dilutions of the observed titre (S2 Fig.).
We found evidence that antigenic seniority and quickly decaying cross-reactivity were important components of the immune response, and obtained measurements for the immunological processes outlined in Fig. 1. Parameter estimates are shown in Table 1. The boosting parameter suggests that primary infection resulted in a log neutralisation titre of around 3 (corresponding to a dilution of 1:40). Our estimate for the exponential decay in cross-reactivity with time was 0.29, suggesting that strains circulating 2.4 y apart had only 50% cross-reactivity. The antigenic seniority parameter controlling suppression of subsequent responses was 0.06, which implies that the response to each new infection was scaled by a factor 0.94 compared with the response to the previous infecting strain. In contrast, we estimated the parameter that controlled boosting of prior responses to be zero.
To further investigate our finding that boosting from antigenic seniority was not required in the model, we compared the observed titre against the earliest strain in each individual’s infection history with their estimated total number of infections. If in reality subsequent infections boost earlier responses, we would expect the titre against the earliest strain to be larger for individuals who have been infected with numerous strains. However, we did not find a significant correlation between the two variables, suggesting that boosting from later infections had little effect on observed neutralisation titres to the first strain (S3 Fig.).
As there may be a trade-off between the short-term and long-term dynamics of influenza infection, we also examined how our estimate for boosting from antigenic seniority changed if we assumed that titres could wane over time after the initial infection [35–38]. First we tested whether the boosting parameter τ1 could be robustly measured if we included waning in the model. We found that there was a trade-off between the two processes in the model: a high degree of boosting was balanced by a larger amount of waning. This suggested that these two mechanisms were not distinguishable given our cross-sectional data (S4 Fig.). We therefore fixed the degree of waning, and estimated the other parameters. Our estimates for cross-reactivity and measurement error remained consistent. We still found evidence for suppression of response via antigenic seniority, and if we assumed an increased amount of waning per year, we obtained a non-zero estimate for antigenic seniority boosting (S1 Table).
In addition, we examined whether broadly cross-reactive antibodies might contribute to observed titres, as has previously been observed during influenza infection [12,39,40]. We extended the original model described in Equations 1 and 2 to incorporate a fixed amount of broad cross-reaction between distant strains (S5 Fig.; details in S1 Text). However, when we fitted this model to data, the parameter estimate for broad cross-reactivity was zero (S2 Table). We therefore recovered the original model formulation and parameter estimates, which suggested that broad cross-reactivity was not required to reproduce the observed data.
We were able to better understand how the model reproduced observed titre values for individual people by considering specific examples. If a participant was infected with a small number of strains in the model, observed titres were predominantly the result of boosting (Fig. 3A). The sharp decay in cross-reactivity in the model meant that a low titre was produced against strains that circulated a number of years before or after the infecting strain. When participants were infected with several similar strains, the expected titre was the sum of contributions from boosting with the test strain and cross-reactivity from related ones (Fig. 3B). Cross-reactivity also led to a high titre against nearby strains, even if the strains were not in the infection history. However, antigenic seniority meant that the contribution to boosting from infection declined with each strain encountered (S6 Fig.). As a result, much of the expected titre against the first strain came from boosting with that strain, whereas titres to later strains have a larger contribution from cross-reactivity (Fig. 3C). Model residuals for these three selected examples were representative of the study population (S7 Fig.).
Because we inferred infection histories for each individual, it was also possible to generate estimates for frequency of infection. Fig. 4A shows the number of A(H3N2) influenza infections per decade at risk, based on the estimated infection histories. The rate of infection decayed initially with age, but was relatively flat after age 30 y, implying that above a certain age, individuals were infected with similar frequency. Fig. 4B shows the distribution of time between two sequential infections, conditional on individuals’ having had at least two infections.
We also tested the ability of the model to predict unseen data. We omitted each of the nine test strains in turn, refitted the model to the remaining eight strains, and used our parameter estimates to predict the omitted data. S8 Fig. shows that although the model captured the general pattern of measured serological response for many strains, the predictive power was highest when test strains were close together in time (S8G–I Fig.).
We have examined how past infections with influenza A(H3N2) strains influence observed cross-sectional neutralisation titres. We found that “antigenic seniority” and quickly decaying cross-reactivity were important components of the immune response. The order in which an individual is infected with influenza strains was therefore important in dictating observed titres to a particular virus: titres appeared to be the result of a combination of strain-specific boosting, cross-reactivity, and suppression of subsequent responses as a result of antigenic seniority (Fig. 5).
Our results emphasise the importance of understanding how currently unobserved mechanisms shape the dynamics of influenza for individuals over the course of their lifetime. Traditionally, analysis of serological data has been descriptive rather than mechanistic. It has therefore been challenging to distinguish between different hypotheses that could describe observed patterns. In particular, we evaluated two antigenic seniority mechanisms that have been proposed as explanations for why individuals exhibit raised titres to strains encountered earlier in life [15]: earlier responses could be boosted by subsequent infections, or subsequent responses could be reduced as a result of prior immunity.
There is an apparent discrepancy between our main parameter estimates for antigenic seniority and empirical observations of boosting of antibodies to early infections. Our baseline results (in the absence of waning) suggest that while there is a reduction in the magnitude of response to later infections, the boosting component of antigenic seniority has a negligible effect on observed long-term titres. However, several studies have found evidence for boosting of existing responses following influenza infection [13,14,19–21].
There were two mechanisms in our model that could potentially lead to increased titres to previously encountered strains after subsequent infections. The first was cross-reaction: in our framework, individuals who had been infected with few strains (Fig. 3A) had lower titres than those who had been infected with several strains that are antigenically related (Fig. 3C). We assumed cross-reaction was symmetric in the model, and made the same contribution to titres regardless of infection order.
In contrast, we did not find evidence for boosting of earlier responses as a result of antigenic seniority. In order to investigate this further, we considered the possibility of waning antibodies as a sensitivity analysis. Unfortunately, with cross-sectional data, the boosting and waning processes were not identifiable. Therefore, we assumed a plausible single overall rate of antibody waning and found evidence for boosting as part of an antigenic seniority process. However, although it was possible to force boosting into the model, we suggest that the identifiability issues between boosting and waning and the discrepancy between the model parameters and observed boosting are both consequences of the different timescales on which these immunological processes occur.
Boosting and waning are both likely to contribute to the hierarchal nature of antibody responses to influenza. But while waning of elevated titres has been observed over periods of less than a year [35–38], it is not clear precisely how boosted antibody responses persist over time in the absence of infection [41]. Based on our model results, we suggest that repeated boosting of long-lived antibody responses in the absence of waning is unlikely: such a process would lead to either extremely high titres in older individuals or very low rates of infection, neither of which seem credible. Therefore, in essence, we believe our model provides a plausible description of the acquisition of a stable set of persistent antibodies.
The apparent discrepancies between observed boosting and relatively low titres to historical strains further highlight the need for studies that take repeated measurements of the serological response of individuals against a panel of historical influenza strains [42]. With such data, the mechanistic model presented here could be expanded to explore both the short- and long-term dynamics of influenza immunity. This would help elucidate the precise role of boosting and suppression in antigenic seniority.
We found that cross-reactivity decayed quickly with time, with a half-life of 2.4 y. Hence, there was little cross-reaction between influenza A(H3N2) strains that circulated several years apart. We also considered a model that included broad cross-reactivity between strains, but when we fitted this model to data, the parameter estimate for broad cross-reactivity was zero, indicating that this additional component was not necessary to reproduce observed serological patterns. However, there is evidence that individuals are capable of producing broadly cross-reactive antibodies following infection with a pandemic strain [12,40], and that individuals can exhibit a longitudinal increase in neutralising titres against pandemic strains that are no longer circulating [13]. Again, this highlights the need for longitudinal studies of serological responses against a panel of historical influenza strains. Such data would make it possible to jointly examine the contribution of broad and strain-specific immune responses, and understand how cross-reactive antibodies and antigenic seniority influence observed serological patterns over multiple timescales.
As well as comparing the effects of different immune mechanisms, we estimated infection histories for each individual in our study population. We used this information to measure how frequency of infection varied with age. Although infections became less frequent as individuals progressed through childhood and young adulthood, they occurred at similar rates for individuals above age 30 y (Fig. 4A). It has been suggested that influenza transmission is driven by intense social contacts among younger age groups [43]. The decline in frequency of infection with age may be the result of age-specific differences in social behaviour. A study conducted in the same area of southern China as our serological survey found intense mixing within the under-20-y age groups, which could mean the force of infection was higher within these groups [44]. Unfortunately, we had limited serological data for very young individuals (the youngest participant in the study was aged 7 y); it would be interesting to see how the frequency of infection changes from birth through childhood.
There are some additional limitations to the work described here. We made no prior assumptions about different age groups’ rate of infection, and hence infection history, in the model. An important next step would be to develop an approach that could measure force of infection from cross-sectional data [45]. This could be explored using a model that accounted for population transmission dynamics as well as serological responses. Moreover, we examined serological data from only 151 participants in southern China. It would therefore be helpful to test similar models of serodynamics against observed titres in other populations [42]. We also focused on responses against a panel of A(H3N2) influenza strains. Unlike group 1 influenza viruses such as A(H1N1), A(H1N1p), and A(H2N2), no group 2 viruses other than A(H3N2) have caused a pandemic; it has been suggested that this is why antibody titres specific to HA stalks might be lower for group 2 viruses than for the more antigenically diverse set of group 1 viruses that have circulated in humans [13]. We also do not distinguish between live infection and vaccination in the model; different routes of exposure could influence the process of antigenic sin/seniority in different ways [12,24].
The model we present offers a novel method for simultaneously investigating immune responses and past infections. Studies looking at the antigenic relationship between different influenza strains typically examine cross-reactivity using ferret sera [9,32]. However, the transmission dynamics of influenza is mediated not just by antigenic change in the virus, but also by underlying immunity in the host population. To analyse the evolutionary trajectory of influenza viruses using human sera, it would be necessary to account for past infections, and how these shape the immune response. We propose that a model of serodynamics, as outlined in this paper, would provide the theoretical foundation required to tackle this problem.
Our results also have implications for the analysis of control measures. By considering how a lifetime of infection shapes cross-sectional sera, we have measured the relative importance of different immune mechanisms and past infections in measured serological responses to influenza. As well as influencing the evolution of influenza, such mechanisms could have an impact on the effectiveness of vaccination programmes [7].
We found that the model had limited capacity to accurately predict the magnitude of observed titres to strains that circulated several years before or after the strains to which the model was fitted (S8 Fig.). This is likely the result of the fast decay in cross-reactivity between strains over time. However, the model could generally predict age-specific trends in titres to unencountered strains, even those far from the strain used for fitting. It also reproduced the observed titre levels accurately when fitted strains were close to the test strains in time. This suggests that sufficient representation of past strains, perhaps from every antigenic epoch, would be needed to reproduce all responses accurately. Further, our results were based on neutralisation titres. Similar results are likely to be obtained using HAI, albeit with lower specificity for lower titre values [15,42]. Also, future studies may be able to take advantage of emerging immunological technology based on high throughput protein microarrays [46] and sequence-based measures of B cell diversity [47].
Using a model of cross-sectional serological responses, we have assessed the relative importance of different immune mechanisms and the timing of influenza infection in shaping observed neutralisation titres across the lifetime of an individual. To our knowledge, these two key factors have not previously been combined to fit immunological data. As well as characterising different aspects of the immune response, we have generated individual-level estimates of the frequency and age distribution of influenza infection from cross-sectional serological data. These results demonstrate the value of interpreting immune responses in the context of a lifetime of infection. Integrating the life course of immunity into future analyses of influenza dynamics could therefore lead to a better understanding of population susceptibility and the potential transmissibility of new seasonal strains.
|
10.1371/journal.pcbi.0030180 | What Are Lightness Illusions and Why Do We See Them? | Lightness illusions are fundamental to human perception, and yet why we see them is still the focus of much research. Here we address the question by modelling not human physiology or perception directly as is typically the case but our natural visual world and the need for robust behaviour. Artificial neural networks were trained to predict the reflectance of surfaces in a synthetic ecology consisting of 3-D “dead-leaves” scenes under non-uniform illumination. The networks learned to solve this task accurately and robustly given only ambiguous sense data. In addition—and as a direct consequence of their experience—the networks also made systematic “errors” in their behaviour commensurate with human illusions, which includes brightness contrast and assimilation—although assimilation (specifically White's illusion) only emerged when the virtual ecology included 3-D, as opposed to 2-D scenes. Subtle variations in these illusions, also found in human perception, were observed, such as the asymmetry of brightness contrast. These data suggest that “illusions” arise in humans because (i) natural stimuli are ambiguous, and (ii) this ambiguity is resolved empirically by encoding the statistical relationship between images and scenes in past visual experience. Since resolving stimulus ambiguity is a challenge faced by all visual systems, a corollary of these findings is that human illusions must be experienced by all visual animals regardless of their particular neural machinery. The data also provide a more formal definition of illusion: the condition in which the true source of a stimulus differs from what is its most likely (and thus perceived) source. As such, illusions are not fundamentally different from non-illusory percepts, all being direct manifestations of the statistical relationship between images and scenes.
| Sometimes the best way to understand how the visual brain works is to understand why it sometimes does not. Thus, visual illusions have been central to the science and philosophy of human consciousness for decades. Here we explain the root cause of brightness illusions, not by modelling human perception or its assumed physiological substrate (as is more typically done), but by modelling the basic challenge that all visual animals must resolve if they are to survive: the inherent ambiguity of sensory data. We do this by training synthetic neural networks to recognise surfaces under different lights in scenes with naturalistic structure. The result is that the networks not only solve this task robustly (i.e., they exhibit “lightness constancy”), they also—as a consequence—exhibit the same illusions of lightness that humans also see. In short, these synthetic systems not only get it right like we do, but also get it wrong like we do, too. This emergent coincidence strongly provides causal evidence that illusions (and by extension all percepts) represent the probable source of images in past visual experience, which has fundamental consequences for explaining how and why we see what we do. The study also suggests the first formal definition of what an illusion is: The condition in which the actual source of a stimulus differs from its most likely source.
| Understanding how we generate accurate perceptions of surfaces is often best informed by understanding why we sometimes do not. Thus, illusions of lightness (and colour) are essential tools to vision research. In many natural environments, light levels vary across space and over time. It is important to be able to perceive surfaces independently of this varying light intensity (and vice versa) in order to forage or predate successfully, for example. (By reflectance, we mean the proportion of incident light reflected by a surface; lightness is the perceived reflectance of a surface; brightness is the perceived intensity of light reaching the eye; and luminance is the actual intensity of the light that reaches the eye with respect to the sensitivity of the human visual system.)
A number of models of lightness perception have been proposed, but most of these fail to deal with complex stimuli or only demonstrate a narrow range of behaviours. For instance, one well-known heuristic model predicts human lightness perceptions by first subdividing stimuli into multiple “local frameworks” based on, for instance, junction analysis, and co-planarity as well as other classic gestalt factors. Then, within each framework, the ratio of a patch's intensity and the maximum intensity in that patch's local framework is used to predict the reflectance, combining a “bright is white” and a “large is white” area rule [1]. These rules are well-defined and effective for simple stimuli (e.g., with two nonzero luminance regions), but the application of the rule has not been studied for more complex images [1]. Indeed, it is hard to see how such a model could be applied to even moderately complex stimuli, much less natural scenes under spatially heterogeneous illumination, without extremely complex edge-classification rules that are as yet undefined. Furthermore, such human-based heuristics provide little insight into the physiological and/or computational principles of vision that are relevant to all visual animals.
More computational approaches, on the other hand, are less descriptive, more quantitative, and make fewer assumptions. For example, artificial neural networks (ANNs) have been trained to extract scene information, such as object shape and movement, from simple synthetic images [2,3]; and a statistical approach using Gibbs sampling and Markov random fields has been used to separate reflectance and illumination from simple images [4]. Most such models, however, are unable to explain brightness contrast and assimilation (e.g., White's illusion) simultaneously without recourse to one or more adjustable weighting factors. One approach that can is the Blakeslee and McCourt filter model [5]. By applying a set of filters (specifically, a bank of oriented difference of Gaussians filters, or ODOG), the model produces results that correspond closely to psychophysical results on a wide range of illusory stimuli. The same model, however, fails to predict the asymmetry of brightness contrast, where darker surrounds cause larger illusions than equally lighter surrounds, as we discuss later. “While these asymmetries are not captured by the ODOG model as it is presently implemented, permitting different gain parameters to be applied to the outputs of independent on-channels and off-channels would constitute a logical first step toward accommodating these differences” [6]. It is also important to stress that the model does not attempt to predict the reflectance of surfaces, only the perceived brightness of a stimulus, and therefore is unable to explain lightness constancy in more natural scenes under spatially heterogeneous illumination. Related machine vision work includes the separation of luminance changes into those caused by shading (including the slant of the surface and direction of incident light), and those caused by paint on the surface, using filters and a mixture of Gaussians [7]; and a localised “mixture of experts” and a set of multiscale filters has been used to extract the intrinsic components of an image, including “de-noising” it [8]. However, these studies do not attempt to explain the human perception of lightness or illusions. Thus, explanations as to why and how we see lightness illusions remain incomplete.
Here we take a different approach to rationalising human illusions and, by extension, lightness perception generally. Rather than modelling human perception or known primate physiology—as is typical of most models—we instead model the empirical process by which vision resolves the most fundamental challenge of visual ecology: the inherent ambiguity of visual stimuli. We make no assumptions about particular physiology or cognition, but instead model the process of development/learning from stimuli with feedback from the environment. This is analogous to the experiential learning of any animal whose behaviour is guided visually, and which must learn to resolve perceptual ambiguity in order to survive.
Fifty ANNs were trained using backpropagation to predict the reflectance of surfaces in synthetic scenes, an example of which is shown in Figure 1A. Each scene consisted of a 3-D matrix of 400 matte surfaces (R) under spatially heterogeneous patterns of illumination (I). As is the case for the human visual system, the trained ANNs did not have direct access to the scenes' reflectance or illumination, but only the product of the two (R • I = S) at each point in space—thus, the luminance stimulus (S) in Figure 1D represents the product of the surface reflectance matrix in Figure 1B and the illumination matrix in Figure 1C. The task was to predict the source reflectance (R) of the stimulus (S) at the centre of each scene without explicit knowledge of the surface's illumination (I).
Surface reflectance matrices (Figure 1B) and illumination matrices (Figure 1C) were created using the “dead-leaves” algorithm, which results in projected images with the same statistical properties as natural images [9]. In 20% of cases, a second surface layer with “gaps” in place of surfaces was placed “in front” of the first surface layer, under independent illumination, equivalent to viewing background objects beyond independently illuminated foreground objects, such as looking through the branches of a tree. See Methods for further details on the ANNs and the “dead-leaves” stimuli.
Note that the ANN training was supervised, meaning that the true target reflectance underlying each stimulus was used by the backpropagation algorithm to estimate errors during learning, which provided feedback for the ANNs. While backpropagation is not physiological in terms of its actual mechanics, the process of altering the network processing according the success and/or failure of its output is equivalent to a visual animal getting feedback from the environment according to the value of its response. Visual systems have evolved to aid survival by allowing animals to respond to the visual environment successfully. This does not necessarily require veridical percepts of the world, but we assume here that generating behaviours that are consistent with surface reflectance, along with other characteristics, will be useful to a visual animal. Animals that generate behaviours that preserve the similarities and differences between surfaces will typically receive some form of feedback from their environment, such as the reward of eating nutritious food or the penalty of eating noxious food. In the same way, “virtual robots” have been shown to develop a form of colour constancy without supervised learning in a visually ambiguous ecology [10]. This feedback can be modelled explicitly using artificial life [10] or reinforcement learning [11], but in this work our focus is not on learning algorithms themselves, but rather on what is encoded. We therefore ignore the temporal credit assignment problem, i.e., the problem of how an animal decides which of its past actions led to a particular reward or penalty. Instead, we consider other sources of uncertainty such as the ambiguity caused by heterogeneous light falling on varied surfaces.
Our emphasis on learning contrasts with “mechanistic” modelling approaches (in the sense defined in [12]), such as the “Anchored Filling-in Lightness Model” [13]. That model describes many of the visual effects discussed here, and more besides, based on neural and anatomical experimental data. It is not derived directly from ecological data in the way that ANNs' behaviour here is, and so cannot give a distal explanation as to why such visual behaviours are found. Similarly, the computational Bayesian approach in [12] uses a parametric model, whose form has been chosen manually, whereas the ANNs used here are a nonparametric model, derived entirely from the data.
After training, each ANN was tested with 10,000 novel images created in the same way as the images in the training set, and the ANN's prediction of the reflectance of each target patch was recorded. The average root-mean-squared (RMS) error for predicted reflectance to the novel test set was 0.171 (with a standard deviation of 0.0016) and the errors approximated a Gaussian distribution (Kolmogorov-Smirnov normality test; p ≈ 0). Thus, trained ANNs—like humans—were able to accurately and robustly predict the reflectance of the central surface from uncertain sensory data; i.e., the ANNs exhibited “lightness constancy” (see related work on depth processing [14], the evolution of visually guided behaviour in virtual robots [15], and distance perception [16]; and on perceiving colour constancy by estimating the illumination of a scene using higher-order statistics [17]). Robust response accuracy, however, varied according to the nature of the stimulus. When, for instance, the central target was viewed against a uniform background with uniform illumination (rather than against a fully articulated surround), the RMS error increased significantly to 0.20 (s.d. 0.015; t-test: p ≈ 0, n = 50). An equivalent decrease in lightness (and colour) constancy in low variance scenes is also evident in human perception [18–21]. The study here suggests that this is because increasing the number of surfaces in a scene (i.e., “articulation,” which is a subset of the more general phenomenon of “cue-combination”) narrows the distribution of possible sources of a stimulus, which has been suggested previously in human studies but never tested directly [20,21].
A basic aspect of human lightness and brightness is that these phenomena do not always accord with stimulus intensity, which is to say we see illusions. The most basic, well-known, and most thoroughly studied illusion is “brightness contrast,” where a central target against a lighter background appears darker than the same target viewed against a darker background (as will be evident to the reader when viewing the two small patches at the middle of the light and dark surrounds in Figure 2A). To test whether trained ANNs also behave in accordance with this illusion, ANNs were presented with “hand-made” stimuli, in which a target stimulus of 0.5 was embedded on uniform surrounds that varied from 0 to 1. The darkest surrounds lead to an average overestimation error of 0.36, whereas the lightest surrounds lead to an average underestimation error of 0.17. Thus, trained ANNs did indeed exhibit brightness contrast. What is more, the data show that they also exhibited an asymmetry in the relative effects of the darker versus lighter surrounds, with the darker surround “carrying” most of the illusion. Remarkably, this latter asymmetry is also evident in human perception [1,6,22]. The anchoring model [1] explains this in terms of a weighted sum of global and local anchoring and “scale normalisation” effects; however, while that model fits the psychophysical data, it is not predictive as to the strength of the effect, because the weight is never explicitly defined. A probabilistic model more similar to the one here also explains the nonlinear relationship between lightness and intensity in terms of possible real-world sources of an ambiguous stimulus [22], if the relative contributions of reflectance and illumination can be estimated. However, the nonlinearity in brightness contrast, which can be inferred from this model, is symmetrical, not asymmetrical as it is here—and in human perception.
Our model suggests a more explicitly data-driven explanation. We express the reflectance R and illumination I as fractions of their potential maximum values, so in all cases 0 < I < 1 and 0 < R < 1. Because the stimulus intensity S = I • R, it is similarly bounded between zero and one. Therefore, the value of S defines the minimum possible illumination and reflectance of a target. As an example of this, suppose that S = 0.7 in some particular stimulus; the darkest possible value of R corresponds to the maximum illumination I = 1, giving R = 0.7 as the minimum lightness possible. If the exact illumination is unknown, then the bounds are 0.7 ≤ R ≤ 1 in this case, and conversely, 0.7 ≤ I ≤ 1. In the extreme, if S = 1, then R = 1 and I = 1 are the only possible sources and the stimulus is totally unambiguous. Conversely, images (or parts of images) with low luminance intensity are more ambiguous—i.e., have a wider range of possible scores for I and R—than high-intensity images. This increased range of possible sources of darker images leads to a greater magnitude of perceptual errors on average, which translates into a larger overestimation of reflectance compared to lighter surrounds on average, assuming that negative values are never predicted. We are not claiming that visual systems must explicitly contain such a model of physics, or that the exact values must be known, but only that past experiences of the consequences of the physics of the environment are encoded in the system, and so behaviour guided by such experiences will lead to the observed patterns of errors.
An important aspect of brightness contrast in humans is that the strength of the illusion is as much a function of stimulus structure as it is of stimulus intensity. For instance, increasing a scene's articulation (as in Figure 2B) increases human perception of brightness contrast considerably [23,24]. Similarly, when presented with targets on two fully articulated surrounds (one light, one dark), the difference in the predicted reflectance of the identical targets was increased (Figure 2B). Also, altering the spatial configuration of a target's surround, without altering the average luminance, can create the illusion of brightness contrast. When the nets were presented with targets on surrounds of identical average intensity, but of differing spatial structure (Figure 2C), they continued to underestimate the target on a local light surround, and overestimate the target in a local dark surround much like humans, specifically outputting R = 0.32 (0.025) and R = 0.74 (0.011), respectively, for the images shown in Figure 2C. The papers summarised in [23] discuss various aspects of articulation in detail, including the effect of both the number of surfaces in a scene and their structural organisation. Similarly here, it is not simply increasing the number of surfaces that leads to better constancy (and so to smaller errors), but the structure of the articulation. More specifically, what matters is past experience regarding the probable source of that articulated information, as has been suggested previously [23].
The ANNs were next tested on other, more complex but well-known brightness contrast–like phenomena, specifically the Vasarely illusion, Mach bands, Chevreul patterns, and the Hermann grid. In the Vasarely illusion (Figure 3A), the corners of each repeated square appear brighter than their immediate surround (which results in what looks like a four-edged star), even though the stimulus is uniform at these junctions. In Mach bands (Figure 3B), a linear gradient appears to be flanked by a highlight at the lightest end of the gradient and a “lowlight” at the gradient's minimum. Neither of these features actually exists in the intensity profile of the stimulus. In Chevreul patterns (Figure 3C), uniform bars appear graded in lightness. And in the Hermann grid (Figure 3D), light spots appear at the central junction of the dark lines where no light dot actually exists. The 50 trained ANNs were presented with each of these stimuli in turn, none of which were presented during training. Their average response is shown in the corresponding row of the right column in Figure 3A–3D. By comparing the stimulus' intensity profile (red line) with the nets' response profile (blue line) at each corresponding point, it is clear that, as before, the networks exhibit responses that are qualitatively similar to human perception in each instance. (Whether they are quantitatively similar to human perception is not relevant, given the inevitable differences in complexity between natural ecology and the “dead-leaves” ecology.)
The results thus far are consistent with the hypothesis that human illusions of lightness are caused by nothing more (or less) than image ambiguity and its empirical—and thus statistical—resolution. The above contrast illusions, however, are also consonant with many other models predicted on, for instance, the statistics of natural images or assumption about low-level and mid-level processing [1,5,24–25]. Indeed, any model that incorporates lateral inhibitory connections, such as centre/surround receptive fields, will predict most of the above phenomena (e.g., [5,25]), which is the typical explanation in most neuroscience textbooks. Few explanations, however, can simultaneously predict both brightness contrast (including its asymmetry) and brightness assimilation—e.g., White's illusion—without recourse to one or more adjustable free parameters [25]. (Important exceptions include the filter model discussed previously [5] and a statistical approach which uses a database of natural scenes to estimate probability distributions over structures in lightness stimuli, including White's stimulus [26].) What makes these two illusions difficult to reconcile simultaneously is that they are diametrically opposed to one another. In brightness contrast, the target on a dark surround appears lighter than the same target on a light surround (Figure 2A), whereas the opposite is true for assimilation in general and White's illusion in particular: the target on the overall darker local surround appears darker (not lighter) than the same target on the overall lighter local surround (Figure 3E; see [1] for an elegant description of these phenomena and their current explanations). White's stimulus can be interpreted as a series of vertical dark and light bars partially obscuring a pair of mid-grey bars on a monochrome background.
Here, the trained ANNs exhibited both brightness contrast and White's illusion (see right column of Figure 3E). As always, the emergent behaviour of the ANNs can be explained in terms of the statistics of their visual experience. Of particular relevance is their experience with the 3-D layering of the surfaces in space. A separate group of ANNs was trained using scenes composed of surfaces in only one depth plane, consisting of the same “dead-leaves” images described in the Methods section, but without the separate mask layer on any of the stimuli. Compared to the main group of ANNs, these lost the “ability” to see White's illusion, but maintained the ability to see lightness constancy, brightness contrast, and related phenomena (unpublished data). Thus, when presented with surfaces at different depth planes under independent illumination, the ANNs learned to ignore information arising from surfaces that were not co-planar with the target; since illumination of each depth-plane is independent, only co-planar information provides statistical information about the probable source of the target. Thus, changing the ecology (by introducing layers using masks) leads directly to a change in behaviour (the ANNs' response to White's stimulus) showing a causal link between the two.
It is important to emphasise, however, that while White's illusion only arises when the networks had experience of 3-D scenes, this is not equivalent to saying that the networks “represented” depth in their post-receptor processing. Indeed, it is highly unlikely that the networks encode depth information explicitly, or indeed contour junction cues, as has been posited for human visual processing, since varying the spatial frequency of the stimulus or the height of the individual test patch varies the strength of the illusory response (see Figure 4A and 4B, respectively) without altering the stimulus' junctions. More specifically, decreasing the spatial frequency of the stimulus and/or target height decreases the ANNs' perception of White's illusion without altering the stimulus' junctions. Remarkably, these latter two observations have also been made of human perception of White's stimulus [5].
Not all human lightness illusions are a consequence of spatial context, and in these cases we found further similarities between the ANN's behaviour and human visual perception. For instance, when viewed in a “void” (i.e., on a black surround), the relationship between a surface's stimulus and its (human-) perceived lightness is not linear, but follows the power law ψ(S) = kSα, where ψ(S) is the perceived lightness, S is the physical intensity of the stimulus, k is a scaling constant, and α is the exponent that describes the shape of the relationship to perceived lightness. For humans, the value of the exponent α typically varies between 0.33 to 0.5 in different studies [27]. When the ANNs are presented with uniform images of increasing intensity, the relationship between target intensity and predicted reflectance also follows a power law with an exponent (α) that equals 0.334—broadly similar to humans.
The ANNs used here are structurally unlike the human visual system: they are smaller and less complex; they lack recurrent connections, spiking, adaptation (after learning is complete), and so on; they are nonhierarchical, and so cannot generate behaviours according to so-called “top-down,” “mid-level,” or “cognitive” influences on “bottom-up” processing. Indeed, these ANNs lack all the proximal mechanisms that are usually thought to be the immediate cause of human visual illusions. Instead, the output of each ANN is driven solely by the statistics of its training history instantiated in the functional architecture of its network. Though sometimes seen as a drawback, this simplicity is taken advantage of to rationalise human illusions, not by modelling what is currently known of human perception and/or primate neurophysiology, but by modelling the inherent ambiguity of human and nonhuman visual ecology that all natural systems must solve to survive, and its empirical resolution. This extends several recent studies that have found relationships between the statistics of images of natural scenes and human perception (e.g., [25–26,28–30]). We can begin to move from correlative to causative explanations.
Perception can be defined as the process of acquiring and organising information from sensors. The input nodes of the ANNs are presented with images in terms of the luminance intensity across space, from which the ANNs must extract scene information, specifically the reflectance of a target patch. This is equivalent to one of the many tasks that the human visual system performs. The Oxford English Dictionary defines an illusion as “something that deceives or deludes by producing a false impression.” Every instrument has measurement errors and the human visual system is no exception. So every percept will have an error associated with it, be it large or small. Errors in visual perception are defined as the difference between what is seen and what the actual physical quality of the retinal stimulus with which the percept is associated is [1], irrespective of whether the physical source is ever known. Given this definition, a so-called “illusory image,” such as the stimuli in Figure 3, is one that induces perceptions that deviate from the underlying reality of the image, a view consistent with recent Bayesian frameworks of constancy (e.g., [12]). There is, however, no absolute threshold on these errors that defines a percept as illusory or non-illusory. We must therefore consider the magnitude of perceptual errors and relate these to the past experiences of the observer.
Returning to the ANNs used earlier, recall that when shown novel “dead-leaves” images, the RMS error was 0.171. Furthermore, approximately 79% of the predictions were within ±0.2 of the target, and just 1% of the errors were greater than ±0.5; i.e., most images were interpreted approximately correctly, but none perfectly. The equivalent error for simultaneous brightness contrast (Figure 2A), with a mid-grey patch on a black background, was 0.36 (s.d. 0.032), an unusually large error. As a specific example, Figure 5 shows the range of all possible reflectances of a single target patch (x-axis) and their relative probabilities (y-axis), for a single “dead-leaves” stimulus. The probabilities are derived from the past experiences of a single ANN, and the peak on the curve corresponds to a reflectance (R) of 0.93. This is by definition the “most likely source” of this particular stimulus in the ANN's past visual experience. If the actual reflectance of the stimulus under consideration is close to the most likely source of the stimulus (i.e., a surface with a reflectance close to 0.93), then the prediction/percept is “correct” and we would say that lightness constancy holds. One would also say that the percept is not an illusion. If, on the other hand, the actual reflectance happened not to be near the most likely source of the stimulus (i.e., more than or less than 0.93), then while the predicted reflectance would have been “correct” most of the time, it would be “wrong” in this particular instance, lightness constancy would have failed, and the percept would be called an illusion. What is more, the further into the tail of “unlikeliness” the source of the stimulus is, the more “illusory” the percept becomes, suggesting that illusions of lightness and lightness constancy exist on a continuum, as opposed to being fundamentally different kinds of phenomena.
It is therefore misleading to describe any stimulus as being an illusion in isolation. Instead, one can describe the true source of a stimulus as being unlikely given the past experiences of a particular observer, and therefore likely to induce an erroneous percept in that observer. Given the similarity of the shared experiences of humans, and our shared genes, it should not be surprising that the patterns of errors that we make are also shared. The exact distribution of errors for human or animal perception is hard to quantify, and the factors leading to more or less lightness constancy are largely unknown [23]. However, it seems clear that most responses are approximately correct, at least where it is possible to measure the true source, although constancy does fail significantly in some cases. The nearest human psychophysical study that we are aware of measures colour constancy for coloured papers under varying illumination [31]. They define a constancy index that ranges from 0 for no constancy to 1 for perfect constancy, the latter meaning that a surface colour is perceived according to the surface spectral properties alone (and not illumination, for example). They measure a colour constancy index of around 0.8, although in many experiments the index was much lower. Constancy can be seen as the inverse of illusions, if we assume that the constant response has a small error under a range of illuminations, and illusions generate large errors. The errors of the ANNs suggest a similar magnitude of constancy, although direct comparison between such different measures is never ideal. We know of no such score for lightness constancy under typical, natural conditions, but it is reasonable to suppose a broadly similar continuous distribution exists for humans, too.
In conclusion, the emergent similarity between human perception and the ANNs' output provides direct support for the view that illusions are caused by (as opposed to merely correlated with) the statistics of past visual experience towards surfaces in space under spatially heterogeneous illumination given ambiguous image data. Because stimulus ambiguity is an inherent challenge of natural visual ecology, illusions must also be inevitable in nature, suggesting that human illusions are common to all visual animals despite vast differences in their underlying neural machinery, which has important consequences for thinking about the biological and computational principles of vision. Evolving or training synthetic systems in ecologically relevant environments provides an important new strategy for uncovering what these principles are that usefully map images to scenes according to the statistics of experience. Finally, the study provides a clear description of what an illusion is, and why we see them: an illusion describes the condition in which the actual source of a stimulus differs from the stimulus' most likely source given the observer's past experience.
The ANNs used here were standard multilayer perceptrons trained via backpropagation. We use multilayer perceptrons because they are known to be universal approximators, capable of learning arbitrary mappings from a finite set of examples. In preliminary experiments, we achieved similar results using support vector regression methods (unpublished data), and believe that any suitable powerful nonlinear multivariate regression tool would work as well. The behaviour we describe is ultimately due to the data, not the learning algorithm.
Each ANN had 400 inputs nodes, one for each pixel of the stimuli; four hidden nodes in one layer; and one output node. The training was supervised, so the target reflectance in the training images was used to estimate errors during the training. The output was therefore the ANN's prediction of the reflectance of the central target patch of the stimulus presented to it. The inputs consisted solely of the stimulus intensity, and not reflection or illumination explicitly. All nodes were fully connected to nodes of their adjacent layers; there were no connections between nodes of the same layer; and connection weights could be positive or negative. Each ANN was initialised with random weights, then trained for 150 iterations with 20,000 training images. These parameters were chosen based on preliminary experiments, and are not critical. Many factors are known to affect the performance of ANNs, such as the number of hidden nodes, the learning rate, the number of training iterations (see Figures S1 and S2), the number of training examples, and so on. Furthermore, these factors tend to interact, making any exhaustive analysis effectively impossible, and making it difficult to guarantee that any particular ANN is “optimal.” However, our aim here is not optimality, but is rather to show that the results described in the paper are robust, and, to demonstrate this, we now briefly analyse some of these parameter settings.
All nonparametric learning systems, including ANNs trained by backpropagation, are prone to “overfitting,” when they accurately model the data that they are trained with, but fail to generalise well to novel data. One conventional solution is to stop training after a fixed number of iterations, before this problem arises, which is why we limit the training algorithm to 150 iterations (see Figure S1).
To see the effect of varying the number of hidden nodes, we trained a series of ANNs, each containing between one and 50 nodes in a single hidden layer. The minimum error corresponds to ANNs with four nodes in their hidden layer (see Table S1). However, a series of t-tests indicate that the other ANNs achieved performances that were not significantly different (p > 0.05 in all cases). Thus, the choice for the number of nodes is somewhat arbitrary, reinforcing the notion that it is the statistics of the training set that are critical, rather than the fine details of the learning algorithm.
Next we considered the number of training records used by the backpropagation algorithm. Again, we trained a series of ANNs with sets of novel “dead-leaves” stimuli. Each ANN had four hidden nodes, but the number of training records varied from 333 to 20,000. As expected, being given more training examples allowed the ANNs to achieve a lower test error, because each independent training example provides extra information about the underlying function (see Table S2). Given the trend of decreasing returns, increasing the number of records above 20,000 would make only a marginal difference, with the cost of longer training times.
Each node of an ANN contains an activation (or “transfer”) function, which takes the sum of the inputs and transforms it, typically rescaling the value to a fixed range. A typical activation function, which we use in the ANNs described in the main paper, is the log sigmoid function, which produces values in the range [0, 1]. The tan sigmoid function, which produces values in the range [−1, +1] and the linear transfer function, which produced unbounded values, were also used in new ANNs for comparison. As the errors in Table S3 show, there is no significant difference between log sigmoid and tan sigmoid functions, as expected. The pure linear activation function, which gives no bounds on the outputs, leads to significantly worse performance. Thus the choice of a particular activation function is not critical, although in the extreme case of a linear function, learning is considerably degraded.
We also tested some of these alternative ANNs with the various “illusion” stimuli used elsewhere in the paper. As a simplified measure of different ANNs responses to the test “illusory” stimuli, we measured each ANN's predicted reflectance for the test patches in the brightness contrast, Hermann grid, and White's stimuli (see Figures 2A, 3D, and 3E, respectively). For each stimulus, we selected two pixels that had identical reflectance values but generate illusory responses in humans. For each pair, we calculated the difference in the ANN's response, such that a score of zero means that they do NOT perceive any illusion, and a positive score corresponds to human perceptions. (This is the same differential measure used in Figure 4C.) The larger the positive score, the stronger the illusion is perceived. Negative scores indicate the “opposite” of human perception. While there is no direct relationship between the magnitudes and human perception, they do provide an indication of the strengths of the illusions for the ANNs. The overall effect is that as training proceeds, the error drops and the strength of the illusions increases (Figure S2). This again shows that the appearance of illusions is causally related to solving the lightness constancy problem.
All experiments were carried out on a standard desktop PC using Matlab 6.5 (Mathworks) and the Matlab Neural Networks toolbox version 4.
A number of 200 × 200 pixel “dead-leaves” images were created following the algorithm presented by Lee et al. [9], which produces images with similar statistics as those that have been found in a wide range of natural scenes. The implementation we used was based on Matlab code provided in the Toolbox Signal (2006) by Gabriel Peyré. Each image was composed of a large number of partially occluding achromatic disks, which can be thought of as a series of “dead leaves” falling on top of each other. The leaf radius is distributed as 1/r3, so these images tend to have a few large “leaves” and many smaller ones, much as with natural scenes. For presentations to the ANNs, random 20 × 20 pixel samples were selected from these large images. The minimum-sized disk was fixed at 0.002 for the reflection maps and 0.01 for the illumination maps. The latter were blurred by filtering with a Gaussian filter of size 8 × 8 with a width of 15. The stimulus matrix presented to the ANNs is defined as S = I • R. Both I and R (and therefore S) are scaled in the range 0…1. Where a second layer was used to create 3-D stimuli (in one-fifth of the training set), the same procedure was used to create the surfaces and the illumination. The layer was then reduced to a series of random horizontal and vertical strips covering an average of 10% of the image opaquely. The remaining 90% was unchanged. The target could be in either layer. We have not carried out any human psychophysical experiments testing responses to these stimuli; however, the algorithm is designed to generate images that are statistically similar to natural scenes, so we assume that human responses would be quite consistent with responses to natural scenes.
Preliminary work showed that if the distribution of the reflectance and illumination maps were very similar, then the ability to resolve lightness constancy in the ANNs was reduced, though not abolished (unpublished data). Presumably, this is because every stimulus was so ambiguous that resolution was increasingly difficult. Given that humans and other animals can solve lightness constancy at least most of the time, the real visual ecology must provide enough information to allow the disambiguation to take place. In our simplified model, this is achieved by ensuring that the distributions of R and I are sufficiently different.
These “dead-leave” images, with heterogeneous light and partial masking, represent a simple model ecology. The size of the distinct surfaces within each scene follows the same distribution as found in natural scenes. The illumination is assumed to come from multiple sources, consistent with some light being reflected from nearby surfaces. The reflectance map is therefore approximately piecewise constant, while the illumination map only changes smoothly, as in [4] and elsewhere. The addition of a second “masking” layer aims to simulate effects such as the viewer looking through the branches of a tree or through a windowframe. Such a simple model could be extended in many ways to make it more natural and realistic, such as added colour, transmittance effects, depth, objects of varied shape with or without attached shadows, sharp shadow edges, and so on. Several of these are the subject of ongoing work, which will allow a wider range of visual behaviours to be studied, such as testing the models' response to colour illusion stimuli. Similarly, we have chosen not to model the eye explicitly, such as defining cone response functions or light adaptation, instead concentrating on the more generic aspects of learning to respond to a visual ecology. |
10.1371/journal.ppat.1003478 | Negative Regulation of Type I IFN Expression by OASL1 Permits Chronic Viral Infection and CD8+ T-Cell Exhaustion | The type I interferons (IFN-Is) are critical not only in early viral control but also in prolonged T-cell immune responses. However, chronic viral infections such as those of human immunodeficiency virus (HIV) and hepatitis C virus (HCV) in humans and lymphocytic choriomeningitis virus (LCMV) in mice overcome this early IFN-I barrier and induce viral persistence and exhaustion of T-cell function. Although various T-cell-intrinsic and -extrinsic factors are known to contribute to induction of chronic conditions, the roles of IFN-I negative regulators in chronic viral infections have been largely unexplored. Herein, we explored whether 2′–5′ oligoadenylate synthetase-like 1 (OASL1), a recently defined IFN-I negative regulator, plays a key role in the virus-specific T-cell response and viral defense against chronic LCMV. To this end, we infected Oasl1 knockout and wild-type mice with LCMV CL-13 (a chronic virus) and monitored T-cell responses, serum cytokine levels, and viral titers. LCMV CL-13-infected Oasl1 KO mice displayed a sustained level of serum IFN-I, which was primarily produced by splenic plasmacytoid dendritic cells, during the very early phase of infection (2–3 days post-infection). Oasl1 deficiency also led to the accelerated elimination of viremia and induction of a functional antiviral CD8 T-cell response, which critically depended on IFN-I receptor signaling. Together, these results demonstrate that OASL1-mediated negative regulation of IFN-I production at an early phase of infection permits viral persistence and suppresses T-cell function, suggesting that IFN-I negative regulators, including OASL1, could be exciting new targets for preventing chronic viral infection.
| Chronic viral infections, such as those of human immunodeficiency virus (HIV) and hepatitis C virus (HCV) in humans, remain serious health problems worldwide, necessitating alternative targets/reagents for better treatment. Although the production of and/or response to type I interferon (IFN-I), a critical antiviral reagent, are known to be dysregulated in chronic viral infections, no serious effort has been performed to determine whether any host IFN-I negative regulator can importantly contribute to inducing and/or maintaining chronic viral infections. In this study, we used a mouse model of chronic viral infection, lymphocytic choriomeningitis virus (LCMV) infection, and asked whether 2′–5′ oligoadenylate synthetase-like 1 (OASL1), a recently defined IFN-I negative regulator, plays a key role in the viral defense against chronic LCMV infection. Our data show that OASL1 suppresses IFN-I production during very early phase of infection, thus inhibits efficient viral control and the induction of functional virus-specific T-cell response, permitting viral persistence. These results indicate that OASL1-mediated suppression of IFN-I production is a critical step for permitting chronic viral infection and suggest that IFN-I negative regulators, including OASL1, could be exciting new targets for preventing chronic viral infection.
| Pattern-recognition receptors (PRRs) displayed on innate immune cells such as dendritic cells (DCs) and macrophages (Macs) sense pathogens by recognizing conserved pathogen-associated molecular patterns (PAMPs) [1], [2]. Major trans-membrane PRRs that sense viruses are the Toll-like receptors (TLRs) such as TLR3, TLR7, and TLR9, and such cytosolic PRRs are retinoic acid-inducible gene I (RIG-I)-like receptors (RLRs) such as RIG-I and melanoma differentiation-associated gene 5 (MDA5) [3]. Upon recognition of cognate ligands, these PRRs initiate various signaling pathways that lead to the production of inflammatory cytokines, including type I interferons (IFN-Is), such as IFN-αs/β, which are critical for inhibiting early viral replication in the host [3], [4]. Additionally, antigen-presenting cells (APCs), particularly DCs, up-regulate co-stimulatory molecules and major histocompatibility complex (MHC) molecules upon viral sensing and induce the differentiation of effector T cells which are key adaptive immune cells required for later viral clearance [5], [6].
The host immune system can effectively induce virus-specific T-cell activation, expansion, and successful generation of memory T cells upon acute viral infections. However, the immune system cannot induce such response upon chronic viral infections such as those of human immunodeficiency virus (HIV), hepatitis C virus (HCV), or Epstein Barr virus (EBV) in humans, and lymphocytic choriomeningitis virus (LCMV) in mice. As a result, the hosts live with a persistent viral load throughout their life-spans and have fundamentally dysfunctional T cells that produce dampened effector cytokines [7], [8]. Various virus-specific T-cell-intrinsic and -extrinsic factors have been known to contribute to inducing and/or maintaining the chronic conditions. Virus-specific T-cell-intrinsic factors include elevated expression of inhibitory receptors such as programmed death-1 (PD-1), T-cell immunoglobulin and mucin protein-3 (TIM-3), cytotoxic T-lymphocyte antigen-4 (CTLA-4), and lymphocyte-activation gene-3 (LAG-3) [9]–[14], whereas virus-specific T-cell-extrinsic factors include altered antigen presentation by impaired DCs [15], enhanced immune suppression by regulatory T (Treg) cells [16], [17], and increased immunosuppressive cytokines such as interleukin-10 (IL-10) [18], [19] and transforming growth factor-β (TGF-β) [20], [21].
In addition, suppression of IFN-I production and response could be a major contributing factor leading to the chronic condition. Suppression may be a result of the reduced number of plasmacytoid DCs (pDCs), a major cellular source of IFN-I upon various viral infections. Indeed, the numbers of pDCs are reduced in humans and mice during infections with HIV, HCV, EBV, and LCMV [22]–[27]. Alternatively, such suppression could be actively regulated by host- and virus-derived negative factors that act on diverse PRR-signaling components and two major transcription factors (TFs), interferon regulatory factor 3 (IRF3) and IRF7 [28]–[30], involved in IFN-I production as well as on IFN-I receptor signaling components involved in the IFN-I response [28]–[30]. To overcome the dysregulated IFN-I production, IFN-I has been used clinically to treat patients infected with certain chronic viruses such as HCV. However, high dose and long-lasting IFN-I treatment is known to be necessary to achieve any therapeutic benefit, implying that there are some underlying mechanisms to negatively regulate the IFN-I response [31]. Therefore, whether host-derived IFN-I negative regulator plays any significant role in promoting viral persistence in the setting of chronic viral infection is quite an important question.
Recently, we showed that 2′–5′ oligoadenylate synthetase (OAS)-like 1 (OASL1), an IFN-stimulated gene (ISG), is a novel translation inhibitor of IRF7, the IFN-inducible IFN-I master TF, and negatively regulates robust IFN-I production during acute viral infections [32]. Therefore, we hypothesized that OASL1 could suppress IFN-I production during chronic LCMV infection by inhibiting IRF7 production and permit persistent viral infection. In the present study, we show that Oasl1 knockout (KO) mice displayed a sustained level of IFN-I during early viral infection with LCMV clone 13 (LCMV CL-13, a chronic virus), controlled viremia quickly and induced better functional T-cell responses compared with wild-type (WT) mice. In addition, we show that IRF7 protein expression was higher in virus-infected Oasl1 KO spleen than in WT spleen and that IFN-I receptor signaling during this early period was necessary for accelerated viral control and the enhanced T-cell immune response in the Oasl1 KO mice. These results indicate that OASL1-mediated negative regulation of IFN-I production at the early phase of the infection is critical in permitting viral persistence.
To investigate the in vivo role of OASL1 in T-cell differentiation and viral defense upon chronic viral infection, we infected both WT and Oasl1 KO mice with LCMV CL-13 and monitored CD8+ T-cell numbers and phenotypes as well as viral titers in the blood for 35 days (d) post-infection (p.i.). The percentage of total CD8+ T cells in peripheral blood mononuclear cells (PBMCs) became much higher in KO mice than in WT mice (>3-fold at 15 d p.i.) upon infection (Fig. 1A and 1B). In addition, the frequency of activated CD8+ T cells expressing CD44 was maintained at a slightly higher percentage in KO mice than in WT mice after infection (Fig. 1A and 1B). The CD44-expressing cells in Oasl1 KO mice at 35 d p.i. were mostly specific to various LCMV epitopes (approximately 80% of CD44+ cells) (data not shown). Notably, the number of LCMV GP33–41 (GP33) peptide-specific CD8+ T cells was more strongly increased in KO mice compared with WT mice (>10-fold at 15 d p.i., Fig. 1C and 1D). These results indicate that virus-specific CD8+ T cells undergo much more massive expansion in Oasl1 KO mice compared with WT mice.
Expression of PD-1 on GP33 peptide-specific CD8+ T cells did not appear to be significantly induced in Oasl1 KO mice after LCMV CL-13 infection, although WT mice exhibited strongly induced PD-1 expression patterns (Figs. 1E and S1A). CD127, a marker of memory T cells, was progressively up-regulated in KO mice but not in WT mice after initial downregulation (Figs. 1E and S1B). These data imply that CD8+ T-cell responses might also be better qualitatively in Oasl1 KO mice upon chronic viral infection. Remarkably, there was complete control of viremia in serum at 15 d p.i. in LCMV CL-13-infected Oasl1 KO mice, but viremia remained continuously high in WT mice serum by 35 d p.i. (Fig. 1F). Consistent with accelerated viral control in KO mice, the mice recovered rapidly from an initial loss in body weight, whereas WT mice did not recover after an even longer period of time (Fig. S2). As expected, Oasl1 KO mice also controlled LCMV Armstrong (LCMV Arm, an acute strain of LCMV) infection much more quickly than WT mice (Fig. S3). Interestingly, the viral titers in Oasl1 KO mice at 4 and 9 d p.i. upon LCMV CL-13 infection were quite comparable to those in LCMV Arm-infected WT mice (Fig. S3). The observation that Oasl1 KO mice control the LCMV CL-13 infection as quickly as the WT mice control the LCMV Arm infection and the known fact that T-cell differentiation occurs efficiently in WT mice upon LCMV Arm infection [33], [34], together imply that rapid viral control in the blood of Oasl1 KO mice upon LCMV CL-13 infection may be one of the major contributing factor for the enhanced CD8+ T-cell differentiation in the KO mice.
We next asked whether the enhanced immune features observed in the blood from Oasl1 KO mice were also present in tissues at a later phase of infection (more than 2 months p.i.). We included another tetramer specific to GP276–286 (GP276), an immunodominant epitope other than GP33, to exclude the possibility that the T-cell responses were virus epitope-specific phenomena. Frequencies and numbers of both GP33- and GP276-specific CD8+ T cells were significantly higher at 75 d p.i. in the spleen and lung of Oasl1 KO mice than in those of WT mice, whereas the frequencies and numbers of such T cells seemed to be lower in the liver of KO mice (Fig. 2A and 2B). The observation that a slightly lower number of virus-specific T cells in the liver (a nonlymphoid tissue) of Oasl1 KO mice is not unusual because frequent antigen exposure to T-cell receptor of CD8+ T cells in the condition of chronic infection is known to enable virus-specific CD8+ T cells to migrate preferentially into nonlymphoid tissues rather than lymphoid tissues [33]. However, PD-1 expression levels in GP33- and GP276-specific CD8+ T cells were significantly lower in the spleen, lung, and liver of Oasl1 KO mice than in those of WT mice (Fig. 2C). Furthermore, we observed much higher frequencies of virus-specific memory CD8+ T cells (CD127+) in all the tissues collected from KO mice at 75 d p.i. (Fig. 2D). Even CD127+ CD62L+ central memory CD8+ T cells were easily detectable in the spleen of KO mice at 130 d p.i., but to a lesser extent in WT mice (Fig. 2E). Together, these results indicate that differentiation of virus-specific CD8+ T cells progresses more efficiently in Oasl1 KO mice than in WT mice during the late period of LCMV CL-13 infection.
To confirm whether virus-specific T cells in KO mice truly demonstrate better function, we assessed the ability of CD8+ and CD4+ T cells to produce effector cytokines such as IFN-γ, tumor necrosis factor alpha (TNF-α), and IL-2 at the individual cell level. After in vitro restimulation of splenocytes collected at 75 d p.i. with GP33–41 or GP276–286 peptide for CD8+ T cells and GP66–80 peptide for CD4+ T cells, the IFN-γ-producing T-cell population was detected at a significantly higher level in the Oasl1 KO sample than in the WT sample (Fig. 3A and 3B). In addition, the percentages of cells producing TNF-α or IL-2 among the IFN-γ-producing cells were much higher in KO mice than in WT mice (Fig. 3A and 3C). The better differentiation and function of virus-specific CD8+ T cells observed in Oasl1 KO mice at 75 d p.i. was maintained at an even later time point (130 d p.i.), although the frequency of LCMV-specific CD8+ T cells became similar to that of WT by this time point (Figs. S4 and S5).
Although viral titers in the spleen, lung, and liver of both WT and KO mice at 75 d p.i. were already below a detectable level, the titer in the kidney (a well-known life-long reservoir of chronic LCMV) was detectable, much less in KO mice than in WT mice, indicating better viral control in KO mice (Fig. 3D, data not shown). In addition, at 130 d p.i., Oasl1 KO mice could eliminate residual viruses completely even in the kidney, whereas WT mice continued to exhibit high viral titers in the kidney (Fig. 3D). These data indicate that the rapid viral control in Oasl1 KO mice upon LCMV CL-13 infection could be responsible for the enhanced multiple immune responses such as increased virus-specific T-cell function and memory T-cell formation in KO mice, a condition that rarely occurs in WT mice during LCMV CL-13 infection. As a result, Oasl1 KO mice completely overcame persistent infection at the late stage of infection.
In the next set of experiments, we attempted to clarify whether the better virus-specific T-cell responses in the absence of OASL1 protein were caused either by virus-specific T-cell-intrinsic change or T-cell-extrinsic change. We also wanted to investigate whether the expanded population of virus-specific CD8+ T cells in Oasl1 KO mice was due to better proliferation or decreased apoptosis of such T cells. To address these questions, we employed LCMV GP33–41 epitope-specific T-cell receptor (TCR) transgenic mice (P14 mice). After purifying Thy1.1+ CD8+ T cells from naïve P14 mice, the P14 Thy1.1+ CD8+ T cells were adoptively transferred into both naïve WT and Oasl1 KO mice. The following day, recipient mice were infected with LCMV CL-13 and sacrificed 5 d p.i. when the virus-specific CD8+ T-cell population expands enormously.
Identical virus-specific P14 CD8+ T cells transferred into WT and KO mice displayed noticeably different responses after the chronic virus infection. The P14 donor cells exhibited approximately 3-fold more expansion in Oasl1 KO mice than in WT mice in vivo (Fig. 4A). When restimulated in vitro with cognate GP33–41 peptide, P14 cells isolated from KO mice at 5 d p.i. generated many more cells co-producing effector cytokines IFN-γ, TNF-α, and IL-2, than those from WT mice (Fig. 4B). These results indicate that virus-specific CD8+ T cells respond better to viral antigens in Oasl1 KO mice both quantitatively and qualitatively. Collectively, these data show that unknown T-cell-extrinsic factors generated in the absence of OASL1 protein play a major role in inciting the donor P14 Thy1.1+ CD8+ T cells to dramatically increase expansion and cytokine production at the early phase of LCMV CL-13 infection.
When the proliferative capacity of T cells was monitored at 5 d p.i. by measuring the expression of Ki67 (a marker of recently proliferated cells) ex vivo, the expression level of Ki67 within the Thy1.1+ CD8+ T-cell population in Oasl1 KO mice was similar to that in WT mice (Fig. 4C). However, when the apoptotic phenotype of Thy1.1+ CD8+ T cells was monitored by measuring the surface expression of Annexin V (AV) and permeability of 7-amino-actinomycin D (7AAD) into cells ex vivo, early apoptotic AV+ 7AAD− cells were present more frequently in WT mice than in KO mice (Fig. 4D). By contrast, the non-apoptotic population (AV− 7AAD−) was much larger in KO mice than in WT mice (Fig. 4D). These results suggest that the extrinsically driven rapid expansion of donor P14 T cells in Oasl1 KO mice is due to better survival rather than proliferation of donor T cells.
To explore which T-cell-extrinsic factors play major roles in the early rapid expansion of CD8+ T cells in Oasl1 KO mice, we measured the levels of major inflammatory cytokines in serum and those of co-stimulatory/MHC molecules expressed on splenic DCs during the early phase of LCMV CL-13 infection. The serum levels of IFN-I were sustained longer in Oasl1 KO mice than in WT mice from 2 d to 3 d p.i. (Fig. 5A). However, the serum levels of other major cytokines such as IL-6, IL-10, TNF-α, and IFN-γ did not show any meaningful difference between WT and Oasl1 KO mice at 2 d p.i. but displayed moderately higher signals in KO mice at 3 d p.i. (Fig. S6). This more upregulation of other cytokines at 3 d p.i. in KO mice is thought to be secondarily caused by early sustained IFN-I produced in KO mice [35]–[39]. The expression levels of co-stimulatory molecules and MHC molecules on different subsets of splenic DCs were not higher in Oasl1 KO mice than in WT mice early after LCMV CL-13 infection (Fig. S7). These results together suggest that IFN-I, a known T-cell survival factor, among extrinsic factors such as cytokines, co-stimulatory molecules, and MHC molecules, sustained longer in Oasl1 KO mice during the very early stage of infection, may play a critical role in initiating massive expansion of virus-specific CD8+ T cells.
We previously reported that Oasl1 KO cells produce more IFN-I upon acute viral infection, which was caused by higher production of IRF7 proteins in KO cells [32]. Thus, we next asked whether chronic LCMV infection also caused higher IRF7 protein production in Oasl1 KO mice to produce the sustained IFN-I. To this end, we first determined the tissue that had the highest IFN-I expression after LCMV CL-13 infection. Among the four tissues that we collected at 2 d p.i., at the time point when the IFN-I level in serum was clearly stronger in KO mice, the spleen showed the most dominant expression of IFN-I mRNA: a 20-fold more IFN-I mRNA signal compared with other tissues of both WT and KO mice (Fig. S8). Importantly, in the spleen, the IFN-I mRNA level was significantly higher (6- to 8-fold) in KO than in WT mice. Thus, we next focused on spleens to measure the expression of two major TFs responsible for IFN-I mRNA expression during early LCMV CL-13 infection. Upon LCMV CL-13 infection, as expected, IRF7 mRNA (but not IRF3 mRNA), similar to IFN-I mRNA and OASL1 mRNA, was strongly induced at 1 d p.i. and declined over the next several days (Fig. 5B). Similar to serum IFN-I levels (Fig. 5A), Oasl1 KO spleens expressed much more IFN-I mRNA during the declining several days. Consistent with our previous report that OASL1 specifically inhibits the translation of IRF7 mRNA upon viral infection, IRF7 protein levels were much higher in Oasl1 KO cells during the days, whereas IRF3 protein levels were similar between WT and KO spleens (Fig. 5C). These results together indicate that the more production of IFN-I in the Oasl1 KO mice during the early phase of LCMV CL-13 infection was caused by high production of IRF7 proteins in Oasl1 KO mice.
We next investigated which cell types are the major source of better IFN-I production in LCMV-CL-13-infected Oasl1 KO mice. Because major cell types known to produce IFN-I after viral infection in vivo are pDCs, conventional DCs (cDCs) including myeloid DCs (mDCs) and lymphoid DCs (lDCs), and Macs [35], we sorted these IFN-I-producing cell types and, as a negative control, polymorphonuclear cells (PMNs) and monocytes (Mos) from splenocytes of WT and Oasl1 KO mice at 2 d p.i. (Fig. S9), followed by measurement of their IFN-I gene expression levels. Although splenic cDCs and Macs of Oasl1 KO mice produced more IFN-Is than WT mice, pDCs defined as CD11cintB220+CD8+/− were the most dominant cellular source for the sustained, higher IFN-I levels observed in the serum of LCMV-CL-13-infected Oasl1 KO mice at the very early stage of infection: pDCs in Oasl1 KO mice had more than 10-fold higher IFN-I mRNA levels than cDCs and Macs, and pDCs in Oasl1 KO mice had a more than 8-fold higher expression of IFN-I mRNAs than those in WT mice (Fig. 5D). However, the expression of OASL1 in these cells was not directly inverse-correlated with IFN-I expression level (Fig. 5D), indicating that other cellular factors also affect IFN-I expression levels (See Discussion).
We next asked whether IFN-I receptor signaling in Oasl1 KO mice is essential for the massive expansion of CD8+ T cells and the accelerated viral control in KO mice. To address this question, we performed in vivo blockade experiments using an antibody against the IFN-α/β receptor 1 (IFNAR-1), which efficiently blocks IFN-I receptor signaling caused by IFN-I [40], [41]. At 1.5 d after LCMV CL-13 infection, the time point just before the serum level of IFN-I began to show a real significant difference between WT and Oasl1 KO mice (Fig. 5A), Oasl1 KO mice were treated with either IFNAR-1 antibody or its isotype control, and WT mice were treated with phosphate-buffered saline (PBS) as another control. IFNAR-1 antibody-treated Oasl1 KO mice and PBS-treated WT mice displayed much more similar patterns in both body weight changes and serum viremia than isotype-control-treated Oasl1 KO mice (Figs. 6A and 6B). Furthermore, the spleen and blood of IFNAR-1 antibody-treated Oasl1 KO mice showed rather typical phenotypes of CD8+ T cells that have been observed in chronically-infected WT mice, such as a moderate frequency of LCMV-specific CD8+ T cells, high expression of PD-1, and impaired induction of effector cytokines (Fig. 6C and 6D). By contrast, isotype control-treated Oasl1 KO mice did not show such chronic infection phenotypes. Collectively, these data indicate that the better virus-specific CD8+ T-cell responses and accelerated virus control in Oasl1 KO mice need IFN-I receptor signaling triggered by sustained IFN-I in KO mice.
In the present study, we showed that Oasl1 KO mice produced sustained level of IFN-I during the very early phase (2–3 d p.i.) of chronic LCMV CL-13 infection, cleared the virus quickly (earlier than 15 d p.i.) from the blood circulation, and induced rapid expansion of virus-specific effector T cells. We also showed that Oasl1 KO mice cleared the virus even from the kidney, a place for long-lasting existence of LCMV CL-13, at the late phase of infection (by 130 d p.i.) and induced the production of polyfunctional virus-specific memory CD8+ T cells. Furthermore, we showed that IFN-I receptor signaling during the very early stage of LCMV CL-13 infection was necessary for the higher T-cell number, efficient T-cell differentiation, and early viral control in Oasl1 KO mice.
The sustained production of IFN-I and more effective viral clearance in Oasl1 KO mice after chronic LCMV CL-13 infection observed in the present study resembles our previous observation that Oasl1 KO mice are more resistant to acute infections with encephalomyocarditis virus and herpes simplex virus-1 and produce higher levels of IFN-I [32]. Because OASL1 is a specific translation inhibitor of IRF7, the IFN-inducible IFN-I master TF, the sustained and stronger expression of IFN-I in Oasl1 KO mice upon LCMV infection is thought to be caused by the release of such inhibition and effective activation of the more translated IRF7 protein in KO mice. Indeed, we observed that the IRF7 protein level in Oasl1 KO mice was much higher during the early phase of chronic LCMV CL-13 infection (Fig. 5C). Because most viruses induce IFN-I and in turn IRF7 expression and can activate the resulting IRF7 protein effectively, Oasl1 KO mice are expected to be resistant to most viral infections that do not produce viral effectors to degrade the function of IRF7. Consistent with this expectation, Oasl1 KO mice turned out to be more resistant to chronic LCMV CL-13 infection as well as acute LCMV Arm infection (Figs. 1F and S3), although LCMV is equipped with nucleoprotein (NP), a virus-derived IFN-I negative regulator that interferes with activation of IRF3 [42]–[44]. In both acute LCMV Arm and chronic LCMV CL-13 infections, OASL1 expression was similarly strongly induced (>20-fold) during the very early phase of infection, and then declined gradually (1–4 d p.i.) (Fig. S10). However, at later time points (9–15 d p.i.), only with LCMV CL-13 infection, OASL1 expression was still marginally up-regulated (Fig. S10), which would reflect the presence of viruses in the mice infected with LCMV CL-13 (Figs. 1F and S3). Therefore, as expected, the mRNA expression of OASL1 (as an ISG) is dominantly regulated by IFN-I induced upon viral infection.
The cellular source of IFN-I upon chronic LCMV CL-13 infection seems to include pDCs and other innate immune cell types such as cDCs and Macs as recently reported [45], [46]. Among these cells, we showed in our study that the critical cellular source for the sustained and higher serum IFN-I level in Oasl1 KO mice at the early stage of infection (2 d p.i. when the IFN-I level in serum became clearly stronger in the KO mice than in WT mice) was pDCs (CD11cintB220+CD8+/−), although cDCs and Macs of Oasl1 KO mice also produced significantly more IFN-I than those of WT mice (Fig. 5D). However, IFN-I expression in these cells was not inversely correlated with OASL1 expression in these cells, indicating that other factors also affect the expression of IFN-I. The pDC-dominant IFN-I production in Oasl1 KO mice at the early stage of LCMV CL-13 infection could be explained by the differences in expressed viral sensors and in IRF7 expression levels among these cells, with the demonstrated role of OASL1 as a specific IRF7 translation inhibitor. It has been well known that cDCs and Macs express viral sensors such as TLR3 and RIG-I/MDA5 that sense dsRNA produced during viral replication, including that of the ssRNA virus LCMV, whereas pDCs dominantly express TLR7 that sense ssRNAs and TLR9 that sense DNAs among known viral sensors [35], [47]. Because the IRF7 expression level in pDCs is much higher than that in cDCs and Macs at a steady state (data not shown) [48], [49], and cell membrane-associated receptor TLR7 can sense the ssRNA viral genome directly even without viral replication upon LCMV infection, pDCs are expected to produce IFN-I initially by sensing the viral genome first and inducing activation of IRF7 protein [4], [45]. After initial production of IFN-I, pDCs would amplify IFN-I production by inducing an IFN-I-mediated positive feedback signaling loop in an autocrine fashion, leading also to accumulation of IRF7 mRNA [4], [50]. In this later phase of IFN-I production, IRF7 protein levels would be much higher in Oasl1 KO pDCs than in WT pDCs because OASL1 protein derived from Oasl1 mRNA induced by the IFN-I-mediated positive feedback signaling loop would inhibit IRF7 translation in WT cells. Thus, Oasl1 KO pDCs could produce much higher levels of IFN-I at the very early stage of LCMV infection: we observed a more than 8-fold higher IFN-I level in KO cells than in WT cells (Fig. 5D).
In addition to reducing viral titers during the early phase of viral infection, such a sustained level of IFN-I in Oasl1 KO mice might act directly and indirectly on CD8+ T cells to enhance their numbers and function. IFN-I receptor signaling in CD8+ T cells triggered by the sustained IFN-I level in KO mice appears to be critical for the massive expansion of virus-specific CD8+ T cells, given that such signaling in CD8+ T cells is known to prevent apoptosis of T cells in vivo, allowing clonal expansion and differentiation upon viral infection [51]–[56]. CD4+ T cells could also be a direct target for IFN-I in KO mice because IFN-I was reported to act directly on CD4+ T cells to sustain their clonal expansion [57] and promote Th1 differentiation in vivo [58]–[60]. In addition to the direct IFN-I effect on T cells, IFN-I was reported to enhance the function of DCs by inducing maturation, such as upregulation of costimulatory molecules and MHC class molecules [61]–[63]. However, a sustained and increased level of IFN-I in Oasl1 KO mice does not seem to critically affect DC maturation in the setting of chronic LCMV infection because there was no apparent difference in the expression of costimulatory molecules and MHCs in DCs isolated from WT and Oasl1 KO mice (Fig. S7). The level of IFN-I produced in WT mice could be sufficient to induce DC maturation, and its higher and sustained level in Oasl1 KO mice might not create any additional effects on DC maturation. It would be interesting, in the future, to determine whether other types of cells influenced by IFN-I also contribute to the better virus-specific CD8+ T-cell response observed in Oasl1 KO mice.
The complete elimination of chronic virus and efficient induction of the adaptive immune response in Oasl1 KO mice observed in this study is extraordinary because WT mice infected with chronic LCMV CL-13 suffer from a sustained viral load throughout their life spans and have fundamentally dysfunctional T cells [34]. There has been considerable number of molecules known to significantly affect viral persistence like an OASL1. For example, functional blockage of IL-10, a well-known immunosuppressive molecule, during the chronic phase of LCMV CL-13 infection reduced viral load; IL-21 receptor signaling in CD8+ T cells was critical for limiting LCMV viral load in vivo; IL-7, the major T-cell survival factor, enhanced LCMV viral clearance; the functional blockage of PD-1, a T-cell-intrinsic signaling inhibitor, led to LCMV viral load reduction [19], [64]–[66]. However, most of these previously characterized molecules, if not all, are not related directly to IFN-I production and its regulation but are mostly involved in controlling adaptive immune responses. Therefore, our report demonstrating the critical role of OASL1, as a direct negative regulator of IFN-I production, in permitting chronic viral persistence is, to our knowledge, the first demonstration of a role of any host IFN-I negative regulator in viral clearance in the setting of chronic infection.
The negative role of OASL1, as an IFN-I negative regulator, in viral clearance suggests that inhibition of the expression/function of OASL1 or IFN-I treatment during the early phase of chronic viral infections could prevent viral persistence. Indeed, a recent study showed that IFN-I treatment in WT mice at 3 d and 5 d p.i. (when the IFN-I level is downregulated after its initial peak) after LCMV CL-13 infection dramatically reduced the LCMV viral load and promoted virus-specific CD8 T-cell responses [46]. This observation is consistent with our data that a sustained level of IFN-I at approximately 2–4 d p.i. in Oasl1 KO mice led to rapid control of viremia and functional virus-specific T-cell responses. However, in the previous study, IFN-I treatment in WT mice later than 1 week of infection did not have any meaningful effect on viral load [46], suggesting that the timing of IFN-I treatment is critical for controlling viremia and restoring the T-cell response and that other factors may degrade IFN-I-mediated responses. Given that the ineffectiveness of exogenous IFN-I at the later phase of chronic virus infection, it is worthwhile to note that Oasl1 expression in pDCs was still sustained at a later phase (30 d p.i.) of chronic LCMV CL-13 infection but not of acute LCMV Arm infection, whereas Ifna/b1 expression in the pDCs returned to the basal level by that phase in both LCMV infections (Fig. S11). Therefore, it is possible that, at the later phase of chronic viral infection, exogenously administered IFN-I itself is not sufficient to induce an adequate level of IFN-I production for viral clearance because the production of IRF7 protein, the key player necessary for going through the IFN-I-mediated positive feedback signaling loop, would still be inhibited by OASL1 protein. In this regard, suppression of OASL1 function or expression during the chronic phase of LCMV infection may help to eliminate the virus by releasing OASL1-mediated suppression of IFN-I production, although we cannot exclude the possibility that other factors, including other IFN-I negative regulators, still promote viral persistence.
Our results showing that reversal of the defective production of IFN-I during chronic virus infection by nullifying OASL1 can inhibit viral persistence raise the possibility that inhibiting the functions of other negative regulators acting on the process of IFN-I production and/or IFN-I receptor signaling pathways would improve the host defense against chronic viral infections. Peptidyl-prolyl cis/trans isomerase NIMA-interacting 1 (Pin 1), which induces degradation of IRF3 protein, and two ubiquitin E3 ligases, tripartite motif-containing 21 (TRIM21) and RTA-associated ubiquitin ligase (RAUL), which induce the degradation of both IRF3 and IRF7 proteins [28], [67], are host factors more directly involved in negatively regulating IFN-I production. Negative host factors involved in IFN-I receptor signaling pathways include inhibitors of the JAK-STAT signaling pathway downstream of the IFN-I receptor such as Suppressor of cytokine signaling (SOCS), Src homology 2-containing protein tyrosine phosphatase (SHP), and protein inhibitors of activated STAT (PIAS) [68], [69]. Because most of the currently known negative factors not only regulate IFN-I but also other cytokines [28], OASL1 is a rather unique host negative factor regulating IFN-I production that specifically functions in IRF7 protein production. In the future, a combinatorial approach for controlling the expression and/or function of OASL1 and regulating diverse other targets might be worthy of investigation. However, blockade of virus-encoded IFN-I suppressors may also be necessary for these approaches to be more effective.
All animal experiments were performed in accordance with the Korean Food and Drug Administration (KFDA) guidelines. Protocols were reviewed and approved by the Institutional Animal Care and Use Committee (IACUC) of the Yonsei Laboratory Animal Research Center (YLARC) at Yonsei University (Permit Number: 2007-0001).
C57BL/6 mice were purchased from the Jackson Laboratory, and LCMV CL-13 epitope-specific TCR transgenic P14 Thy1.1 mice were obtained from the Emory Vaccine Center. Oasl1 KO mice were derived from our previous study [32]. All mice were maintained in the specific pathogen-free facility of the YLARC at Yonsei University. Six- to ten-week-old littermate female or male C57BL/6 WT and Oasl1 KO mice were infected with 2×106 plaque-forming units (PFUs) of LCMV CL-13 or LCMV Arm diluted in serum-free RPMI medium per 20 g of mouse body weight by intravenous (i.v.) injection. For serum virus titration, three to five drops of blood were collected at the indicated time points p.i., and the serum was directly stored at −70°C. For tissue virus titration, small pieces of the spleen, lung, liver, and kidney, were put in Dulbecco's modified Eagle's medium (DMEM) containing 1% fetal bovine serum (FBS) (Gibco) and stored at −70°C. The tissues were later homogenized completely using a homogenizer (Kinematica) before titration. Viral titers from sera or homogenized samples were determined by plaque assay on Vero cells as previously described [33].
PBMCs were isolated from peripheral blood using density gradient centrifugation underlaid with histopaque-1077 (Sigma-Aldrich). Lymphocytes from tissues, including the spleen, lung, and liver, were isolated as previously described [9]. Lungs and livers were perfused with ice-cold PBS before collection for lymphocyte isolation. For phenotypic analysis of virus-specific CD8+ T cells derived from peripheral blood and tissues, single-cell suspensions were stained with CD4 (RM4-5), CD8a (53-6.7), CD44 (IM7), CD62L (MEL-14), CD127 (A7R34), and PD-1 (RMP1-30) antibodies in the presence of each tetramer. H-2Db tetramers bound to GP33–41 and GP276–286 peptides were generated and used as previously described [70]. Pieces of spleen were digested with 1 mg/ml of type II collagenase (Worthington Biochemicals) and 1 mg/ml of bovine pancreatic DNase I (Sigma-Aldrich) as previously described for the analysis of splenic DCs, Macs, and PMNs/Mos at the indicated time points after LCMV CL-13 infection [71]. The resulting splenocytes were mixed with CD16/CD32 (2.4G2) antibodies and stained with CD3e (145-2C11), CD19 (1D3), CD49b (DX5), CD11c (HL-3), CD11b (M1/70), F4/80 (BM8), CD8a (53-6.7), and B220 (RA3-6B2) antibodies. Splenocytes were co-stained with CD40 (3/23), CD80 (16-10A1), CD86 (GL1), H-2Kb (AF6-88.5), I-A/I-E (M5/114.15.2), and PD-L1 (MIH5) antibodies, or their isotype antibodies to detect various accessory molecules in different subset of DCs. After 5 h of in vitro restimulation of splenocytes with 0.2 µg/ml of LCMV GP33–41 or GP276–286 peptide to induce a CD8+ T-cell response, and with 1.0 µg/ml of GP66–80 peptide to induce a CD4+ T-cell response, in the presence of Golgi plug/Golgi stop (BD Biosciences), intracellular staining for cytokines was performed using IFN-γ (XMG1.2), TNF-α (MP6-XT22), and IL-2 (JES6-5H4) antibodies. To measure the proliferation capability ex vivo, splenocytes were stained with a Ki67 antibody (B56). For direct analysis of apoptosis ex vivo, splenocytes were briefly incubated with Annexin V and 7AAD (BD Biosciences). All antibodies were purchased from BD Biosciences except for CD127, CD11b, F4/80, I-A/I-E, PD-L1 (eBioscience), and PD-1 antibodies (BioLegend). The Live/Dead fixable dead cell Stain kit (Invitrogen) was used to remove the dead cell population in most staining procedures, except in ex vivo apoptosis staining. All stained samples were read using FACSCalibur and FACSCantoII instruments (BD Biosciences) and analyzed using FlowJo software (Tree Star).
The splenocytes digested with type II collagenase and DNase I were incubated with a cocktail of biotin-conjugated antibodies and anti-biotin beads (Miltenyi Biotech) and were subsequently depleted of T, B, and natural killer (NK) cells. The resulting splenocytes were mixed with CD16/CD32 (2.4G2) antibody and incubated with CD11c beads (Miltenyi Biotech). After magnetic separation, the bead-attached cells were stained with CD11c (HL-3), CD8a (53-6.7), and B220 (RA3-6B2) antibodies for sorting cDCs and pDCs, and the unattached cells were stained with CD11b (M1/70) and F4/80 (BM8) antibodies for sorting Macs and PMNs/Mos. The stained cells were sorted using a FACSAriaII instrument (BD Biosciences). For adoptive transfer of P14 Thy1.1+ CD8+ T cells, cells were isolated from the spleen of B6 P14 Thy1.1+ mice using a CD8+ T-cell isolation kit (Miltenyi Biotec). Mice were infected with LCMV CL-13 at 1 d after the adoptive transfer of 5×103 P14 Thy1.1+ CD8+ T cells via the tail vein into naïve WT or Oasl1 KO mice.
Sera were obtained from WT and KO mice at the indicated time points after LCMV CL-13 infection for measurement of cytokines. The IFN-I levels in the sera were measured using VeriKine Mouse Interferon Alpha and Beta ELISA kits (PBL interferon source). Other cytokine levels in the sera were measured using BD CBA mouse inflammation kits (BD Biosciences) according to the manufacturer's instructions. For immunoblot analysis, snap-frozen tissues were homogenized in NP-40 buffer [50 mM Tris-HCl (pH 8.0), 150 mM NaCl, and 1% NP-40] with a protease inhibitor ‘cocktail’ (1 mM PMSF, 10 µg/ml aprotinin, 5 µg/ml pepstatin and 5 µg/ml leupeptin) by sonication, and the tissue lysate was centrifugated at 17,000× g for 10 min at 4°C for removal of tissue debris. Equal amounts of proteins were analyzed by immunoblotting using IRF7 (51-3300, Zymed), IRF3 (51-3200, Zymed), β-actin (A2228, Sigma-Aldrich) and OASL1 [32] antibodies, and signals developed with Amersham ECL reagents were detected using the ImageQuant LAS 4000 system (GE Healthcare).
To block IFN-I receptor signaling in vivo, 0.5 mg (per mouse) of anti-mouse IFNAR-1 monoclonal antibody (clone MAR1-5A3; BD Biosciences) was administered to Oasl1 KO mice 1.5 d p.i. by i.v. injection. PBS and mouse IgG1 (MOPC-21, isotype control) were injected into WT and KO mice, respectively, at the same time point. Body weight was monitored every 2, 3, or 4 d, and serum was collected at 6, 8, and 15 d p.i. All mice were sacrificed at 35 d p.i., and their spleens were harvested and analyzed.
Tissues from animals were snap-frozen in liquid nitrogen after perfusion with 5 mM ethylenediaminetetraacetic acid in PBS and sorted cells were pelleted by centrifugation before lysis. Total RNA from the tissues and cells were purified using TRIzol RNA Isolation Reagent (Invitrogen), and cDNAs were synthesized using SuperScript II Reverse Transcriptase according to the manufacturer's protocol (Invitrogen). The expression levels of individual genes were measured by quantitative PCR (qPCR) performed using the CFX96 real-time PCR detection system (Bio-Rad) with the following gene-specific forward (F) and reverse (R) primers: Oasl1 F: CCAGGAAGAAGCCAAGCACCATC and R: AGGTTACTGAGCCCAAGGTCCATC; Irf7 F: CAGCAGCAGTCTCGGCTTGTG and R: TGACCCAGGTCCATGAGGAAGTG; Irf3 F: CTGGACGAGAGCCGAACGAG and R: TGTAGGCACCACTGGCTTCTG; Ifna2 F: AAGGACAGGCAGGACTTTGGATTC and R: GATCTCGCAGCACAGGGATGG; Ifna5/6/13 F: AGGACTCATCTGCTGCATGGAATG and R: CACACAGGCTTTGAGGTCATTGAG; Ifnb1 F: CCACTTGAAGAGCTATTACTG and R: AATGATGAGAAAGTTCCTGAAG; Gapdh F: GGCAAATTCAACGGCACAGTCAAG and R: TCGCTCCTGGAAGATGGTGATGG. The relative mRNA expression normalized to the Gapdh signal was shown. The qPCR reaction conditions were as follows: after initial denaturation of the template for 3 min at 95°C, 45 thermal cycles of 15 sec at 95°C, 25 sec at 60°C, and 30 sec at 72°C were run in a final volume of 20 µl using SYBR Green I dye for PCR product detection (Qiagen) according to the manufacturer's instructions.
Statistical analysis of all presented data was performed using a two-tailed unpaired Student's t-test using Prism 5.0 software (GraphPad). A P value less than 0.05 was considered statistically significant.
The list of GenBank accession numbers for the major genes and proteins that are mentioned in the text are as follows: OASL1, NP_660210 (NM_145209); TLR3, NP_569054 (NM_126166); TLR7, NP_573474 (NM_133211); TLR9, NP_112455 (NM_031178); RIG-I, NP_766277 (NM_172689); MDA5, NP_001157949 (NM_001164477)); IFN-α, NP_034632 (NM_010502); IFN-β, NP_034640 (NM_010510); PD-1, NP_032824 (NM_008798); TIM-3, NP_599011 (NM_134250); CTLA-4, NP_033973 (NM_009843); LAG-3, NP_032505 (NM_008479); IRF3, NP_058545 (NM_016849); IRF7, NP_058546 (NM_016850); Pin 1, NP_075860 (NM_023371); TRIM21, NP_001076021 (NM_001082552); RAUL, NP_001003918 (NM_001003918); IFNAR, NP_034638 (NM_010508); PIAS, NP_001159421 (NM_001165949).
|
10.1371/journal.pbio.1002427 | Excessive Osteocytic Fgf23 Secretion Contributes to Pyrophosphate Accumulation and Mineralization Defect in Hyp Mice | X-linked hypophosphatemia (XLH) is the most frequent form of inherited rickets in humans caused by mutations in the phosphate-regulating gene with homologies to endopeptidases on the X-chromosome (PHEX). Hyp mice, a murine homologue of XLH, are characterized by hypophosphatemia, inappropriately low serum vitamin D levels, increased serum fibroblast growth factor-23 (Fgf23), and osteomalacia. Although Fgf23 is known to be responsible for hypophosphatemia and reduced vitamin D hormone levels in Hyp mice, its putative role as an auto-/paracrine osteomalacia-causing factor has not been explored. We recently reported that Fgf23 is a suppressor of tissue nonspecific alkaline phosphatase (Tnap) transcription via FGF receptor-3 (FGFR3) signaling, leading to inhibition of mineralization through accumulation of the TNAP substrate pyrophosphate. Here, we report that the pyrophosphate concentration is increased in Hyp bones, and that Tnap expression is decreased in Hyp-derived osteocyte-like cells but not in Hyp-derived osteoblasts ex vivo and in vitro. In situ mRNA expression profiling in bone cryosections revealed a ~70-fold up-regulation of Fgfr3 mRNA in osteocytes versus osteoblasts of Hyp mice. In addition, we show that blocking of increased Fgf23-FGFR3 signaling with anti-Fgf23 antibodies or an FGFR3 inhibitor partially restored the suppression of Tnap expression, phosphate production, and mineralization, and decreased pyrophosphate concentration in Hyp-derived osteocyte-like cells in vitro. In vivo, bone-specific deletion of Fgf23 in Hyp mice rescued the suppressed TNAP activity in osteocytes of Hyp mice. Moreover, treatment of wild-type osteoblasts or mice with recombinant FGF23 suppressed Tnap mRNA expression and increased pyrophosphate concentrations in the culture medium and in bone, respectively. In conclusion, we found that the cell autonomous increase in Fgf23 secretion in Hyp osteocytes drives the accumulation of pyrophosphate through auto-/paracrine suppression of TNAP. Hence, we have identified a novel mechanism contributing to the mineralization defect in Hyp mice.
| X-linked hypophosphatemia (XLH) is the most frequent form of inherited rickets in humans. A mouse model of XLH, known as Hyp, is characterized by exceptionally low serum phosphate and vitamin D levels, increased serum levels of the hormone fibroblast growth factor-23 (Fgf23), and impaired bone mineralization. Fgf23 is secreted from two classes of bone cells known as osteoblasts and osteocytes. Fgf23 increases urinary phosphate excretion and suppresses vitamin D hormone production in the kidney. Although Fgf23 is known to be responsible for lower blood phosphate and vitamin D hormone levels in Hyp mice, its putative role as a signaling factor causing impaired mineralization has not been explored. We recently reported that Fgf23 is a suppressor of tissue nonspecific alkaline phosphatase (Tnap) gene expression via FGF receptor-3 (FGFR3) signaling in osteoblasts, leading to inhibition of mineralization through accumulation of the TNAP substrate pyrophosphate. Pyrophosphate is a potent inhibitor of mineralization. Using a combination of cell culture and animal models, we report that the increase in osteocyte Fgf23 secretion of Hyp mice leads to FGFR3-mediated suppression of TNAP with subsequent accumulation of pyrophosphate. Hence, we have identified a novel signaling mechanism by which excessive osteocytic secretion of Fgf23 contributes to the mineralization defect in Hyp mice.
| X-linked hypophosphatemia (XLH) is the most frequent form of inherited rickets in humans. XLH is caused by inactivating mutations in the phosphate-regulating gene with homologies to endopeptidases on the X-chromosome (PHEX) [1–3]. Similarly, a loss-of-function deletion in Phex, the murine homologue of PHEX, leads to an XLH-like phenotype in Hyp mice, a well-known animal model for XLH [4–6]. PHEX/Phex is predominantly expressed in bone and teeth and at lower levels in muscle, skin, brain, and lungs [7,8]. Both XLH patients and Hyp mice are characterized by hypophosphatemia, impaired bone mineralization, inappropriately low serum vitamin D hormone (1,25(OH)2D3), and increased circulating intact fibroblast growth factor-23 (FGF23) [9–11]. FGF23 is a phosphaturic hormone, mainly produced by osteoblasts and osteocytes in response to increased extracellular phosphate and circulating 1,25(OH)2D3 [12]. In renal proximal tubules, FGF23 suppresses the membrane expression of the type II sodium-phosphate cotransporters Npt2a and Npt2c, which are necessary for the urinary reabsorption of phosphate [13]. In addition, FGF23 suppresses the renal proximal tubular expression of 1α-hydroxylase [14], the key enzyme responsible for vitamin D hormone production. Fgf23 requires the obligatory coreceptor α-Klotho (Klotho) to bind to the ubiquitously expressed fibroblast growth factor receptor 1c (FGFR1c) [15,16]. Hence, the hormonal actions of Fgf23 are restricted, at least at physiological concentrations, to tissues expressing Klotho such as proximal and distal tubules in the kidney, parathyroid gland, choroid plexus in the brain, and sinoatrial node in the heart [13,17].
The molecular mechanisms why loss of PHEX/Phex function leads to increased FGF23 secretion in osteoblasts and osteocytes are still incompletely understood. PHEX is an ectoenzyme thought to be involved in the proteolytic processing of extracellular matrix (ECM) proteins. Earlier studies in Hyp mice revealed aberrant processing of SIBLING (Small Integrin-Binding Ligand, N-linked Glycoprotein) proteins such as matrix extracellular phosphoglycoprotein (MEPE) [18], causing accumulation of acidic serine- and aspartate-rich MEPE-associated motif (ASARM) peptides. ASARM peptides are potent inhibitors of mineralization, and are thought to be at least partially responsible for the mineralization defect observed in Hyp mice [19]. Another substrate of PHEX is the ECM protein osteopontin (OPN), a well-known mineralization inhibitor that binds to hydroxyapatite (HA) crystals and blocks the deposition of HA onto ECM [20]. In addition to ASARM peptides, OPN was shown to be increased in bones of Hyp mice [21]. Because the mineralization defect present in Dentin matrix protein-1 (Dmp-1)-deficient mice also leads to overexpression of Fgf23 [22], it is currently thought that osteocytes respond to impaired mineralization by increased Fgf23 secretion. It is interesting to note in this context that ablation of Fgfr1 in bone partially rescues the excessive Fgf23 secretion in Hyp mice, suggesting that Fgfr1-mediated signaling may somehow be involved in the mechanism how osteocytes sense mineralization in the surrounding matrix [23]. In addition, long-term inhibition of FGFR by a pan-FGFR inhibitor in Hyp and Dmp-1-deficient mice leads to normalization of serum phosphate and calcium and improves mineralization [24]. Besides disturbed mineralization, defective phosphate sensing in osteoblasts has also been implicated to play a role in the augmented Fgf23 secretion in Hyp mice [25].
Although the exact mechanism driving Fgf23 secretion in Phex and Dmp-1-deficient models has remained elusive thus far, several lines of evidence suggest that increased circulating Fgf23 is a major pathogenetic factor in XLH patients and Hyp mice, leading to hypophosphatemia and subsequently impaired bone mineralization. Firstly, extraskeletal overexpression of FGF23 also causes hypophosphatemia and osteomalacia [26,27]. Secondly, ablation of Fgf23 in Hyp mice recapitulates the Fgf23-null phenotype [28,29]. Thirdly, treatment of Hyp mice with anti-Fgf23 antibodies normalizes serum phosphate and vitamin D hormone levels, decreases osteoid volume, and improves bone mineralization [9,15,24]. All these findings suggest that excessive Fgf23 secretion is the major driving force behind the Hyp phenotype.
However, because osteoblasts isolated from Hyp mice fail to mineralize in a normal fashion in vitro [30], and dietary phosphate supplementation attempting to correct hypophosphatemia did not rescue the osteomalacia in Hyp mice [31], it is likely that the mineralization defect in Hyp mice has at least two components, namely systemic hypophosphatemia plus independent alterations in the ECM [32]. The relative contribution of local accumulation of ASARM peptides and of OPN in the ECM versus the endocrine phosphaturic effect of Fgf23 to the osteomalacia observed in Hyp mice is currently unclear.
We recently discovered that FGF23 suppresses tissue nonspecific alkaline phosphatase (TNAP) transcription and leads to decreased local inorganic phosphate (Pi) production as well as accumulation of pyrophosphate (PPi) by a Klotho-independent, FGFR3-mediated signaling axis in osteoblasts [33]. PPi is another well-known inhibitor of mineralization produced by osteoblasts and osteocytes. Increased levels of PPi in the ECM are known to impair the mineralization process by binding to HA crystals [34–36]. Conversely, absence of PPi in the ECM either via genetic ablation of its intracellular-to-extracellular transporter progressive ankylosis (ANK) [37] or ablation of ectonucleotide pyrophosphatase/phosphodiesterase 1 (ENPP1) [38], an enzyme which produces PPi from ATP, results in hypermineralization of bones. Increased levels of PPi in the ECM can be a consequence of two different mechanisms, either increased production and transportation of PPi to the ECM, or decreased hydrolysis of PPi by TNAP, leading to accumulation of PPi in the ECM.
Based on our recent finding that Fgf23 is a regulator of Tnap transcription, we hypothesized that excessive Fgf23 secretion in Hyp osteocytes could locally contribute to defective mineralization by suppressing TNAP and increasing PPi concentrations. Here, we report that the PPi concentration is indeed increased in Hyp bones, and that Tnap expression is decreased in Hyp-derived osteocyte-like cells ex vivo and in vitro. In addition, we show that blocking of increased Fgf23-FGFR3 signaling in Hyp-derived osteocyte-like cells partially restores the suppression of TNAP expression, phosphate production, and mineralization in vitro. Thus, we have identified a novel mechanism contributing to the defective mineralization in Hyp mice.
It is well known that Hyp mice are characterized by hypophosphatemia, hypocalcemia, impaired bone mineralization, and increased serum Fgf23 [9]. This was confirmed in our study. Three-month-old male Hyp mice used in our experiments were hypophosphatemic and hypocalcemic, exhibited elevated serum alkaline phosphatase (ALP) activity and increased serum intact Fgf23 (Fig 1A) and showed impaired bone mineralization as evidenced by widened osteoid seams and enlarged osteocyte lacunae in histological bone sections (Fig 1B). Since it was previously reported that OPN protein expression is increased in Hyp mice [21], we quantified OPN protein expression in femur extracts from wild-type (WT) and Hyp mice by western blotting. As shown in Fig 1C, OPN protein expression was higher in Hyp femur extracts compared to WT mice. Immunohistochemistry confirmed increased OPN protein expression in Hyp compared to WT bones (Fig 1C, lower panel). According to our hypothesis, the concentration of PPi should be increased in Hyp bones. To initially test whether this hypothesis may be worth pursuing, we quantified the amount of PPi in WT and Hyp mice femur extracts. As shown in Fig 1D, the PPi concentration in Hyp bones was indeed higher compared to WT bones.
Our hypothesis predicts that increased Fgf23 secretion in bone cells from Hyp mice would suppress TNAP and would subsequently lead to accumulation of PPi. However, in contrast to our hypothesis, serum ALP in Hyp was actually higher compared to WT mice (Fig 1A). However, we reasoned that the inhibitory effect of Fgf23 on TNAP might be cell-specific in bone and might only occur in osteocytes where the Fgf23 concentrations in the extracellular fluid are probably highest. To investigate a potential cell-specific effect of Fgf23 on osteoblasts and osteocytes, we isolated osteoblast- and osteocyte-rich fractions from femurs of WT and Hyp mice, using a sequential digestion technique [39]. To confirm the successful isolation of osteoblasts and osteocytes, we analyzed the mRNA abundance of the osteoblast-specific marker osteocalcin (Ocn) [40], and of the osteocyte-specific marker sclerostin (Sost) [41]. Fractions (F) 1 and 2 were discarded because of the high contamination with other cell types. In both WT and Hyp mice, F-3 showed higher levels of Ocn mRNA expression compared to the other fractions (Fig 2A), suggesting that this was an osteoblast-rich fraction. Conversely, Sost mRNA expression was low in F-3, F-4, and F-5, and increased 5- to 40-fold in F-6/7 and F-8/9, respectively (Fig 2A). Based upon these results we considered fractions 3–5 as osteoblast-rich, and F-6/7 and F-8/9 as osteocyte-rich in both WT and Hyp femurs (Fig 2A).
Analysis of gene expression in osteoblast-rich fractions revealed lower mRNA expression of Ank and Enpp1 and ~50-fold higher Tnap expression in Hyp- versus WT-derived F-3 (Fig 2B). The most pronounced differences between WT and Hyp mice were observed in F-3. Interestingly, Fgf23 and Opn mRNA expression remained unchanged between the genotypes in all three fractions, confirming an earlier report that osteoblastic Fgf23 production is not different between WT and Hyp mice [39]. The mRNA expression of Fgfr1 and Fgfr3 was lower in Hyp- versus WT-derived F-3 and/or F-5. In contrast, the mRNA abundance of Ank, Enpp1, Opn, Fgf23, Fgfr1, and Fgfr3 was distinctly increased in Hyp-derived osteocyte-rich fractions F-6/7 and/or F-8/9 (Fig 2C). Most interestingly, Tnap mRNA abundance was decreased by 80%–90% in Hyp-derived osteocyte-rich fractions F-6/7 and F-8/9, relative to WT-derived osteocytes (Fig 2C). In accordance with the mRNA data, western blotting analysis of the pooled protein samples showed unchanged TNAP protein abundance in osteoblast-rich fractions F-3/4/5 but lower TNAP protein abundance in osteocyte-rich fractions F-6/7/8/9 isolated from Hyp mice, relative to WT controls (Fig 2D).
Collectively, these results corroborate the notion that osteocytes are the major source of the increased circulating Fgf23 levels in Hyp mice. Furthermore, our data show that there is not only increased OPN expression, but also increased mRNA expression of PPi-regulating factors such as Ank and Enpp1 in Hyp-derived osteocytes. Collectively, the observed changes in gene expression of Ank, Enpp1, and Tnap in Hyp-derived osteocyte-rich cell fractions are able to explain the accumulation of PPi in the ECM of Hyp bones.
To further examine whether osteoblasts and osteocytes isolated from Hyp mice differentially express PPi-regulating genes and Tnap in a cell autonomous fashion, we moved from the above mentioned ex vivo approach to an in vitro model. To this end, we isolated calvarial osteoblasts from newborn WT and Hyp mice, and differentiated the cells up to 22 d. At day 12, only little mineralized nodule formation was observed, and cells expressed maximum levels of Ocn mRNA, whereas at day 22, more mineralized nodules were formed and the mRNA expression of Sost was highest (Fig 3A). Therefore, we considered cells harvested at day 12 as differentiated osteoblasts and cells harvested at day 22 as osteocyte-like cells.
mRNA expression analysis at day 12 revealed no significant differences in Ank, Enpp1, and Opn expression between WT osteoblasts and Hyp osteoblasts (Fig 3B). Fgfr1 mRNA abundance was up-regulated, whereas Fgfr3 mRNA expression was lower in Hyp versus WT osteoblasts (Fig 3B). Fgf23 mRNA abundance was higher in Hyp compared to WT osteoblasts already at day 12 (Fig 3B). We further analyzed if this increase in Fgf23 mRNA expression led to increased Fgf23 secretion in the cell culture medium. We found ~60-fold higher concentrations of intact Fgf23 in the culture medium of Hyp osteoblasts (Fig 3B). Despite increased Fgf23 mRNA expression and secretion, Tnap mRNA expression was increased in Hyp compared with WT osteoblasts (Fig 3B). Assessment of TNAP protein expression using western blotting analysis and of TNAP enzyme activity using BCIP/NBT staining showed similar levels of protein expression and activity in WT and Hyp osteoblasts at day 12 (Fig 3B). It is well known that TNAP is responsible for Pi production in vitro during differentiation by cleaving β–glycerophosphate, a component of the differentiation medium [42]. Therefore, we assessed Pi concentration in the culture medium as readout for TNAP enzyme activity. Pi concentration in the medium of Hyp osteoblasts was lower than that of WT osteoblasts (Fig 3B). We currently don’t have a good explanation for the discrepancy between Tnap mRNA expression and enzyme activity at the 12-day time point in Hyp-derived osteoblasts. As another readout for TNAP enzyme activity, we analyzed PPi concentration in the cell culture medium. However, no significant changes in PPi concentration were observed between cell culture medium from WT and Hyp osteoblasts at day 12 (Fig 3B).
In analogy to the osteocyte-rich fractions isolated from Hyp femurs, mRNA abundance of Ank, Enpp1, Opn, Fgf23, and Fgfr3 was increased, whereas mRNA expression of Tnap was decreased, in Hyp compared to WT osteocyte-like cells differentiated for 22 d (Fig 3C). BCIP/NBT staining and western blotting analysis confirmed the decreased TNAP protein expression and activity in Hyp versus WT osteocyte-like cells (Fig 3C). In accordance with decreased Tnap expression, Pi concentration was lower and PPi concentration was higher in the cell culture medium from Hyp versus WT osteocyte-like cells (Fig 3C). Similar to our findings in osteoblast-like cells, intact Fgf23 in the culture medium was increased in Hyp-derived osteocyte-like cells (Fig 3C). Taken together, our data suggest that the up-regulation in Ank, Enpp1, and Opn, as well as the downregulation in Tnap mRNA and protein expression are cell autonomous effects in Hyp osteocyte-like cells.
To validate our ex vivo and in vitro finding that TNAP is suppressed in Hyp-derived osteocyte-like cells, we examined TNAP enzyme activity in osteoblasts and osteocytes in sections of femurs from WT and Hyp mice. As shown in Fig 4A, TNAP enzyme activity was profoundly suppressed in osteocytes, but not in osteoblasts, of Hyp compared with WT mice, corroborating the ex vivo and in vitro data.
A puzzling finding in our in vitro experiments with calvarial cells was that the up-regulated Fgf23 secretion observed in both osteoblast- and osteocyte-like cells from Hyp mice suppressed Tnap mRNA and protein abundance only in osteocyte-like cells but not in osteoblasts. To rule out that this finding was due to the calvarial origin of the cells, we isolated osteoblasts from femurs of newborn WT and Hyp mice and differentiated them for 12 and 22 d. Similar to calvarial cells, osteocalcin expression was higher at day 12, whereas Sost expression was higher at day 22 compared to day 12, consistent with a differentiated osteoblast-like phenotype at day 12 and an osteocyte-like phenotype at day 22 (S1A Fig). At day 12, femoral osteoblasts isolated from Hyp mice showed increased Tnap mRNA expression, decreased phosphate production, but unchanged BCIP/NBT staining relative to WT cells (S1B and S1C Fig). After 22 d of differentiation, Tnap mRNA abundance, phosphate production, and BCIP/NBT staining were decreased in Hyp versus WT cells (S1B and S1C Fig). Fgf23 mRNA abundance was increased in Hyp versus WT cells at day 12 and 22 (S1B Fig). To test the differential sensitivity of calvarial versus femoral osteoblasts and osteocytes to recombinant FGF23 (rFGF23), we treated WT and osteoblasts and osteocyte-like cells with different doses of rFGF23, and monitored Tnap mRNA expression. The rFGF23-induced suppression of Tnap mRNA expression was generally similar in calvarial versus femoral osteoblasts and osteocyte-like cells from WT and Hyp mice (S2 Fig). However, in line with a lower sensitivity of Hyp-derived osteoblasts, higher concentrations of rFGF23 were needed to suppress Tnap mRNA in Hyp femoral and calvarial osteoblasts, relative to osteocytes (S2 Fig). Hence, the differences in TNAP expression and the response to pharmacological treatment with rFGF23 between osteoblasts and osteocyte-like cells isolated from Hyp mice were similar in cells derived from calvarial and femoral origin.
We previously reported that FGF23 inhibits Tnap transcription via FGFR3 [33]. The abovementioned increase in Fgfr3 mRNA in Hyp osteocyte-rich fractions and osteocyte-like cells relative to WT cells would be consistent with the notion that the up-regulation of FGFR3 during osteocytic differentiation is the pivotal process making Hyp osteocytes more responsive to the suppressive effect of Fgf23 on Tnap transcription. To confirm the up-regulation of Fgfr3 mRNA during osteocytic differentiation in vivo, we performed in situ mRNA expression analysis in frozen femur sections, employing laser capture microdissection (LCM), a technique which we recently developed [43]. We found that Fgf23 mRNA abundance was ~3-fold higher in WT osteocytes than in WT osteoblasts (Fig 4B). Relative to WT osteoblasts, Fgf23 mRNA expression was ~6–7-fold higher in Hyp osteoblasts and osteocytes (Fig 4B). In accordance with our in vitro and ex vivo data, Tnap mRNA expression was increased in Hyp osteoblasts, but suppressed in Hyp osteocytes, relative to WT osteoblasts and osteocytes, respectively (Fig 4B). Fgfr1 mRNA expression was higher in osteoblasts and osteocytes of Hyp mice, relative to WT controls (Fig 4B). Notably, Fgfr3 mRNA abundance was ~5-fold higher in WT osteocytes than in WT osteoblasts, whereas Fgfr3 mRNA expression was suppressed in Hyp versus WT osteoblasts, but profoundly up-regulated in Hyp osteocytes (Fig 4B). Hyp osteocytes showed ~70-fold higher Fgfr3 mRNA abundance than Hyp osteoblasts. Taken together, these findings support the notion that the distinct up-regulation in Fgfr3 mRNA expression during osteocytic differentiation especially in Hyp mice is permissive to the Fgf23-mediated suppression of Tnap transcription in vitro and in vivo.
Next, we examined whether the changes in the mRNA expression of PPi-regulating genes in Hyp osteocyte-like cells are causatively linked to increased Fgf23 secretion. To this end, we isolated osteoblasts from newborn Hyp and WT mice, and treated osteocyte-like cells differentiated for 22 d with either neutralizing anti-FGF23 antibody (FGF23 AB) or an FGFR3 inhibitor for 24 h. In analogy to the experiments shown in Fig 3, Hyp osteocyte-like cells expressed distinctly lower Tnap mRNA, and showed lower Pi but higher PPi concentrations in cell culture medium, relative to WT cells (Fig 5A and 5B). With the exception of an up-regulation in Opn mRNA abundance, treatment of osteocyte-like cells with either FGFR3 inhibitor or FGF23 AB did not have significant effects in WT cells, but increased Tnap and Opn, and lowered Ank and Enpp1 mRNA expression in Hyp osteocyte-like cells, relative to vehicle-treated cells (Fig 5A and 5B). However, both treatments did not restore Tnap mRNA expression in Hyp osteocyte-like cells to WT control levels. In concordance with the increased Tnap mRNA expression after inhibition of Fgf23 signaling, treatment of Hyp osteocyte-like cells with either FGFR3 inhibitor or FGF23 AB increased Pi and decreased PPi concentrations in the cell culture medium (Fig 5A and 5B).
Finally, to determine if longer term inhibition of Fgf23 signaling in Hyp osteocyte-like cells translates into a more complete correction of TNAP activity and PPi levels, we treated osteocyte-like cells with either FGFR3 inhibitor or FGF23 AB for 96 h and subsequently assessed PPi concentration and TNAP enzyme activity, using NBT/BCIP staining for the latter. As shown in Fig 5C, both treatments did not alter ALP staining or PPi concentration in WT osteocyte-like cells. However, treatment with either FGFR3 inhibitor or FGF23 AB increased ALP staining in Hyp osteocyte-like cells compared to vehicle-treated cells. The increase in TNAP activity was accompanied by a profound increase in Pi concentration and normalization of PPi levels in cell culture medium from Hyp osteocyte-like cells treated with either FGFR3 inhibitor or FGF23 AB (Fig 5C).
Taken together, our data provide evidence that increased Fgf23-FGFR3 signaling inhibits TNAP activity in Hyp osteocytes, causing PPi accumulation which in turn contributes to the mineralization defect observed in Hyp mice. However, inhibition of Fgf23 signaling did not completely normalize TNAP mRNA expression and enzyme activity, suggesting that other, still unknown factors are involved in the regulation of TNAP in Hyp osteocytes. Of note, inhibition of Fgf23 signaling increased phosphate concentrations in the cell culture medium beyond the levels found in WT osteocyte-like cells (Fig 5C). This finding may suggest that the increased phosphate production after inhibition of Fgf23 signaling could not be adequately used for mineralization in Hyp osteocyte-like cells due to the presence of additional inhibitors of mineralization, most likely OPN and ASARM peptides due to Phex deficiency.
Hyp osteocyte-like cells are not only characterized by suppressed Tnap expression but also by increased abundance of genes associated with PPi production such as Enpp1 and Ank. To exclude the role of altered PPi production after inhibition of Fgf23 signaling in Hyp osteocyte-like cells, we treated WT osteocyte-like cells with rFGF23. We previously showed that treatment of WT osteoblasts with rFGF23 does not alter the expression of Enpp1 and Ank [33]. Treatment of WT osteoblasts with rFGF23 suppressed Tnap mRNA expression and increased PPi concentrations in the culture medium, independent of changes in expression of Enpp1 or Ank (Fig 6A).
To further validate the link between Fgf23, TNAP, and PPi in vivo, we treated WT mice with rFGF23. A 5-d treatment with rFGF23 suppressed Tnap mRNA expression (Fig 6B) and significantly increased PPi concentrations in bones of WT mice (Fig 6B). In line with our in vitro data, rFGF23 treatment did not alter mRNA levels of Ank and Enpp1 in bones. Taken together, these data corroborate the notion that extracellular FGF23 is independently associated with PPi levels in bone through its suppressive effect on TNAP.
Finally, to test whether excessive secretion of Fgf23 is responsible for the decreased Tnap expression and PPi accumulation in Hyp bones in vivo, we analyzed bones from Hyp mice in which Fgf23 was specifically deleted in cells of the osteoblastic lineage. To this end, we used a novel mouse model (Fgf23Δ/flox/Col2.3cre+) carrying a germline-deleted Fgf23 allele together with a floxed Fgf23 allele. Fgf23Δ/flox mice were mated with type 1 collagen 2.3 kb promoter-cre mice, resulting in deletion of Fgf23 in the osteoblast lineage [44]. Fgf23Δ/flox/Col2.3cre+ mice were mated with Hyp mice to obtain Hyp/Fgf23Δ/flox/Col2.3cre+ mice, a Hyp mouse model with conditional deletion of Fgf23 in osteoblasts and osteocytes. Analysis of TNAP enzyme activity in femur sections of 3-mo-old WT, Hyp, and Hyp/Fgf23Δ/flox/Col2.3cre+ mice showed that the suppression of TNAP enzyme activity in Hyp osteocytes was rescued in Hyp/Fgf23Δ/flox/Col2.3cre+ mice (Fig 7). TNAP activity was similar in osteoblasts at the bone surface in all genotypes, corroborating the notion that Fgf23 does not contribute to the regulation of TNAP activity in osteoblasts of Hyp mice (Fig 7). In a subset of these mice, we were able to quantify PPi in distal femurs. Bone PPi concentration was 1.18 ± 0.009 μmol/mg in WT (n = 4), 1.49 ± 0.007 μmol/mg in Hyp (n = 2), and 0.91 ± 0.002 μmol/mg in Hyp/Fgf23Δ/flox/Col2.3cre+ mice (n = 2). Collectively, these data suggest that increased Fgf23 secretion is indeed responsible for the suppression of TNAP expression and subsequent PPi accumulation in Hyp bones.
In the current study, we identified a novel mechanism contributing to the defective mineralization in Hyp mice. Our data indicate that besides its endocrine role as phosphaturic hormone, excessive osteocytic Fgf23 secretion has an additional para-/autocrine role in the development of osteomalacia in Hyp mice by suppressing TNAP activity in osteocytes, which in turn leads to accumulation of PPi, a potent inhibitor of mineralization. We hypothesize that the cell-specific suppression of Tnap in osteocytes but not osteoblasts of Hyp mice is based upon the profound up-regulation of Fgfr3 expression during osteocytic differentiation. This model is shown in Fig 8. Moreover, we demonstrated that conditional deletion of Fgf23 in cells of the osteoblastic lineage rescued the suppressed TNAP activity in osteocytes of Hyp mice in vivo, and that blocking of the cell-autonomous increase in Fgf23-FGFR3 signaling in Hyp-derived osteocyte-like cells improved TNAP activity and phosphate production, and decreased PPi concentration in vitro. However, inhibition of Fgf23 signaling did not fully correct the mineralization defect in vitro, suggesting that increased local Fgf23 production is only partially responsible for impaired mineralization in Hyp mice.
The XLH and Hyp phenotypes are caused by loss-of-function mutations in PHEX/Phex. It was previously thought that increased ASARM peptides were largely responsible for the cell autonomous mineralization defect observed in osteoblasts isolated from Hyp mice, and partially for the osteomalacia found in Hyp mice [18]. Phex binds to and proteolytically cleaves free ASARM peptides [45,46], and also degrades OPN [21]. ASARM peptides and OPN are increased in bones of Hyp mice [19], and known to impair mineralization in vivo and in vitro [45,47]. It is interesting to note in this context that transgenic overexpression of PHEX under different promoters only partially rescued the osteomalacia in Hyp mice (PHEX-tg/Hyp) [8,48]. The current study may provide a possible explanation why osteomalacia was not fully corrected in the majority of studies with PHEX-tg/Hyp mice. In this regard, Fgf23 levels remained significantly higher in Phex-Tg/Hyp mice [49], thus TNAP activity may have remained suppressed in these mice, causing impaired mineralization via PPi accumulation.
An interesting aspect of our study was the striking differences between the expression profiles of osteoblasts and osteocytes isolated by sequential digestion from Hyp bones. Whereas Tnap mRNA and protein expression was strongly suppressed in Hyp-derived osteocyte-like cells, Tnap mRNA expression was ~50-fold increased in Hyp relative to WT osteoblast-like cells. Vice versa, we found increased Fgf23 expression only in osteocytes but not in osteoblasts of Hyp mice as evidenced by sequential digestion of bones. This finding is in agreement with earlier reports [29,39]. However, LCM-based in situ expression profiling of osteoblasts and osteocytes revealed increased Fgf23 mRNA abundance not only in osteocytes, but also in osteoblasts in Hyp bones. Therefore, the relative contribution of osteoblasts and osteocytes to the increased circulating Fgf23 levels in Hyp mice is not entirely clear. Furthermore, in agreement with earlier studies [39], we found that Fgfr3 mRNA abundance is profoundly up-regulated during osteocytic differentiation especially in Hyp mice, supporting the notion that the higher membrane abundance of FGFR3 in Hyp osteocytes versus osteoblasts forms the basis for the cell type-specific suppression of Tnap transcription by Fgf23. Collectively, these results underscore the biological differences between osteocytes and osteoblasts in Hyp mice and suggest that the increased ALP activity in the serum of Hyp mice and XLH patients more likely reflects changes in bone surface cells rather than osteocytes. Although we did not assess TNAP expression in newly embedded osteocytes at bone-forming surfaces, we speculate that suppression of TNAP with subsequent accumulation of PPi may also occur in osteoid seams at the bone surface, not only in osteocyte lacunae. Up-regulation of FGFR3 expression in newly embedded osteocytes may lead to Fgf23-mediated suppression of TNAP, which may in turn result in accumulation of PPi and subsequent inhibition of mineralization in the widened osteoid seams of Hyp mice in addition to increased concentrations of ASARM peptides and OPN.
TNAP, an ectoenzyme, is responsible for the local production of Pi for mineralization via hydrolyzing PPi in the ECM. Tnap loss-of-function mutants are characterized by impaired bone mineralization via accumulation of PPi [50]. Furthermore, TNAP-deficient osteoblasts fail to mineralize in vitro [51], underscoring the pivotal importance of TNAP for bone mineralization. We previously reported that Fgf23-FGFR3 signaling suppresses TNAP transcription and activity, causing PPi accumulation and inhibition of mineralization in vitro [33]. Here, we showed that inhibition of Fgf23-FGFR3 signaling in Hyp osteocyte-like cells by treatment with either an FGFR3 inhibitor or an anti-FGF23 antibody improved TNAP activity and decreased PPi concentration in vitro. In addition, bone-specific deletion of Fgf23 in Hyp/Fgf23Δ/flox/Col2.3cre+ mice rescued the suppressed TNAP activity in osteocytes of Hyp mice. Although bony PPi concentrations were not quantified after treatment of Hyp mice with anti-Fgf23 antibodies [9] or a pan-FGFR inhibitor [24], it is likely that systemic anti-FGF23 treatment or pan-FGFR inhibition also, at least partially, corrects the increased PPi concentration in bone. This idea is indirectly supported by our finding that a 5-d treatment of WT mice with rFGF23 suppressed TNAP expression and increased PPi in bone, suggesting that circulating FGF23 levels are able to modulate bony PPi metabolism. The latter findings may also have implications for tumor-induced osteomalacia (TIO), because our data suggest that excessive extraskeletal production of FGF23 may also lead to PPi accumulation in bone. However, due to the low affinity of the FGFR3 signaling pathway [33], this mechanism may only become operative at high circulating FGF23 levels.
Vitamin D hormone levels are inappropriately low in Hyp mice and in XLH patients due to the FGF23-mediated suppression of renal 1α-hydroxylase [15]. The vitamin D hormone not only governs intestinal absorption of calcium and phosphate [52], but also inhibits bone mineralization by stimulating the transcription of Enpp1, Ank, and Opn [33,35]. In line with low vitamin D hormone levels in Hyp mice, the mRNA abundance of Ank and Enpp1 was almost undetectable in Hyp-derived F-3 osteoblasts. On the contrary, the mRNA abundance of Ank and Enpp1 was higher in Hyp- than WT-derived osteocyte-rich cell fractions ex vivo. Therefore, inappropriately low vitamin D hormone levels cannot account for the changes observed in Ank and Enpp1 expression in Hyp-derived osteocyte-like cells. Furthermore, inhibition of Fgf23 signaling partially corrected the increased Ank and Enpp1, but not the increased Opn mRNA expression, in Hyp-derived osteocyte-like cells in our experiments. Collectively, our data and the work of others suggest that Phex deficiency [39,53], via only partially known signaling pathways at present time, induces a complex pattern of altered gene regulation in which increased Fgf23 transcription is only a portion of the pathophysiology.
In conclusion, we have found that the mineralization defect in bones of Hyp mice and in cultures of Hyp-derived osteoblasts is not only due to local accumulation of ASARM peptides and OPN but also due to the Fgf23-driven accumulation of PPi, another potent mineralization inhibitor. Clearly, more work is required to disentangle the complex interactions between Phex deficiency, Fgf23 secretion, and para-/autocrine Fgf23 feedback signaling in osteocytes of Hyp mice. A more complete understanding of these aspects of osteocyte biology may help to design novel treatments for the mineralization defects observed in diseases associated with excessive osteocytic Fgf23 secretion such as XLH or chronic kidney disease.
All animal studies were approved by the Ethical Committee of the University of Veterinary Medicine, Vienna and by the Austrian Federal Ministry of Science and Research and were undertaken in strict accordance with prevailing guidelines for animal care and welfare (permit number BMWF-68.205/0037-II/3b/2013). Both WT controls and Hyp mice were on C57BL/6 background and were kept on normal mouse chow (Ssniff, Soest, Germany). As described [44], a conditional Fgf23 mouse model that harbored alleles with floxed exon 2 was developed through standard gene targeting. An Fgf23 null allele (Δ) created by mating to the global ella-cre transgenic line was bred onto the flox-Fgf23 background to produce Fgf23Δ/flox mice. Fgf23Δ/flox mice were crossed with type 1 collagen 2.3-kb promoter-cre mice, resulting in Fgf23Δ/flox/Col2.3cre+ mice by standard mating strategies; this line was mated onto the Hyp genetic background to obtain Hyp/Fgf23Δ/flox/Col2.3cre+ mice [44]. Genotyping of the mice was performed by multiplex PCR using genomic DNA extracted from the tail. The mice were kept at 24°C with a 12 h/12 h light/dark cycle and were allowed free access to food and tap water. All experiments were performed on 3-mo-old males. Some WT mice received daily intraperitoneal injections of vehicle (phosphate-buffered saline with 2% DMSO) or 10 μg recombinant human FGF23 R176/179Q (rFGF23, kindly provided by Amgen, Thousand Oaks, CA) per mouse for 5 days, and were killed 8–12 hours after the last injection. At necropsy, the mice were exsanguinated from the abdominal vena cava under anesthesia with ketamine/xylazine (67/7 mg/kg i.p.) for collection of serum and bones.
Serum calcium, phosphorus, and ALP activity were analyzed using a Cobas c111 analyzer (Roche). Intact Fgf23 in serum and culture medium was determined by ELISA (Immutopics).
Primary osteoblast-rich and osteocyte-rich cell fractions were isolated as previously described [54]. Briefly, both femurs were collected, carefully defleshed, the epiphysis was cut off, and bone marrow was flushed out using HBSS calcium-free and magnesium-free medium (Life Technologies). Subsequently, the washed femurs were minced into small pieces using scissors and digested with 1.25 mg/ml type II collagenase (Invitrogen) at 37°C. Cells released after the first two digestions of 15 min each were discarded. Cells released after the next three consecutive digestions of 20 min each were collected after passing through a 100-μm nylon cell strainer as Fraction 3 (F-3), Fraction 4 (F-4) and Fraction 5 (F-5), respectively. Digested bones were washed once, and treated with 4 mM EGTA in HBSS calcium-free and magnesium-free medium for 20 min at 37°C. Cells released after this treatment were collected, and bones were again digested using 1.25 mg/ml type II collagenase for 20 min at 37°C. Cells released after this digestion were collected and combined with the previous fraction and named F-6/7. Thereafter, bones were again treated with 4 mM EGTA for 20 min and subsequently with 1.25 mg/ml type II collagenase for 20 min, cells were collected as before, and named F-8/9.
LCM was performed as described previously [43]. Briefly, distal femurs of 3-mo-old WT and Hyp mice were snap-frozen in liquid nitrogen with OCT compound (Sakura Finetek, Zoeterwoude, Netherlands). Four-μm-thick cryosections of were cut on a cryotome (Leica Kryostat 1720), using the cryotape method as described [55]. Cryosections were quickly stained with HistoStain (Arcturus). Osteoblasts and osteocytes (~100–200 cells per sample each) in the cancellous bone of the distal femoral metaphysis were dissected based on their typical morphology, using a Veritas (Arcturus) LCM system.
Total RNA was isolated directly after collection of the bone cell fractions using Tri-Reagent (Ambion, Thermo Fisher Scientific) according to the manufacturer's protocol. RNA quantity was determined using a Nanodrop photometer (Thermo Scientific). For LCM-harvested osteoblasts and osteocytes, total RNA was extracted using the SPLIT RNA Extraction Kit (Lexogen), and RNA quality was determined by Agilent RNA 6000 Pico Chip (Agilent Technologies). cDNA synthesis was performed using the High capacity cDNA reverse transcription kit (Applied Biosystems). Quantitative RT-PCR was performed on a Rotor-Gene 6000 (Corbett Life Science) using 5X HOT Firepol Evagreen qPCR mix plus (Solis BioDyne). A melting curve analysis was done for all assays. Primer sequences are available on request. Efficiencies were examined based on a standard curve. Expression of target genes was normalized to the expression of the housekeeping gene glyceraldehyde-3-phosphate-dehydrogenase (Gapdh).
Isolated mouse femurs were fixed in 4% paraformaldehyde at 4°C overnight and were processed and embedded in methylmethacrylate as described previously [56]. Midsagittal sections of the distal femurs were prepared using a HM 355S microtome (Microm, Walldorf, Germany), and were stained with von Kossa/McNeal [57]. Sections were analyzed using a Zeiss Axioskop II microscope.
Calvariae were aseptically harvested from 3-d-old mice, minced and incubated with digestion medium (α-MEM medium, 2 mg/ml type II collagenase (Invitrogen) and 2% Penicillin-Streptomycin) at 37°C in a water bath for 4 h. Bone fragments were washed with PBS and cultured in α-MEM medium supplemented with 2% Penicillin-Streptomycin and 10% calf serum (PAA). A similar protocol was followed using femora of 3-d-old mice to obtain femural osteoblast cultures. After 90% confluence, cells were grown in the presence of osteoblastic differentiation medium (50 μg/ml ascorbic acid and 10 mM β-glycerophosphate) for 12–22 d as specified. The differentiated cells were treated with various concentrations of recombinant human FGF23 carrying the R176/179Q stabilizing mutations (rFGF23, kindly provided by Amgen Inc., Thousand Oaks, CA, US) for 24 h, 20 ng/ml rat anti-FGF23 antibody (kindly provided by Amgen Inc., Thousand Oaks, CA, US), or 25 nM FGFR3 inhibitor PD173074 (Sigma) for 24 or 96 h. At the various time points following treatment, cell culture supernatant and samples for RNA isolation were collected and stored at −80°C. For BCIP/NBT staining, cells were fixed using acetone and methanol (30:70) for 5 min at −20°C, and stained using TNM buffer (100 mM Tris-HCl, pH 9.5, 100 mM NaCl, 5 mM MgCl2) containing 0.175 mg/mL 3-bromo-4-chloro-3-indolyl phosphate (BCIP, Sigma) and 0.45 ng/mL nitrotetrazolium blue chloride (NBT, Sigma) for 45 min at room temperature. Stained cells were photographed using a stereomicroscope (Stemi SV6; Zeiss), and the percent area of positive staining was measured using Image J software.
For immunohistochemistry, 5-μm-thick undecalcified sections were obtained from plastic-embedded femurs as described [56]. Sections were deplastified, incubated for 15 min in 3% hydrogen peroxide in PBS to block endogenous peroxidase activity, and, after blocking with 10% rabbit serum, incubated with rabbit anti-OPN antibody (Abcam, 1:300) at 4°C overnight. After washing, sections were incubated for 2 h with biotinylated goat anti-rabbit secondary antibody (1:2,000, Vector). Finally, the sections were counterstained with Mayer's hematoxylin. Negative control was performed by omitting the primary antibody. For TNAP staining, deplastified bone sections were incubated with vector red ALP staining kit (Vector Laboratories) according to the manufacturer's protocol. Fluorescent images of TNAP and DAPI were obtained using appropriate filter sets. Fluorescence measurements were obtained using Image J software as described previously [58]. Fluorescence along the bone surface was marked manually and quantified using Image J for obtaining relative fluorescence of osteoblasts. At least 15 osteocytes per image and a total of 6 images per animal were chosen for the quantification of relative fluorescence in osteocytes. Relative fluorescence in osteocytes was normalized to cell number. The sections were analyzed using a Zeiss Axioskop 2 microscope.
Proteins from femurs were isolated using a previously described protocol [59]. Briefly, femurs were carefully defleshed and bone marrow was flushed out. After demineralization (300 μl of 1.2 M HCl at 4°C overnight), proteins were isolated using 6M guanidine-HCL in 100 mM Tris buffer, pH 7.4, at 4°C for 72 h. Extracted proteins were concentrated using ethanol precipitation and re-dissolved in 8M urea buffer. Protein concentration was determined using a BCA assay (Thermo Scientific).
Proteins were solubilized in Laemmli sample buffer, fractionated on SDS–PAGE (50 μg/well), and transferred to a nitrocellulose membrane (Thermo Scientific). Immunoblots were incubated overnight at 4°C with polyclonal rabbit anti‐OPN (1:2,000, Abcam), polyclonal goat anti-TNAP (1:2,000, R&D Systems) and monoclonal mouse anti-β-actin (1:5,000, Sigma) in 2% (w/v) bovine serum albumin (BSA, Sigma) in a TBS‐T buffer [150 mM NaCl, 10 mM Tris (pH 7.4/HCl), 0.2% (v/v) Tween-20]. After washing, membranes were incubated with horseradish peroxidase-conjugated secondary antibodies (Sigma). Specific signal was visualized by ECL kit (Amersham Life Sciences). The protein bands were quantified by Image Quant 5.0 software (Molecular Dynamics).
PPi was extracted from whole femurs with 1.2 M HCl at 4°C overnight. HCl was evaporated at 99°C, and samples were resuspended in deionized water. The amount of PPi extracted from bone or in cell culture supernatant was quantified using the PPiLight Inorganic Pyrophosphate Assay (LONZA) according to the manufacturer's protocol. Sodium PPi tetrabasic decahydrate (Sigma) was used as standard.
Statistics were computed using PASW Statistics 17.0 (SPSS Inc., Chicago, IL, US). The data were analyzed by two-sided t test (two groups) or one-way analysis of variance (ANOVA) followed by Student-Newman-Keuls multiple comparison test (>2 groups). p-Values of less than 0.05 were considered significant. Data represent mean values ± SD.
|
10.1371/journal.ppat.1004235 | cGMP and NHR Signaling Co-regulate Expression of Insulin-Like Peptides and Developmental Activation of Infective Larvae in Strongyloides stercoralis | The infectious form of the parasitic nematode Strongyloides stercoralis is a developmentally arrested third-stage larva (L3i), which is morphologically similar to the developmentally arrested dauer larva in the free-living nematode Caenorhabditis elegans. We hypothesize that the molecular pathways regulating C. elegans dauer development also control L3i arrest and activation in S. stercoralis. This study aimed to determine the factors that regulate L3i activation, with a focus on G protein-coupled receptor-mediated regulation of cyclic guanosine monophosphate (cGMP) pathway signaling, including its modulation of the insulin/IGF-1-like signaling (IIS) pathway. We found that application of the membrane-permeable cGMP analog 8-bromo-cGMP potently activated development of S. stercoralis L3i, as measured by resumption of feeding, with 85.1±2.2% of L3i feeding in 200 µM 8-bromo-cGMP in comparison to 0.6±0.3% in the buffer diluent. Utilizing RNAseq, we examined L3i stimulated with DMEM, 8-bromo-cGMP, or the DAF-12 nuclear hormone receptor (NHR) ligand Δ7-dafachronic acid (DA)—a signaling pathway downstream of IIS in C. elegans. L3i stimulated with 8-bromo-cGMP up-regulated transcripts of the putative agonistic insulin-like peptide (ILP) -encoding genes Ss-ilp-1 (20-fold) and Ss-ilp-6 (11-fold) in comparison to controls without stimulation. Surprisingly, we found that Δ7-DA similarly modulated transcript levels of ILP-encoding genes. Using the phosphatidylinositol-4,5-bisphosphate 3-kinase inhibitor LY294002, we demonstrated that 400 nM Δ7-DA-mediated activation (93.3±1.1% L3i feeding) can be blocked using this IIS inhibitor at 100 µM (7.6±1.6% L3i feeding). To determine the tissues where promoters of ILP-encoding genes are active, we expressed promoter::egfp reporter constructs in transgenic S. stercoralis post-free-living larvae. Ss-ilp-1 and Ss-ilp-6 promoters are active in the hypodermis and neurons and the Ss-ilp-7 promoter is active in the intestine and a pair of head neurons. Together, these data provide evidence that cGMP and DAF-12 NHR signaling converge on IIS to regulate S. stercoralis L3i activation.
| Human parasitic nematodes, including Strongyloides stercoralis, cause extensive morbidity in the developing world. The infectious form of S. stercoralis is a developmentally arrested third-stage larva (L3i), which resumes development into a parasitic adult upon entering a host. The molecular mechanisms controlling the developmental arrest and activation of L3i are not well understood. The free-living nematode Caenorhabditis elegans has a morphologically similar developmentally arrested third-stage dauer larva, which is regulated by four canonical dauer signaling pathways. Using C. elegans as a model, we hypothesized that cyclic guanosine monophosphate (cGMP) signaling would be important for L3i activation and would also regulate downstream insulin/IGF-1-like signaling (IIS). Indeed, we found that the membrane-permeable cGMP analog 8-bromo-cGMP stimulated L3i activation, accompanied by an increase in transcripts of putative agonistic insulin-like peptides (ILPs), which encode the ligands for IIS. Using the C. elegans model, we also hypothesized that DAF-12 nuclear hormone receptor (NHR) signaling would be downstream of IIS during L3i activation. Surprisingly, we found that during L3i activation, parallel cGMP and DAF-12 NHR signaling pathways co-regulate the downstream IIS pathway via modulation of ILPs. Together, these data help to further elucidate the pathways governing S. stercoralis L3i development.
| Parasitic nematodes infect approximately one in four persons globally, with the vast burden of disease concentrated in tropical and developing regions [1]. The parasitic nematode Strongyloides stercoralis infects an estimated 30–100 million people worldwide [2]; in corticosteroid-treated or human T-cell lymphotropic virus 1 (HTLV-1) infected persons, infection with S. stercoralis can result in hyperinfection and potentially fatal disseminated strongyloidiasis [3]. Like many soil-transmitted helminths, the infectious form of S. stercoralis is a developmentally arrested third-stage larva (L3i), which is non-feeding, long-lived, and stress-resistant [4]. S. stercoralis L3i exhibit thermotaxis and chemotaxis in response to a range of host-like cues [5], including host body temperature [6], carbon dioxide [7], sodium chloride [8], and urocanic acid [9]. Upon entering a suitable host, L3i quickly activate and resume feeding and development [4]. However, the molecular mechanisms by which S. stercoralis L3i sense and transduce host cues and subsequently initiate resumption of development are poorly understood.
The free-living nematode Caenorhabditis elegans has a facultative developmentally arrested third-stage larva, known as the dauer larva, which forms during stressful conditions including high temperature, low food abundance, and high dauer pheromone levels; when conditions improve, C. elegans exits dauer and resumes reproductive development [10], [11]. The molecular pathways regulating dauer entry have been well studied and include: a cyclic guanosine monophosphate (cGMP) signaling pathway, an insulin/IGF-1-like signaling (IIS) pathway, a dauer transforming growth factor β (TGFβ) signaling pathway, and a DAF-12 nuclear hormone receptor (NHR) regulated by dafachronic acid (DA) steroid ligands (Figure 1) [12], [13]. We have demonstrated that components of these four pathways are present in S. stercoralis [14], that elements of the IIS pathway control L3i arrest and activation [15], [16], and that Δ7-DA is a potent activator of L3i [17]. However, a role for cGMP signaling in regulating S. stercoralis L3i development has not been previously examined, nor have the epistatic relationships of these pathways been explored.
In C. elegans, cGMP is a second messenger for many chemosensory seven-transmembrane G protein-coupled receptors (7TM GPCRs), which act as sensors for a wide variety of environmental stimuli and regulate the developmental switch controlling dauer versus reproductive development. Indeed, the C. elegans genome encodes over 1,300 chemosensory 7TM GPCRs [18], [19], which signal through the 21 G protein α (Gα) subunits, two G protein β (Gβ) subunits, and two G protein γ (Gγ) subunits encoded in the genome [20]. One of the primary functions of these G proteins is to regulate guanylyl cyclases, including Ce-DAF-11, which in turn produce cGMP [21]. Downstream effects of cGMP signaling are mediated, in part, by heteromeric cyclic-nucleotide gated ion channels, including the one formed by Ce-TAX-4 and Ce-TAX-2, which regulates chemosensation, thermosensation, and dauer development [22]–[24].
Perhaps one of the best studied examples of chemosensory 7TM GPCRs regulating development through cGMP signaling is the sensing of dauer pheromone, which is a measure of population density in C. elegans [25]. Dauer pheromone, a complex mixture of ascarosides [26]–[29], is continuously secreted by C. elegans, and high concentrations potently induce dauer formation [11], [30]. Several chemosensory 7TM GPCRs sense specific or combinations of dauer-inducing ascarosides; these include Ce-SRBC-64 and Ce-SRBC-66 [31], Ce-SRG-36 and Ce-SRG-37 [32], and Ce-DAF-37 and Ce-DAF-38 [33]. At least two of these chemosensory 7TM GPCRs signal through the Gα subunits Ce-GPA-2 and Ce-GPA-3 [31]. When dauer pheromone levels are high, Ce-GPA-2 and Ce-GPA-3 inhibit the transmembrane guanylyl cyclase encoded by Ce-daf-11, thereby decreasing concentrations of the second messenger cGMP [31]. Constitutively-activated forms of Ce-GPA-2 or Ce-GPA-3, as well as inactivating mutations in Ce-daf-11, result in dauer constitutive (Daf-c) phenotypes, while mutations that inactivate Ce-GPA-2 or Ce-GPA-3 result in decreased dauer entry under dauer-inducing conditions [34]–[37].
Exogenous application of the membrane-permeable cGMP analog 8-bromo-cGMP rescues the Daf-c phenotype of Ce-daf-11 mutants, but does not rescue the Daf-c phenotypes of Ce-tax-4, IIS pathway, or dauer TGFβ pathway mutants [36]. Furthermore, addition of 8-bromo-cGMP increases abundance of Ce-ins-7 and Ce-daf-28 transcripts [38], which both encode agonistic insulin-like peptides (ILPs) in the IIS pathway [39], [40]. Signaling mediated by cGMP also regulates the dauer TGFβ pathway, since Ce-daf-11 acts cell autonomously to regulate Ce-daf-7 expression in ASI chemosensory neurons [41], [42]. These data suggest that cGMP signaling acts upstream of both IIS and TGFβ signaling in regulating C. elegans dauer development.
Genetic epistatic analysis of the four pathways regulating dauer development has placed Ce-DAF-12 NHR signaling downstream of cGMP, IIS, and dauer TGFβ signaling [12], [34], [43]–[46]. Ce-DAF-12 is an NHR regulated by DA steroid ligands, the presence of which promotes reproductive growth and development [47], [48]. Exogenous application of Δ7-DA can rescue the Daf-c phenotype of both IIS and dauer TGFβ pathway mutants [47]. Additionally, both IIS and dauer TGFβ pathways regulate Ce-DAF-9, which encodes a cytochrome P450 that catalyses the rate-limiting final step in DA biosynthesis [49], [50]. Together, these data suggest that Ce-DAF-12 NHR signaling is the downstream effector for C. elegans dauer development.
While the pathways regulating C. elegans dauer development are well-studied, the mechanisms regulating L3i developmental arrest and activation in parasitic nematodes are comparatively not well understood. In particular, the roles of cGMP signaling and the epistatic relationships of canonical dauer pathways in regulating S. stercoralis L3i arrest and activation have not been examined. Previous work has demonstrated that increased cGMP signaling activates L3i in Ancylostoma and Nippostrongylus hookworms (clade 9B) [51]–[54], which are closely related to C. elegans (clade 9A) [54]; however, this has not been demonstrated in S. stercoralis (clade 10B) [54], where parasitism is thought to have arisen independently of the hookworms [55]. Additionally, next-generation deep-sequencing of the transcriptome (RNAseq) of S. stercoralis revealed increased transcript levels of many cGMP signaling pathway components in L3i [14]. Together, these data led us to hypothesize that chemosensory 7TM GPCRs sense host cues and that cGMP pathway signaling is important for transducing host-like cues in S. stercoralis L3i as well as triggering resumption of development once inside the host.
Although we and others have investigated the role of canonical dauer pathways in regulating L3i arrest and activation [14], [56], we are not aware of any studies examining the epistatic relationship of these pathways in parasitic nematodes. While studies examining the role of dauer TGFβ signaling in parasitic nematodes have largely concluded that this pathway regulates L3i development differently from Ce-DAF-7 [57]–[62], studies examining IIS and DAF-12 NHR signaling suggest these pathways function similarly in regulating both L3i and dauer development [15], [17], [63]–[65]. Therefore, we also hypothesized that the epistatic ordering of cGMP, IIS, and DAF-12 NHR signaling pathways regulating C. elegans dauer development would be retained in S. stercoralis during L3i activation.
In this study, we sought to determine the roles of and the relationships between canonical dauer pathways in the regulation of S. stercoralis L3i activation. We found that during L3i activation, parallel cGMP and DAF-12 NHR signaling pathways co-regulate the downstream IIS pathway via modulation of ILPs. Understanding the mechanisms of L3i activation may lead to new or improved therapies for parasitic nematode infections.
The S. stercoralis PV001 and UPD strains were maintained in prednisone-treated beagles in accordance with protocols 702342, 801905, and 802593 approved by the University of Pennsylvania Institutional Animal Care and Use Committee (IACUC). All IACUC protocols, as well as routine husbandry care of the animals, were carried out in strict accordance with the Guide for the Care and Use of Laboratory Animals of the National Institutes of Health.
The S. stercoralis PV001 and UPD strains were maintained and cultured as previously described [16], [66], [67].
In vitro activation of S. stercoralis L3i was performed as previously described [16], [68] with the following adaptations. All conditions were supplemented with antibiotics (100 U/ml penicillin, 10 µg/ml streptomycin, and 12.5 µg/ml tetracycline). M9 Buffer was used as the medium for both the experimental conditions and the negative control [69], except where indicated.
For the titration of 8-bromo-cGMP, L3i were isolated from six-day-old charcoal coprocultures (incubated at 25°C) by the Baermann technique at 27°C. L3i were subsequently washed twice in deionized water and incubated in M9 buffer supplemented with antibiotics for three hours at room temperature before distribution amongst the different conditions. The positive control was composed of Dulbecco's Modification of Eagle's Medium (DMEM) (supplemented with L-glutamine, 4.5 g/L glucose, and sodium pyruvate) (Corning, www.cellgro.com), 10% naive canine serum, and 12.5 mM L-glutathione reduced (CAS 70-18-8) (Sigma-Aldrich, www.sigmaaldrich.com). The negative control was M9 buffer. The experimental conditions, using 8-bromo-cGMP (CAS 51116-01-9) (Sigma) at 800 µM, 200 µM, 100 µM, and 50 µM, were carried out in M9 buffer. Each condition consisted of three wells in a 96-well plate, with approximately 100 L3i in 100 µl total volume in each well. L3i were incubated at 37°C in 5% CO2 in air for 21 hours; 2.5 µl of fluorescein isothiocyanate (FITC; CAS 3326-32-7) (Sigma) dissolved in N,N-dimethylformamide (CAS 68-12-2) (Sigma) at 20 mg/ml and incubated for ≥ one month was then added to each well, and the cultures were incubated an additional three hours at 37°C and 5% CO2 in air (24 hours total). L3i for each condition were pooled and washed five times in 14 ml of M9 buffer, with centrifugation at 75× g for five minutes at room temperature. L3i were then mounted on glass slides with grease-edged cover-slips, immobilized by a 20-second heat-shock at 60°C or with 10 mM levamisole (CAS 16595-80-5) (Sigma), and viewed by fluorescence microscopy. Only L3i with FITC in the pharynx were scored as “positive” for feeding. Three biological replicates were performed, and the mean percentage of L3i feeding in each condition with the standard deviation was plotted in Prism version 5.03 (GraphPad Software, Inc., http://www.graphpad.com/).
For L3i activation kinetics, conditions included both 200 µM 8-bromo-cGMP in M9 buffer and host-like cues, consisting of DMEM, 10% canine serum, and 3.75 mM L-glutathione reduced. L3i were isolated as previously described. L3i in all conditions were incubated at 37°C and 5% CO2 in air for a total of 24 hours, with chemical cues added at appropriate intervals such that L3i were exposed to activating compounds for 4, 6, 12, 18, or 24 hours total. L3i were incubated with FITC and scored for feeding as previously described. Three biological replicates were performed, and the mean percentage of L3i feeding in each condition with the standard deviation was plotted in Prism.
For titration of Δ7-DA, conditions included Δ7-DA at 800 nM, 400 nM, 200 nM, 100 nM, and 50 nM in M9 buffer as well as an M9 buffer negative control. Additionally, 100 µM LY294002, a phosphatidylinositol-4,5-bisphosphate 3-kinase (PI3 kinase) inhibitor previously demonstrated to inhibit L3i activation [16], in 1.3% dimethyl sulfoxide (DMSO) was mixed with 400 nM Δ7-DA in M9 buffer; the negative control was 1.3% DMSO in M9 buffer. L3i were isolated as previously described. L3i in all conditions were incubated at 37°C and 5% CO2 in air for a total of 24 hours. L3i were incubated with FITC and scored for feeding as previously described. Four biological replicates were performed, and the mean percentage of L3i feeding in each condition with the standard deviation was plotted in Prism.
For RNAseq analysis, PV001 strain L3i were isolated from seven-day-old charcoal coprocultures (incubated at 25°C) by the Baermann technique at 27°C. L3i were subsequently washed twice in deionized water and incubated in M9 buffer supplemented with antibiotics for three hours at room temperature. An aliquot of these L3i (no stimulation control) was pelleted and frozen in 100 µl TRIzol reagent (Life Technologies, www.lifetechnologies.com). L3i were then distributed amongst the following conditions, each supplemented with antibiotics: M9 buffer, DMEM (supplemented with L-glutamine, 4.5 g/L glucose, and sodium pyruvate), 200 µM 8-bromo-cGMP in M9 buffer, and 400 nM Δ7-DA in M9 buffer. We found that removal of the reduced glutathione and naive canine serum components from the DMEM-based biochemical mixture had no effect on L3i activation; thus, fresh DMEM without these additives was used. L3i were incubated at 37°C and 5% CO2 in air for a total of 24 hours using 500 L3i in 100 µl of liquid in each well of a 96-well round-bottom plate with 24 wells per condition. One well from each condition was used to assess the percentage of L3i ingesting the FITC dye, as previously described. L3i from the remaining 23 wells were pooled and the pellet frozen at −80°C in 100 µl TRIzol reagent. This protocol was repeated for a total of four biological replicates. The mean percentage of L3i feeding in each condition with the standard deviation was plotted in Prism.
S. stercoralis L3i were activated for RNAseq analysis as described, with a total of four biological replicates for each of the five conditions. The L3i pellet in TRIzol was thawed and ground using a pestle, and total RNA was extracted per the manufacturer's protocol. RNA concentrations were determined using a NanoDrop 2000 spectrophotometer (Thermo Scientific, www.nanodrop.com). Total RNA was additionally quantified using the Bioanalyzer 2100 (Agilent Technologies, Inc., http://www.agilent.com), and only samples with an RNA integrity number (RIN) greater than 8.0 were used.
Libraries were constructed using the TruSeq RNA Sample Preparation Kit v2 (Illumina, Inc., http://www.illumina.com) according to the manufacturer's protocol. For each library, 500 ng of total RNA, diluted to 10 ng/µl in de-ionized water, was used as starting material. Polyadenylated RNA enrichment was performed first using olido-dT beads and eluted polyadenylated RNA fragmented at 94°C for eight minutes to approximately 200±65 (standard deviation) bp. Subsequently, first and second strand cDNA was synthesized; unique adapters for each replicate were then ligated. dsDNA fragments with ligated adapters were enriched using 14 cycles of PCR. Libraries were assessed for fragment size distribution using the Bioanalyzer 2100.
The concentration of the dsDNA adapter-ligated libraries was then determined by quantitative PCR (qPCR) with the Kapa SYBR Fast qPCR Kit for Library Quantification (Kapa Biosystems, Inc., http://www.kapabiosystems.com) using the manufacturer's protocol. Three dilutions, at 1∶4,000, 1∶8,000, and 1∶16,000, were used to calculate the concentration of each of the libraries using a calibration curve of Kapa standards. Each library was then diluted to 10 nM and pooled in equal volume quantities.
Pools were sequenced on the Illumina HiSeq 2000 with 100 bp paired-end reads, with image analysis and base calling performed with HiSeq Control Software. Raw flow-cell data was processed and demultiplexed using CASAVA (Illumina) for each of the 21 samples. Raw RNAseq reads are available in the ArrayExpress database (www.ebi.ac.uk/arrayexpress) under accession number E-MTAB-2192.
Raw reads from each L3i activation sample were independently aligned to S. stercoralis genomic contigs (6 December 2011 draft; ftp://ftp.sanger.ac.uk/pub/pathogens/HGI/) using TopHat2 version 2.0.9 (http://tophat.cbcb.umd.edu/) [70], which utilized the Bowtie2 aligner version 2.1.0 (http://bowtie-bio.sourceforge.net/bowtie2/index.shtml) [71] and SAMtools version 0.1.19 (http://samtools.sourceforge.net/). Default parameters were used, but with the following options: mate inner distance of 25; mate standard deviation of 50; minimum anchor length of 6; minimum intron length of 30; maximum intron length of 20,000; micro exon search; minimum segment intron of 30; and maximum segment intron of 20,000. Aligned reads from each developmental stage were inspected using the Integrated Genome Viewer (IGV) version 2.3.20 (http://www.broadinstitute.org/igv/).
Additional S. stercoralis developmental stages used for RNA isolation, dsDNA library construction and sequencing (ArrayExpress accession number E-MTAB-1164; http://www.ebi.ac.uk/arrayexpress/experiments/E-MTAB-1164), and read alignment to the S. stercoralis draft genome (6 December 2011 version) have been described previously [14]. De novo assembly of RNAseq reads from S. stercoralis developmental stages has also been described previously [14]. These seven developmental stages include: free-living females (FL Female), post-free-living first-stage larvae (PFL L1), infectious third-stage larvae (L3i), in vivo activated third-stage larvae (L3+), parasitic females (P Female), post-parasitic first-stage larvae (PP L1), and post-parasitic third-stage larvae (PP L3).
Transcript abundances of manually annotated S. stercoralis genes were calculated using Cufflinks version 2.0.2 (http://cufflinks.cbcb.umd.edu/) as fragments per kilobase of coding exon per million fragments mapped (FPKM), with paired-end reads counted as single sampling events [72]. FPKM values for coding sequences (CDS) (Data S1, S2, S3), along with ±95% confidence intervals, were calculated for each gene using Cuffdiff version 2.0.2 (Data S4). FPKMs and 95% confidence intervals were plotted in Prism. Significant differences in FPKM values between developmental stages and p-values were determined using Cuffdiff version 2.0.2, a program with the Cufflinks package [73], [74]; p-values less than 0.05 were considered statistically significant.
BLAST searches of the S. stercoralis (ftp://ftp.sanger.ac.uk/pub/pathogens/HGI/) genomic contigs, as well as S. stercoralis de novo assembled transcripts (ArrayExpress accession number E-MTAB-1184; http://www.ebi.ac.uk/arrayexpress/experiments/E-MTAB-1184), using C. elegans protein sequences (Data S5) were performed using Geneious version 6.0 (www.geneious.com) set to the least restrictive parameters. A total of 76 C. elegans chemosensory 7TM GPCRs, four from each of the 19 families (Data S5), were used to BLAST search the S. stercoralis genomic contigs, resulting in a total of 227 hits. The 21 C. elegans Gα proteins, the two Gβ proteins, and the two Gγ proteins (Data S5), were used to search both the S. stercoralis genomic contigs and de novo assembled transcripts.
BLAST hits were manually annotated using aligned reads from all seven developmental stages by a combination of IGV and Geneious. Putative S. stercoralis homologs were identified through reverse BLAST searches using NCBI's pBLAST (http://blast.ncbi.nlm.nih.gov/Blast.cgi) [75] against C. elegans sequences. Putative S. stercoralis homologs of chemosensory 7TM GPCRs were also checked for transmembrane domains using TMHMM Server version 2.0 (http://www.cbs.dtu.dk/services/TMHMM/) and assigned to a C. elegans superfamily based on reverse BLAST search results. Manually annotated S. stercoralis transcripts were used to determine predicted protein sequences.
Phylogenetic analysis of C. briggsae, C. elegans, and S. stercoralis Gα proteins was performed by alignment of protein sequences with Clustal W and a BLOSUM matrix using Geneious (Data S6). A neighbor-joining phylogenetic tree, with 1000 iterations of bootstrapping, was then constructed using Geneious. S. stercoralis Gα protein-encoding genes were named by relationship to C. briggsae and C. elegans proteins (Figure S1).
Plasmids for microinjection of S. stercoralis were constructed by Gateway cloning technology (Life Technologies). In general, DNA sequences between the start codons of genes of interest and the stop codons of the genes immediately upstream were used as promoters in Ss-ilp reporter transgene constructs. The Ss-ilp-1 promoter region, containing 2,327 bp 5′ of the start site, was PCR amplified from S. stercoralis genomic DNA, using the primers Ssilp1-1F and Ssilp1-1R, and recombined into pDONR P4-P1R, forming pPV483. The Ss-ilp-6 promoter region, containing 2,566 bp 5′ of the start site, was PCR amplified from S. stercoralis genomic DNA, using the primers Ssilp6-1F and Ssilp6-1R, and recombined into pDONR P4-P1R, forming pPV484. The Ss-ilp-7 promoter region, containing 1,270 bp 5′ of the start site, was PCR amplified from S. stercoralis genomic DNA, using the primers Ssilp7-1F and Ssilp7-1R, and recombined into pDONR P4-P1R, forming pPV485. The 870 bp coding sequence and stop codon for enhanced green fluorescent protein (egfp) was PCR amplified from pJA257 (Addgene, www.addgene.org), using the primers EGFPGW-1F and EGFPGW-1R, and recombined into pDONR 221, forming pPV477. The Ss-era-1 terminator, consisting of 598 bp 3′ of Ss-era-1 [76], was PCR amplified from pAJ08 (Addgene), using the primers Ss-era-1-1FattB2r and Ss-era-1-1RattB3, and recombined into pDONR P2R-P3, forming pPV475. Primer sequences are listed in Data S7. S. stercoralis genomic DNA was prepared from mixed-stage worms using the Qiagen DNeasy Blood and Tissue kit (www.qiagen.com). All pDONR plasmid inserts were confirmed by complete sequencing.
S. stercoralis ILP promoter::egfp reporter plasmids were constructed by LR recombination reactions using Gateway. The Ss-ilp-1 promoter::egfp::Ss-era-1 terminator plasmid was constructed by recombining the plasmids pPV483, pPV477, and pPV475 into pDEST R4-R3, forming pPV487. The Ss-ilp-6 promoter::egfp::Ss-era-1 terminator plasmid was constructed by recombining the plasmids pPV484, pPV477, and pPV475 into pDEST R4-R3, forming pPV488. The Ss-ilp-7 promoter::egfp::Ss-era-1 terminator plasmid was constructed by recombining the plasmids pPV485, pPV477, and pPV475 into pDEST R4-R3, forming pPV489. All pDEST plasmid inserts were confirmed by complete sequencing.
S. stercoralis was transformed by gonadal micro-injection of adult free-living females as previously described [66]. A mix of 50 ng/µl of either pPV487, pPV488, or pPV489 and 20 ng/µl of pAJ08 (Addgene) as a co-injection marker was micro-injected into the distal gonad of gravid females. Injected females were paired with an equal number of males and incubated on an NGM agar plate, with E. coli OP50 as a food source, at 22°C. The F1 post-free-living progeny were screened for fluorescence both 48 and 72 hours after microinjection. Larvae were screened for expression of fluorescent reporter transgenes using an Olympus SZX12 stereomicroscope with coaxial epifluorescence (www.olympus.com). Each transgenic larva was subsequently mounted on a 2% agarose pad (Lonza, www.lonza.com), anesthetized with 10 mM levamisole (Sigma), and examined in detail using an Olympus BX60 compound microscope equipped with Nomarski Differential Interference Contrast (DIC) optics and epifluorescence. Specimens were imaged using a Spot RT Color digital camera and Spot Advanced v5.1 image analysis software (Diagnostic Instruments, Inc., www.spotimaging.com). Captured images were processed using GIMP version 2.6 (www.gimp.org) and Microsoft PowerPoint 2007 (www.microsoft.com).
We previously discovered that in S. stercoralis, transcripts of genes encoding putative cGMP signaling proteins are up-regulated in L3i, and so we speculated that cGMP signaling may be important in L3i for relaying host cues and controlling resumption of development upon entering a permissive host [14]. Since chemosensory 7TM GPCRs are known to regulate cGMP signaling and to be crucial for C. elegans' response to environmental cues [18], [19], [77], including the sensing of ascarosides [31]–[33], we hypothesized that homologs in S. stercoralis might play a role in sensing environmental and host cues, especially in L3i. Thus, we surveyed homologs of chemosensory 7TM GPCRs in S. stercoralis to determine whether the transcripts are developmentally regulated in a manner consistent with a role in sensing host cues.
Using reciprocal BLAST searches, with four disparate members from each of the 19 C. elegans chemosensory 7TM GPCR families used to conduct the initial search [19], we identified a total of 85 genes in the S. stercoralis genome that encode putative chemosensory 7TM GPCRs (Table 1). These 85 putative S. stercoralis chemosensory 7TM GPCRs almost certainly compose an incomplete list of the total number of S. stercoralis chemosensory 7TM GPCRs; however, this list likely includes the majority of these receptors encoded in the genome, given the open parameters of the BLAST searches. The chemosensory 7TM GPCRs from S. stercoralis were assigned by sequence homology to the SRA, SRG, SRSX, and STR superfamilies of 7TM GPCRs (Table 1). Although we included four members from each of the C. elegans SRW, SRZ, and SRBC superfamilies in our BLAST searches, we did not find homologs from any of these superfamilies in S. stercoralis. This search also identified other conserved classes of 7TM receptors (data not shown). We determined which of the seven C. elegans superfamilies each of the S. stercoralis chemosensory 7TM GPCRs were homologous to using BLAST scores (Table 1). However, we were unable to assign S. stercoralis homologs to specific C. elegans families due to large predicted protein sequence differences.
Utilizing RNAseq data from seven different S. stercoralis developmental stages [14], we determined the transcript abundance patterns for each of the 85 putative S. stercoralis chemosensory 7TM GPCR homologs (Table 1). Surprisingly, nearly all of the transcript abundance profiles fit into one of only four patterns: transcripts detected only in L3i, only in in vivo activated L3+, in both L3i and L3+, or in none of the developmental stages examined. Furthermore, the normalized transcript abundance, calculated as FPKM, for nearly all of these transcripts was lower than for many other genes we have examined in S. stercoralis.
In C. elegans, as well as other metazoans, chemosensory 7TM GPCRs signal through heterotrimeric G proteins to intracellular effectors [20], [78]. G proteins are composed of three separately transcribed peptides: the Gα subunit that interacts directly with the 7TM GPCR and confers functional specificity, the Gβ subunit, and the Gγ subunit [21]. The C. elegans genome contains 21 Gα subunit-, two Gβ subunit-, and two Gγ subunit-encoding genes; the promoters for the majority of the Gα subunit-encoding genes are active in chemosensory amphidial neurons [20]. Two of the C. elegans Gα subunit-encoding genes, Ce-gpa-2 and Ce-gpa-3, play a role in larval commitment to dauer development [31], [37]. Our lab has previously identified orthologs of these two genes in S. stercoralis, named Ss-gpa-2 and Ss-gpa-3, the transcripts of which are at a maximum in L3i [14], [79]. Additionally, the promoter for Ss-gpa-3 is active in the amphidial neurons of transgenic S. stercoralis post-free-living larvae, suggesting that it plays a role in relaying chemosensory cues [80]. Thus, we sought to identify all the G proteins in S. stercoralis and examine their transcript abundance patterns to determine whether the transcripts of other G proteins in the parasite are also at a maximum in L3i.
Using reciprocal BLAST searches, we identified a total of 14 Gα subunit-, two Gβ subunit-, and two Gγ subunit-encoding genes in S. stercoralis (Table 2). By comparing the putative protein sequences for the S. stercoralis Gα subunits with C. elegans and C. briggsae sequences in a phylogenetic analysis, we were able to determine the Gα gene class [78] for each of the S. stercoralis Gα subunits as well as their orthologous relationships with their Caenorhabditis spp. counterparts (Table 2, Figure S1). Notably absent from the S. stercoralis genomic contigs, as well as the de novo assembled transcripts, were orthologs for gpa-1, -8, -9, -11, -14, -15, and -16.
We then examined the transcript abundance patterns for each of the G protein-encoding genes, using RNAseq data from seven S. stercoralis developmental stages [14]. We found that for many of the nematode-specific Gα subunit orthologs [78], transcript abundance reached a peak in L3i (Table 2, Figure S2). In contrast, transcripts for Ss-gpa-4 were detected in all developmental stages examined, except L3i. Transcripts from the highly conserved Gβ subunit and Gγ subunit genes were found in all developmental stages examined (Table 2, Figure S3).
In previous studies, we observed an increase in transcripts encoding guanylyl cyclases in S. stercoralis L3i [14], suggesting that when L3i encounter a permissive host, one of the downstream effects of chemosensory 7TM GPCR signaling through heterotrimeric G proteins is an increase in the second-messenger cGMP. A similar pathway—and accompanying increase in cGMP—has been described in C. elegans in response to odorants [81]. Additionally, other research groups have used the membrane-permeable analog of cGMP, 8-bromo-cGMP, to test whether increases in cGMP can activate L3i in place of host-like cues in Ancylostoma caninum [52], Ancylostoma ceylanicum [51], and Nippostrongylus brasiliensis [53]. In these three hookworm species (clade 9B) [54], which are closely related to C. elegans (clade 9A), 8-bromo-cGMP activates L3i at 5 mM for A. caninum and A. ceylanicum and at 500 µM for N. brasiliensis.
Parasitism in S. stercoralis (clade 10B) is thought to have arisen independently of the hookworm species [55]. For this reason, we asked whether increases in cGMP can also activate L3i in place of host-like cues in this parasite. To this end, we applied 8-bromo-cGMP to S. stercoralis L3i and assessed activation in an in vitro feeding assay [16], [68]. We incubated S. stercoralis L3i in a range of 8-bromo-cGMP concentrations in M9 buffer, M9 buffer as a negative control, and a mixture of biochemical host-like cues, consisting of DMEM supplemented with 10% canine serum and 12.5 mM reduced glutathione, as a positive control, for a total of 24 hours at 37°C and 5% CO2 in air; we then assessed resumption of feeding, a hallmark of activation, by ingestion of a FITC fluorescent dye.
S. stercoralis L3i were activated by 8-bromo-cGMP, with a concentration of 200 µM resulting in 85.1% (±2.2%, SD) of L3i feeding compared to 0.6% (±0.3%, SD) for the M9 buffer negative control (Figure 2A). At much higher concentrations of 8-bromo-cGMP (i.e. 2 mM or greater), we observed L3i that were radially constricted in alternating segments along the longitudinal axis and that had a compromised cuticle, as evidenced by permeability to the FITC dye (data not shown).
To compare the kinetics of activation of S. stercoralis L3i by host-like cues to L3i activated by 8-bromo-cGMP, we examined the frequency of feeding over a 24-hour time course. We determined the percentage of L3i feeding after incubation in 200 µM 8-bromo-cGMP or a mixture of DMEM, 10% canine serum, and 3.75 mM reduced glutathione, for 4, 6, 12, 18, and 24 hours at 37°C and 5% CO2 in air. We found that 8-bromo-cGMP activated L3i more rapidly than the mixture of biochemical host-like cues (Figure 2B).
To investigate changes in the abundance of transcripts during L3i activation, we utilized RNAseq to examine the differences in parasites exposed to five different conditions. To control for the influence of host-like temperature on L3i activation, we used L3i exposed to neither thermal nor biochemical host-like cues (no stimulation control) and L3i exposed to thermal cues only (M9 buffer control). These two control conditions were compared to L3i stimulated with both thermal and chemical host-like cues. Titration of Δ7-DA in M9 buffer demonstrated that frequency of L3i feeding was maximal at 400 nM (93.3±1.1%, SD; Figure 3); at concentrations of Δ7-DA 2 µM or greater, we observed a large percentage of L3i that were immobilized, curled, or dead (as evidenced by cuticle permeability to the FITC dye). Thus, we used the following conditions, in addition to thermal cues, to stimulate L3i: DMEM alone, 200 µM 8-bromo-cGMP, and 400 nM Δ7-DA. We hypothesized that these three activating conditions (DMEM, 8-bromo-cGMP, and Δ7-DA) would exhibit different activation profiles, since they target different parts of the L3i activation pathway.
Using four biological replicates, we assessed the percentage of L3i feeding, as measured by ingestion of FITC dye, in each of the conditions incubated for 24 hours (Figure 4A). RNA was extracted from each of the five conditions for each of the four biological replicates, and RNAseq libraries were constructed. We then determined the normalized transcript abundance (FPKM) for each gene of interest.
Studies in C. elegans have revealed that cGMP pathway signaling lies upstream of IIS and regulates transcript levels of Ce-daf-28 and Ce-ins-7, both proposed to be agonists of the Ce-DAF-2 insulin-like receptor [38], [39]. Previously, we identified seven ILPs in S. stercoralis and noted that the transcripts for several of these are developmentally regulated [14]. In A. caninum, 8-bromo-cGMP activation has been correlated with the transcriptional profile observed with serum stimulation; however, no canonical dauer pathway component transcripts were examined in this study [82]. To determine whether cGMP signaling also regulates transcripts encoding ILPs during activation of S. stercoralis L3i, we utilized RNAseq to examine changes in transcript levels for Ss-ilp-1 through -7 following L3i activation with 8-bromo-cGMP (Figure 4).
Several of the S. stercoralis ILP-encoding transcripts were significantly regulated during L3i activation (Figure 4). Based on their predicted protein sequences and transcript abundance patterns in different developmental stages, we previously hypothesized that both Ss-ilp-1 and Ss-ilp-6 encode agonistic ILPs while Ss-ilp-7 encodes an antagonistic ILP for the DAF-2 insulin-like receptor [14]. Compared to the no stimulation control, Ss-ilp-1 transcripts were increased over 20-fold in L3i stimulated with 8-bromo-cGMP (p<0.01). Additionally, Ss-ilp-3 and Ss-ilp-6 transcripts increased significantly (8-fold and 11-fold, respectively) in L3i stimulated with 8-bromo-cGMP compared to the no stimulation control (p<0.01). By contrast, levels of Ss-ilp-7 transcripts following 8-bromo-cGMP activation were unchanged relative to the no stimulation or M9 buffer controls. Interestingly, DMEM-mediated activation of L3i resulted in modulation of ILP-encoding transcripts similar to activation with 8-bromo-cGMP.
Studies in C. elegans have shown that cGMP pathway signaling also regulates the dauer TGFβ pathway, including transcript levels of the single dauer TGFβ ligand-encoding gene Ce-daf-7 [41]. We previously described seven daf-7-like genes in S. stercoralis, named Ss-tgh-1 through -7, and noted that Ss-tgh-1, -2, and -3 transcripts were only detected in L3i [14], [61]. We examined the changes in transcript abundance for Ss-tgh-1 through -7 upon L3i stimulation with 8-bromo-cGMP (Figure S4). We observed significant decreases in transcript abundance for Ss-tgh-1, -2, and -3 in 8-bromo-cGMP-treated worms in comparison to the M9 buffer control (p<0.01). Interestingly, Ss-tgh-6 was up-regulated ≥14-fold in 8-bromo-cGMP-treated worms in comparison to either the no stimulation or M9 buffer controls (p<0.01).
Epistatic analysis in C. elegans has placed DAF-12 NHR signaling downstream of the cGMP, IIS, and dauer TGFβ pathways with respect to regulation of dauer development [12], [34], [83]. Operating under the assumption that this epistatic relationship also exists during S. stercoralis L3i activation, we hypothesized that activation of L3i with Δ7-DA would not modulate ILP transcripts, since NHR signaling is downstream of IIS in C. elegans. To our surprise, Δ7-DA regulated ILP-encoding transcripts in a manner similar to 8-bromo-cGMP, which is upstream of IIS in C. elegans (Figure 4). We found that Ss-ilp-1, Ss-ilp-3, and Ss-ilp-6 transcripts were significantly increased (14-fold, 7-fold, and 10-fold, respectively) in L3i stimulated with Δ7-DA in comparison to the no stimulation control (p<0.01), similar to their regulation by 8-bromo-cGMP. These results suggest that NHR signaling may be upstream of IIS during S. stercoralis L3i activation.
We hypothesized that regulation of ILP-encoding transcripts by Δ7-DA was either a general phenomenon that would be observed with any activating condition or that IIS signaling was downstream of NHR signaling in S. stercoralis during L3i activation. The only two stimuli other than Δ7-DA that, to our knowledge, result in L3i feeding (DMEM and 8-bromo-cGMP) are both predicted to signal upstream of both IIS and NHR signaling; thus, we were unable to directly test whether regulation of ILP-encoding transcripts by Δ7-DA was a general phenomenon observed with any activating condition. However, we hypothesized that if NHR signaling was indeed upstream of IIS, we would be able to block Δ7-DA-mediated activation of L3i by inhibiting IIS. In previous work, we demonstrated that the PI3 kinase inhibitor LY294002 blocks L3i activation at 100 µM [16]. Using our in vitro L3i activation assay, we found that 100 µM LY294002 in the presence of 400 nM Δ7-DA resulted in 7.6% (±1.6%, SD) of L3i feeding in comparison to 93.3% (±1.1%, SD) in 400 nM Δ7-DA alone. Thus, this PI3 kinase inhibitor almost completely blocked the effect of Δ7-DA-mediated activation in comparison to 0.3% (±0.2%, SD) feeding in the M9 buffer+DMSO negative control (Figure 3).
Activation of S. stercoralis L3i by administration of 8-bromo-cGMP or Δ7-DA results in regulation of ILP-encoding transcripts similar to that observed in third-stage larvae three days after infection of a permissive host [14]. Furthermore, this regulation of ILP-encoding transcripts is accompanied by a phenotype, namely resumption of feeding. C. elegans ILPs regulate IIS not only by changes in transcript abundance, but also by localization of their expression in specific tissues. For this reason, given the evidence that IIS appears to play a critical role in mediating L3i arrest and activation, we identified the tissues in which the promoters of several S. stercoralis ILP-encoding genes are active. In C. elegans, the promoters of genes encoding ILPs are often active in the nervous system, intestine, and/or gonad [39], [84]. These tissues are important regulators of dauer development, longevity, and responsiveness to environmental stresses [39], [84], [85]. Previously, we reported that the transcript abundances of three S. stercoralis ILPs, Ss-ilp-1, -6, and -7, are differentially regulated during post-free-living development [14]. To determine whether these three S. stercoralis ILPs are expressed in similar tissues as C. elegans ILPs, we made promoter::egfp reporter constructs for Ss-ilp-1, -6, and -7 and expressed these in transgenic S. stercoralis post-free-living larvae.
We observed EGFP under control of the Ss-ilp-1 promoter in the hypodermis/body wall as well as a single pair of head neurons (Table 3, Figure 5A–D). The promoter activity for Ss-ilp-6 was similar to Ss-ilp-1, with EGFP observed in the hypodermis/body wall and head neurons; however, Ss-ilp-6 promoter activity was observed in several pairs of head neurons and tail neuron(s), while Ss-ilp-1 promoter activity was limited to a single pair of head neurons (Table 3, Figure 5E–H). EGFP under control of the Ss-ilp-7 promoter was localized to the intestine as well as a single pair of head neurons with a single process that extended dorsally almost to the anterior portion of the intestine (Table 3, Figure 5I–L). The location and shape of this pair of neurons is most consistent with SIAV in C. elegans.
In this study, we sought to both describe the upstream components that regulate the second messenger cGMP in S. stercoralis, including chemosensory 7TM GPCRs and heterotrimeric G proteins, and determine whether cGMP pathway signaling regulates S. stercoralis L3i activation. Additionally, we sought to elucidate the epistatic relationships between cGMP signaling, IIS, and DAF-12 NHR signaling pathways during L3i activation. We hypothesized that the cGMP-regulated chemosensory pathway may be one of the first to transduce host cues when S. stercoralis L3i encounter a permissive host. This hypothesis was based on our previous observation that the transcripts of multiple cGMP pathway components are increased in S. stercoralis L3i, suggesting that this pathway may be “poised” to transduce host cues [14], and studies demonstrating that exogenous application of 8-bromo-cGMP activates L3i of hookworm species [51]–[53]. We therefore sought to describe the components of a chemosensory 7TM GPCR signaling pathway in S. stercoralis and determine whether cGMP signaling regulates L3i activation as well as IIS and other signaling pathways.
Using RNAseq data from seven S. stercoralis developmental stages and draft S. stercoralis genomic contigs [14], we identified and characterized the developmental transcript profiles for 85 chemosensory 7TM GPCRs predicted in the S. stercoralis genome. The majority of chemosensory 7TM GPCR-encoding transcripts were found in L3i and/or L3+ (Table 1) at abundances that were low compared to other S. stercoralis transcripts. These data strongly suggest that these receptors act in a few chemosensory cells to sense host cues. Other chemosensory 7TM GPCR-encoding transcripts in S. stercoralis, which were not observed in this study, may be present in life stages such as the free-living male or autoinfective L3 (L3a), which have not yet been interrogated by RNAseq. In these stages, the encoded chemosensory 7TM GPCRs might transduce chemical signals important in mate finding or migration within the host, respectively.
The paucity of chemosensory 7TM GPCR-encoding genes in S. stercoralis in comparison to C. elegans is not entirely surprising given the evolutionary distance separating S. stercoralis (clade 10B) and C. elegans (clade 9A) [54] and the large differences in the number of chemosensory 7TM GPCRs among even closely related nematode species [19]. The differences in the number of chemosensory 7TM GPCR genes within the well-studied Caenorhabditis genus is illustrated by the fact that there are approximately 40% more chemosensory 7TM GPCR-encoding genes in the C. elegans genome (1,646 genes) than in the C. briggsae genome (1,151 genes) [19]. Thus, the chemosensory 7TM GPCR family of receptors appears to have a great deal of evolutionary plasticity in terms of absolute number and ligand specificity. Additionally, S. stercoralis, like many parasitic nematodes, predominately resides inside a nutrient-rich host or in bacteria-rich feces and thus does not need to continually navigate and adapt to a complex external environment with limited resources. The need for chemosensory 7TM GPCRs in S. stercoralis is mainly limited to L3i sensing a host, larval migration within the host, and free-living male and female mate attraction during heterogonic development. These differences in lifestyle between S. stercoralis and C. elegans may partially account for the smaller number of 7TM GPCR-encoding genes in S. stercoralis. Whether this reduction in the number of 7TM GPCR-encoding genes is a feature of all nematode parasites or only S. stercoralis is currently unknown.
We used a similar strategy to identify G proteins in S. stercoralis; however, the sequence divergence between S. stercoralis and C. elegans of the proteins these genes encode is far less than for the chemosensory 7TM GPCRs. By phylogenetic analysis, we were able to identify the C. elegans ortholog for each of the S. stercoralis Gα-, Gβ-, and Gy-encoding genes (Table 2, Figures S1, S2, & S3). Interestingly, S. stercoralis appears to have only 14 Gα-encoding genes in comparison to the 21 in C. elegans. This observation is congruous with the smaller number of chemosensory 7TM GPCRs in S. stercoralis, since fewer receptors would need fewer signal transduction molecules. Many of the nematode-specific Gα-encoding genes have transcripts that are at their peak in L3i (Table 2, Figure S2). Along with our previous observation that the Ss-gpa-3 promoter is active in amphidial neurons [80], these data are consistent with a role for S. stercoralis Gα subunits in relaying environmental and host chemosensory cues in L3i.
Using a previously established S. stercoralis L3i feeding assay [16], [68], we demonstrated that exogenous application of the membrane-permeable cGMP analog 8-bromo-cGMP stimulates L3i activation (Figure 2A) with a higher potency than that observed in experiments with other parasitic nematodes [51]–[53]. Furthermore, 8-bromo-cGMP activated L3i more quickly than a mixture of host-like biochemical cues (Figure 2B). These data suggest that increases in endogenous cGMP levels accompany S. stercoralis L3i activation upon encountering a permissive host. Since previous work has only demonstrated 8-bromo-cGMP activation in hookworm species (clade 9B) [51]–[53], which are closely related to C. elegans (clade 9A) [54], these findings from the distantly related S. stercoralis (clade 10B), where parasitism is thought to have arisen independently from the hookworm species [55], suggest a more broadly conserved mechanism of L3i activation in parasitic nematodes.
Although cGMP signaling appears to be involved in L3i activation of both hookworms and S. stercoralis, this is a departure from the role of this pathway in regulating C. elegans dauer arrest when dauer pheromone is present. In C. elegans, elevated levels of ascarosides, which accompany high population density, bind chemosensory 7TM GPCRs that activate the inhibitory G proteins Ce-GPA-2 and Ce-GPA-3 that repress the guanylyl cyclase Ce-DAF-11, ultimately decreasing cGMP levels and promoting dauer entry [31]. While ascaroside pheromones have been detected in many nematode species [86]–[88], it is difficult to envision a role for these compounds in regulating L3i development for the parasitic nematodes of many warm-blooded animals, particularly in species where all post-parasitic larvae invariably developmentally arrest in the infectious form [89]. In S. stercoralis, where post-parasitic larvae can facultatively develop to a single generation of free-living males and females, the post-free-living larvae invariably develop to L3i regardless of population density. In these cases, an L3i-promoting ascaroside seems unlikely. However, a dauer-like pheromone that regulates L3i formation has recently been described in the parasitic nematode Parastrongyloides trichosuri, which can undergo multiple rounds of free-living replication outside its animal host [90]. Thus, cGMP pathway signaling may have several roles in free-living and parasitic nematodes, including modulation of L3i and dauer development by transduction of favorable (e.g., host or food) as well as unfavorable (e.g., dauer pheromone) environmental cues.
The second aim of this study was to determine the epistatic relationships of cGMP signaling, IIS, and DAF-12 NHR signaling in regulating S. stercoralis L3i activation. We hypothesized that, as in C. elegans, cGMP signaling would be upstream of IIS and that DAF-12 NHR signaling would be downstream of IIS [12], [34]. In C. elegans, cGMP signaling regulates both the IIS pathway as well as the dauer TGFβ pathway, including modulation of their cognate peptide ligands [38], [39], [41]. Using RNAseq, we demonstrated that activation of S. stercoralis L3i by 8-bromo-cGMP was accompanied by a dramatic increase in Ss-ilp-1 and Ss-ilp-6 transcripts (Figure 4B and G). In previous work, we described Ss-ILP-1 as a putative IIS agonist, due to protein sequence similarities with C. elegans agonistic ILPs and a decrease in Ss-ilp-1 transcript levels during parasitic development [14]; thus, our data suggest that Ss-ilp-1 transcripts increase immediately following L3i activation, but then decrease again during development to the parasitic female, consistent with a role as an agonistic ILP. We similarly described Ss-ILP-6 as a putative IIS agonist due to protein sequence similarities with C. elegans agonistic ILPs and an increase in Ss-ilp-6 transcripts in third-stage larvae that were activated inside a permissive host for three days [14]; thus, a similar regulation of Ss-ilp-6 transcripts by administered 8-bromo-cGMP reinforces our assertion that stimulation with this compound mimics early in vivo L3i activation and that Ss-ilp-6 encodes an IIS agonist important in L3i activation.
We also observed modulation of dauer TGFβ ligand transcripts during 8-bromo-cGMP-stimulated S. stercoralis L3i activation. Although developmental regulation of dauer TGFβ pathway homologs and an increase in the number of dauer TGFβ ligands from one to seven in S. stercoralis suggests a different role for this pathway than in C. elegans, the dauer TGFβ pathway does appear to be important in L3i, since three of the TGFβ ligands have transcripts only detected in this developmental stage [14]. Our observation that Ss-tgh-1, -2, and -3 transcripts are all decreased following L3i activation (Figure S4B-D) is consistent with these previous findings, which infer a role for TGFβ signaling in parasitic nematodes that is opposite to its apparent role in C. elegans dauer regulation [57]–[61]. Together, our RNAseq results suggest that, as in C. elegans, cGMP pathway signaling is upstream of both IIS and dauer TGFβ signaling in S. stercoralis.
In C. elegans, DAF-12 NHR signaling is downstream of IIS in regulating dauer entry, as evidenced by genetic epistatic analysis and rescue of the daf-c phenotype in daf-2(e1368) worms by Δ7-DA [34], [44], [47]. Therefore, we hypothesized that Δ7-DA-mediated L3i activation would not involve modulating ILP transcripts or IIS. Surprisingly, we found that the profiles of ILP transcript abundance were almost identical in 8-bromo-cGMP-mediated and Δ7-DA-mediated L3i activation (Figure 4). We reasoned that either modulation of ILP transcripts levels, and thus IIS, is a non-specific feature of S. stercoralis L3i activation or that our assumption about pathway ordering was incorrect. To test this, we utilized LY294002, a potent inhibitor of PI3 kinases such as Ss-AGE-1, which blocks L3i feeding when activating biochemical host-like cues are present [16]. We found that LY294002 almost completely abolishes Δ7-DA-mediated L3i activation (Figure 3). While we cannot entirely account for off-target effects of LY294002, which inhibits all PI3 kinases, only three classes of PI3 kinases are present in nematodes and the class I PI3 kinase Ce-AGE-1 is exclusively associated with dauer development. Ss-AGE-1 is the ortholog of Ce-AGE-1 and the sole class I PI3 kinase in S. stercoralis [16]. Therefore, it is almost certainly the mediator of LY294002's effect. Together, the data from RNAseq and chemical inhibitor studies strongly suggest that DAF-12 NHR signaling acts upstream of IIS during L3i activation.
Since our data strongly suggest that regulation of IIS by ILPs is crucial for L3i activation, we identified the tissues in which ILP promoters are active during post-free-living development. Using previously established techniques [15], [76], [80], we expressed S. stercoralis ILP promoter::egfp reporter constructs in the post-free-living generation. We found that Ss-ilp-1 and Ss-ilp-6 promoters are active in head neurons as well as the hypodermis/body wall (Figure 5A–H); in previous work, we hypothesized that both of these ILPs act as IIS agonists [14]. Additionally, we found that the Ss-ilp-7 promoter is active in both a single pair of head neurons as well as the intestine (Figure 5I–L). We previously hypothesized that Ss-ilp-7 acts as an IIS antagonist [14]. The promoter activity of S. stercoralis ILPs in hypodermal, neuronal, and intestinal tissues is consistent with anatomical locations of C. elegans ILP promoter activity [84].
Together, our data suggest a model of S. stercoralis L3i activation in which parallel cGMP and DAF-12 NHR signaling co-regulate the downstream IIS pathway via modulation of ILPs. This model differs in several significant respects from current models of canonical dauer pathway regulation in C. elegans. First, C. elegans decreases cGMP pathway signaling in response to dauer pheromone, resulting in dauer arrest [31], [36]. In contrast, we hypothesize that in S. stercoralis L3i, host compounds bind chemosensory 7TM GPCRs, which activate G proteins that in turn activate guanylyl cyclases that increase cGMP levels; this ultimately triggers L3i to activate and resume development, in part through increased IIS via increases in agonistic ILPs [14], [16]. Second, epistatic analysis in C. elegans has demonstrated that dauer development is regulated by upstream cGMP signaling that regulates IIS, which in turn regulates downstream DAF-12 NHR signaling [12]. However, the data in this study suggest that both cGMP and DAF-12 NHR signaling lie upstream of IIS in regulating L3i activation, further emphasizing the importance of IIS in S. stercoralis L3i development.
|
10.1371/journal.pbio.1001243 | Sequential Analysis of Trans-SNARE Formation in Intracellular Membrane Fusion | SNARE complexes are required for membrane fusion in the endomembrane system. They contain coiled-coil bundles of four helices, three (Qa, Qb, and Qc) from target (t)-SNAREs and one (R) from the vesicular (v)-SNARE. NSF/Sec18 disrupts these cis-SNARE complexes, allowing reassembly of their subunits into trans-SNARE complexes and subsequent fusion. Studying these reactions in native yeast vacuoles, we found that NSF/Sec18 activates the vacuolar cis-SNARE complex by selectively displacing the vacuolar Qa SNARE, leaving behind a QbcR subcomplex. This subcomplex serves as an acceptor for a Qa SNARE from the opposite membrane, leading to Qa-QbcR trans-complexes. Activity tests of vacuoles with diagnostic distributions of inactivating mutations over the two fusion partners confirm that this distribution accounts for a major share of the fusion activity. The persistence of the QbcR cis-complex and the formation of the Qa-QbcR trans-complex are both sensitive to the Rab-GTPase inhibitor, GDI, and to mutations in the vacuolar tether complex, HOPS (HOmotypic fusion and vacuolar Protein Sorting complex). This suggests that the vacuolar Rab-GTPase, Ypt7, and HOPS restrict cis-SNARE disassembly and thereby bias trans-SNARE assembly into a preferred topology.
| Cellular components often travel between organelles in vesicular entities. This intracellular traffic usually involves production of a vesicle containing cargo from one organelle membrane, movement of the vesicle to its destination, and then fusion of the vesicle with the target organelle. Thus, membrane fusion is a fundamental process required for these intracellular trafficking events. SNARE proteins and SM proteins mediate this fusion process. SNAREs form complexes that are either located on the same membrane or vesicle (called cis-SNARE complexes) or bridge two membrane compartments or vesicles (trans-SNARE complexes). The cis-SNARE complexes must be activated before trans-SNARE complexes can form and allow the membranes to fuse. We investigated the mechanism of cis-SNARE activation and trans-SNARE formation by studying the fusion of highly purified yeast vacuoles. We found that cis-SNARE activation involves the selective removal of a single SNARE protein from a pre-existing cis-SNARE complex, which is replaced by a similar SNARE originating from the other fusion partner. The activated cis-SNARE complexes depended on SM proteins for their stability. Thus, we have shown that the preferred topology of trans-SNARE formation is determined by cis-SNARE–SM protein interactions.
| Cognate combinations of v- and t-SNAREs mediate membrane docking in every vesicular transport step in the endomembrane system. v- and t-SNAREs bind each other in coiled-coil complexes containing four helices [1]. Three of these helices (termed Qa, Qb, and Qc) are provided by t-SNAREs, while one (termed R) is provided by the v-SNARE [2],[3]. SNAREs from the two fusing membranes form trans-complexes between the membranes [4],[5]. It is clear that these trans-complexes are necessary for subsequent fusion; however, the pathway leading to their establishment is less well understood. Important observations have been obtained from experiments with purified SNAREs, studied either in detergent solution or after reconstitution into artificial lipid bilayers [4]. SNAREs form four-helix QabcR complexes. They are a substrate for the ATP-driven chaperone Sec18/NSF, which completely disrupts pure SNARE complexes [6],[7]. Singular SNAREs can re-associate into complexes. Depending upon the method chosen, this re-association can occur in different topologies.
In most in vitro studies with reconstituted SNAREs in proteoliposomes, fusion-active combinations of SNAREs were found to require the Qa, Qb, and Qc-SNARE in one membrane and the R-SNARE on the other vesicle [4],[8]–[10]. The preference for this contribution is consistent with the fact that Qabc-containing SNARE complexes can form without an R-SNARE [3]. Thus, co-reconstituting these three Q-SNAREs into one liposome likely favors the formation of this Qabc complex, which is a suitable receptor for the R-SNARE coming from the other fusion partner after mixing of the vesicles. However, similar reconstitution approaches with endosomal SNAREs yielded functional combinations of SNAREs in several different topologies [11], suggesting that different distributions of Q- and R-SNAREs over the two membranes can in principle form functional complexes. Studies with single SNARE molecules indicated even greater flexibility. For example, they showed the existence of anti-parallel associations of SNAREs. Thus, SNAREs can associate in multiple topologies [11]–[14].
An important question arising from these observations is whether in a cellular environment SNAREs have the liberty to associate in such variable topologies, or whether SNARE-associated tethering and docking proteins, such as the Rab-GTPases and their effector proteins, might control cis-SNARE disruption and guide subsequent SNARE association in a preferred topology in trans. Research on reconstituted, purified SNAREs has only recently begun to incorporate SNARE-associated proteins, such as SM-proteins, Rabs, and Rab-effectors into the reconstitutions [10],[15]–[16]. However, these co-reconstitutions have not yet been used to explore a possible effect of these components on the assembly pathway and topology of SNARE complexes. In physiological membranes, the issue of cis-SNARE complex disassembly and its transformation into trans-complexes has been studied on endosomes [17]. Endosomes contain the cognate set of endosomal SNAREs necessary for their fusion, as well as SNAREs for fusion at the plasma membranes. Cognate as well as non-cognate combinations of SNAREs could be co-precipitated from endosomal fractions, leading to the conclusion that there is promiscuity in cis-SNARE complex assembly.
Cis-SNARE complexes can accumulate as a product of a preceding fusion reaction or by spontaneous re-association of separated SNAREs in cis. They are reactivated by Sec18/NSF and its cofactor Sec17/α-SNAP [6],[18]. These chaperones are generally believed to completely disrupt SNARE complexes, liberating the individual SNARE for subsequent reassembly into trans-SNARE complexes [5],[19]. This reassembly and the ensuing fusion depend on Rab-GTPases and SM proteins [20]. These conserved factors can interact and cooperate with further compartment-specific factors, e.g. Rab effectors (tether factors) and lipids such as phosphatidyl-inositol-3-phosphate, Munc13, or complexin [21]. Addition of purified NSF completely disrupts complexes of purified SNAREs as well as SNARE complexes on isolated endosomal or vacuolar membranes [22]. Furthermore, the t-SNARE subunits SNAP25 and syntaxin1 reside in non-overlapping membrane patches in plasma membrane sheets of PC12 cells, suggesting that these two SNAREs remain spatially separated on this membrane [23].
We have used the cell-free fusion of purified vacuoles as a model reaction to follow the fate of cis-SNARE complexes and their conversion into trans-complexes [24]. Yeast vacuoles harbor five SNAREs that are necessary for their fusion [22]. The Qa SNARE Vam3, the Qb SNARE Vti1, the Qc SNARE Vam7, and the R-SNARE Nyv1 form the trans-SNARE complexes required for fusion. An additional R-SNARE, Ykt6, can be associated with these SNAREs. However, Ykt6 was not found in trans-SNARE complexes [25] and a significant fraction of Ykt6 leaves the vacuolar membrane upon priming [26].
In this study, we took three approaches in order to investigate the transition from cis- into trans-SNARE complexes in vacuole fusion. First, we tested which subunits of the trans-SNARE complex are contributed by one or the other fusion partner. Second, we created inactive SNARE mutations in the two fusion partners in combinations that allowed us to test the functional significance of the observed SNARE association. And third, we analyzed priming, the activation of cis-SNARE complexes, asking whether activation of cis-SNARE complexes indeed leads to complete disruption of cis-SNARE complexes. We tested, whether incomplete dissociation of cis-SNARE complexes might prejudice the trans-association of SNAREs in a certain topology.
In the past years the topology of trans-SNARE formation mainly has been addressed by employing liposome fusion systems, in which recombinantly expressed SNAREs have been reconstituted and tested for fusion activity [4],. While most of these studies gave evidence for a preferred Qabc-R topology, others indicated an alternative possibility of trans-SNARE formation [11],[13]. Moreover, experiments conducted under more physiological conditions suggested a preferred QbcR-Qa trans-SNARE topology [27],[28]. To unravel these contradictory observations, we decided to investigate the topology of trans-SNARE formation in the vacuolar fusion system. To accommodate recent reports that oxidation might affect SNARE function [29],[30], we strictly decided to work in all following experiments under reducing conditions by adding DTT to fusion reactions and detergent extracts. Indeed, by using non-reducing SDS-PAGE, we noticed that some vacuolar proteins change their migration behavior, suggesting that oxidation might occur during fusion and detergent extraction (Figure S1).
We used differential tagging of vacuolar SNAREs to probe the topology of the trans-SNARE complexes formed during vacuole docking. In agreement with published observations [31], we noticed that tagging vacuolar SNAREs on their cytoplasmic N-terminus interferes with fusion activity (C. Peters, unpublished results). Consequently, we fused all tags to the C-termini, which for the membrane-anchored SNAREs are at the lumenal face of the vacuolar membrane. All tagged SNAREs were expressed from their authentic loci in the genome under the control of their native promoters, i.e., no non-tagged allele of the respective SNARE was left in these cells (Table S1). The tagged strains were viable and grew normally. The expression levels of the proteins on vacuoles isolated from the tagged strains were normal (Figure S2), and their fusion activities were also comparable to those of untagged vacuoles (Figure S3).
Vacuole docking depends on trans-complex formation between Vam3-Qa, Vti1-Qb, Vam7-Qc, and Nyv1-R [5],[19]. Vam3-Qa, Vti1-Qb, and Nyv1-R are integral membrane proteins, whereas Vam7-Qc is anchored to vacuoles by the phosphatidylinositol-binding PX domain [32]. Both fusion partners carry the same set of SNAREs, but vacuoles from strains expressing differently tagged SNAREs can be mixed in vitro. Differential peptide tagging thus allows the investigation to distinguish cis-associations occurring within the same membrane from trans-associations between SNAREs originating from the apposed fusion partners. Starting a fusion reaction with ATP produces trans-associations, which lead to fusion and hence are converted into post-fusion cis-complexes. In order to prevent this conversion, it is desirable to block fusion at a late stage. We noted that after an initial incubation for 5 min at 27°C, subsequent cooling of the reaction to 7°C efficiently suppresses fusion; we used this simple technique to accumulate docked vacuoles (Figure S4A). We tested whether the vacuoles could prime and dock by two-stage incubations, exploiting the fact that completion of priming renders the further course of fusion resistant to antibodies to Sec18p, while completion of docking renders it resistant to anti-Ypt7 [33],[34].
We incubated vacuoles under fusion conditions at 27°C for a 5 min period with control buffer or antibodies to Sec18p and Ypt7p, respectively, in order to stop further priming and docking in the presence of ATP (Figure S4B). Then, the reaction continued either at 27°C or 7°C for 30 min. In the absence of inhibitors, vacuoles arrested at 7°C efficiently completed fusion during the second incubation at 27°C for 30 min. They also did so in the presence of anti-Ypt7p or anti-Sec18p during the second incubation, suggesting that the initial pre-incubation at 7°C had rendered them resistant and permitted completion of priming and docking. If those two inhibitors already were present during the first incubation, no significant fusion was observed (Figure S4B). This suggests that at 7°C, the reaction passes the priming and docking stages and arrests at a productive intermediate stage beyond docking. Fusion inhibition at 7°C was not due to decreased reporter maturation, since adding Triton X-100 to the vacuoles, allowing fusion-independent maturation of pro-ALP, did not result in significantly different ALP activities at 7°C and 27°C (Figure S4C).
We therefore used this 7°C incubation in order to accumulate trans-SNARE complexes and probe their topology (Figure 1). We mixed the Nyv1-HA(R) vacuoles either with Vam3-VSV(Qa), Vam7-VSV(Qc), or Vti1-VSV(Qb) vacuoles to test for a QbcR-Qa topology. We mixed Vam3-HA(Qa) vacuoles either with Nyv1-VSV(R), Vam7-VSV(Qc), or Vti1-VSV(Qb) to probe for a Qabc-R topology. After a 7°C incubation with ATP for 30 min, a time that is sufficient for complete docking [33],[34], the membranes were solubilized and immunoprecipitated against the HA tag. The degree of trans-association between the HA-tagged and VSV-tagged strains was assayed by Western blotting.
The observed results fell into two categories. The trans interactions among Vam3-Nyv1 (Qa-R), Vam3-Vam7 (Qa-Qc), and Vam3-Vti1 (Qa-Qb) increased from −ATP to +ATP fusion reaction. The increase was sensitive to the docking inhibitor GDI (Figure 1B, Figure 1D, and Figure 1F; Text S1). The trans interactions of Nyv1-Vam7 (R-Qc) and Nyv1-Vti1 (R-Qb) were comparatively much weaker than the Nyv1-Vam3 (R-Qa) or the Vam3-Vam7 (Qa-Qc) and Vam3-Vti1 (Qa-Qb) interactions (Figure 1A, Figure 1C, and Figure 1E). We also looked for homophilic interactions by tagging the same SNARE in both fusion partners with different tags, e.g., Vti1-HA (Qb) on one vacuole and Vti1-VSV (Qb) on the other. We could not detect any homophilic trans-interactions between Vti1-HA-Vti1-VSV (Qb-Qb), Vam3-HA-Vam3-VSV (Qa-Qa), or Nyv1-HA-Nyv1-VSV (R-R) (unpublished data). Thus, we did not obtain any indications that trans-SNARE complexes might multimerize.
In sum, our observations suggest that Nyv1-R, Vam7-Qc, and Vti1-Qb are retained in a partial cis-SNARE complex that incorporates Vam3-Qa from the other fusion partner. The resulting trans-SNARE complex hence should show a preferred QbcR-Qa topology, i.e., the Qb, Qc, and R-SNARE are predominantly contributed from the same membrane, whereas the syntaxin-like SNARE Qa might act alone on the other fusion partner. It should be noted that we consider this as a preferred topology, since a certain amount of trans interactions between Nyv1-Vam7 (R-Qc) and Nyv1-Vti1 (R-Qb) can also be observed (Figure 1C and Figure 1E).
In principle these complexes can emerge not only from trans-SNARE pairing, but also from cis associations that might occur after fusion of the two differentially labeled vacuoles, or after solubilization of the membranes. In addition to the controls described above, two observations in the immunoprecipitation experiments argue against this and show that these SNARE complexes connected the apposed membranes before fusion. First, the efficiency of the coprecipitations was much lower if the membranes had been incubated in the absence of ATP, which prevents SNARE priming and fusion [5],[18]. Second, the Ypt7p inhibitor, GDI, reduced the associations. Thus, the trans associations depend on docking. These criteria argue in favor of the genuine existence of trans-SNARE complexes as displayed in Figure 1. Additionally, we excluded a random SNARE association occurring in the solubilizate by mixing primed detergent extracts of differently tagged versions of SNAREs, and found no random intermixing of these SNAREs into preexisting QbcR complexes (Figure S5).
A Qa-QbcR topology differs from the generally held Qabc-R model that a trans-SNARE complex assembles from a Qabc SNARE subcomplex from one fusion partner and a single R-SNARE from the other fusion partner [3]. Therefore, we sought to test whether the Qa-QbcR topology of the trans-SNARE complex corresponds to a functional restriction of SNARE requirements during vacuole fusion. To this end, we analyzed fusion reactions between vacuoles carrying combinations of SNARE mutations that allow the investigation to distinguish between the Qabc-R and Qa-QbcR topologies (Figure 2). In an in vitro vacuole fusion assay, one can distinguish the two fusion partners because one vacuole type (BJ3505) contains a pro-alkaline phosphatase in the lumen, while the other contains the maturation enzyme (DKY6281). These two vacuole populations are separately prepared and mixed in vitro. Fusion between them generates mature alkaline phosphatase, whose activity serves to quantify fusion [24]. Despite the fact that the two fusion partners have different content, they have an identical pool of SNAREs in this homotypic fusion system. Topologically restricted trans-SNARE complexes can, hence, form in two orientations in a wildtype situation (Figure S6). Therefore, combinations of at least two mutations have to be distributed over the two fusion partners to circumvent this problem (Table S1). Vam3-Qa, Vam7-Qc, and Nyv1-R genes can be deleted without compromising viability. Since Vti1-Qb is essential, we used the conditional vti1-1(Qb) allele expressing a Vti1-Qb protein that is inactivated at 40°C but remains functional at 25°C [35]. Cells can, therefore, be grown with functional Vti1-Qb. This Vti1-Qb can then be inactivated by shifting the cells to 40°C during vacuole isolation (Figure 2A). If both fusion partners carried the vti1-1 allele, fusion was blocked after pre-incubation at 40°C because neither fusion partner retains a functional Qb-SNARE, as shown earlier [22]. If only one fusion partner carried vti1-1(Qb), but the other had the wildtype allele, fusion still proceeded efficiently, showing activities that were similar after preincubation at 40°C to those after preincubation at 25°C. After 40°C preincubation of the WT/vti1-1(Qb) combination, Vti1-Qb remained functional only on the wildtype side (Figure 2A; Figure S6). This permits assembly of functional trans-SNARE complexes, but only in one orientation. In this situation we can ask which side contributes a certain SNARE subunit and functionally discriminate trans-SNARE topologies.
To discriminate between the different topology models, we deleted the Nyv1-R gene in one fusion partner and inserted the vti1-1(Qb) allele into the other. For this combination, the Qabc-R model predicts fusion because the nyv1Δ-R vacuole can provide a complete Qabc t-SNARE and the vti1-1(Qb) vacuole can provide an R-SNARE. The Qa-QbcR model, in contrast, predicts inhibition because neither fusion partner can provide the necessary QbcR combination in one membrane (Figure S6). In the experiment nyv1Δ-R vacuoles fused with vti1-1(Qb) vacuoles after preincubation at 25°C, but displayed reduced fusion efficiency (60%) after preincubation at 40°C, which induces the mutant phenotype (Figure 2B; Figure S6). This result stands in support of a preferred Qa-QbcR topology.
We created a second combination of mutations that allowed us to discriminate between the two models by mutating Vti1-Qb and Nyv1-R in the same membrane (Figure 2C; Figure S6). According to the Qabc-R hypothesis, such vti1-1(Qb) nyv1Δ-R vacuoles should not even fuse with a wildtype vacuole because they can neither provide a functional R-SNARE nor a functional Qabc-SNARE (Figure S6). The Qa-QbcR model predicts fusion for this combination because the vti1-1(Qb) Δnyv1-R vacuole can provide Qa SNARE, which can pair with QbcR from the wildtype partner. We observed that the vti1-1(Qb) nyv1Δ-R mutant vacuoles fused almost equally well with wildtype vacuoles after pretreatment at 40°C or 25°C. One might invoke redundancy with other R-SNAREs to explain the remaining activity. This is unlikely, because only 10% residual activity remained when vti1-1(Qb) nyv1Δ-R vacuoles were incubated with fusion partners lacking functional Nyv1-R or Vti1-Qb (Figure 2C; Figure S6). Fusion between vti1-1(Qb) nyv1Δ-R and wildtype vacuoles, hence, depended on Vti1-Qb and Nyv1-R, and did not result from substitution by another, non-vacuolar R-SNARE.
According to the Qa-QbcR model, a vacuole containing an inactive Qa and Qb should be fusion incompetent, even in combination with a wildtype fusion partner. Such a double mutant has neither a functional Qa nor a functional QbcR combination (Figure S6). The Qabc-R model, in contrast, predicts fusion because the wildtype vacuole could provide a complete Qabc SNARE and the Qa/Qb double mutant still carries a functional R-SNARE. We tested this combination by inserting temperature sensitive vti1-1(Qb) and vam3tsf-Qa alleles [35],[36] into the same strain (Figure 2D; Figure S6). Although Vam3-Qa is not essential, we had to use the vam3tsf-Qa allele because vam3Δ-Qa vacuoles also lack Vam7-Qc and, hence, are not suitable for this type of analysis. The cells were grown at 25°C and subjected to a brief 40°C or 25°C treatment during the spheroplasting step of vacuole isolation. Isolated mutant vacuoles were then fused to wildtype vacuoles. In this situation, when only one of the two mutations was present, preincubation at 40°C hardly reduced fusion activity (20%) in comparison to preincubation at 25°C. A more severe effect was observable when the two mutations were distributed over the two fusion partners (reduction of fusion efficiency of about 50%). However, when vti1-1(Qb) and vam3tsf-Qa mutations were combined in the same vacuole, the severest fusion defects became apparent. Already at 25°C, the double mutant vacuoles retained only 50% of the activity of the single mutants. Upon brief pretreatment at 40°C, fusion was almost completely suppressed. This result is consistent with the Qa-QbcR model.
To exclude a priming defect caused by the double SNARE mutation in the vti1-1 (Qb) vam3tsf-Qa mutant, we tested SNARE-complex stability by immunoprecipitation of Nyv1-R from detergent extracts of wildtype and mutant vacuoles incubated at restrictive temperature. While wildtype vacuoles showed the persistence of a QbcR complex, mutant vacuoles displayed an unstable complex but were able to prime (Figure S7A). As a further control, we tested the influence of a nyv1Δ-R mutation in the vam3tsf-Qa background. As expected, this mutant fused as well as the single vam3tsf-Qa at permissive temperature, but lost fusion activity at restrictive temperature due to inactivation of R and Qa SNARE on the same membrane (Figure 2D; Figure S6). Additionally, we measured reporter loading of the different SNARE-mutants by incubating DKY and BJ vacuoles in the presence of Triton X-100. We found that in the presence of the vam3tsf-Qa mutation, BJ vacuoles contained only 50% of pro-ALP loading. We accommodated this by doubling the incubation time in developing buffer (Figure S7B).
Based on the observation that both biochemical and functional analysis revealed a preferred QbcR-Qa topology in vacuolar trans-SNARE formation, we asked whether a stabilization of a primed Vam7/Vti1/Nyv1 (QbcR) complex in cis could prejudice the topology of the trans-SNARE complexes that form during subsequent docking.
Isolated vacuoles contain cis-SNARE complexes that are activated and disrupted by Sec18 upon addition of ATP [22]. We confirmed this result when working under similar conditions. However, as mentioned in the beginning of the Results section, we performed our fusions and immunoprecipitations in the presence of DTT and investigated how cis-SNARE complexes behaved under this condition. In order to monitor the assembly of cis-SNARE complexes, we first precipitated Vam3-Qa from detergent extracts. Specifically, we were interested in determining whether a persistent post-priming Qabc (Vam3-Qa, Vti1-Qb, Vam7-Qc) complex might be established after the NSF-mediated priming process, as predicted by the current model of trans-SNARE formation [3]. Although this Qabc-complex formation has been demonstrated for recombinant proteins and is routinely used in liposome fusion assays, it has not yet been detected on physiological membranes.
SNARE-activation was started by addition of ATP. Vacuoles that did not receive ATP could not activate their cis-SNARE complexes, and hence served as a negative control. After 5 min, EDTA was added in order to stop further hydrolysis of ATP. This short period of ATP exposure only allows cis-SNARE assembly, since trans-SNARE formation depends on docking and needs longer time to occur [24]. In vacuoles incubated without ATP, Vti1-Qb and Vam7-Qc co-fractionated with Vam3-Qa as expected (Figure 3, left panel). Consistent with earlier experiments [5],[22] a substantial part of Vti1-Qb, Vam7-Qc, and Nyv1-R were released from Vam3-Qa in the presence of ATP. Surprisingly, when Nyv1-R was precipitated from wildtype vacuoles, we did not observe this instability for a QbcR complex (Figure 3, right panel). In contrast to Vam3-Qa, Vam7-Qc and Vti1-Qb remained tightly associated with Nyv1-R in the presence of ATP. We quantified the difference in persistence of QbcR and Qabc complexes in the presence of ATP (Figure 3). Vam3-Qa lost about 50% of associations with all other SNAREs, whereas Nyv1-R only was separated from Vam3-Qa (at a rate of about 50%) and retained association with Vam7-Qc and Vti1-Qb at an extent of almost 100%. We interpreted this result as a preferred generation of a stable post-priming QbcR complex instead of an expected Qabc complex, although this effect was not absolute, since also a substantial part of Qabc-complexes sustained ATP exposure.
Is maintenance of cis-SNARE associations relevant to the establishment of trans-SNARE complexes and to subsequent fusion? In order to address this question, we tested three different conditions that destabilize cis-SNARE interactions for their effect on trans-SNARE pairing and fusion.
First, we used excess rSec18 (Text S1) as a tool to specifically destroy cis-SNARE complexes and correlated this with the inhibitory effect of excess rSec18 on vacuolar fusion during the priming phase. In vacuole fusion, priming (cis-SNARE activation) and docking (trans-complex formation) can be distinguished by determining the time point at which a fusion reaction becomes resistant to the addition of different inhibitors [18],[24]. We tested the effect of excess of Sec18/NSF on the priming or docking phase. We used antibodies to Sec17/α-SNAP, which inhibits priming (acts on priming phase of fusion curve, 0–15 min), and antibodies to the vacuolar Rab-GTPase Ypt7 or GDI, which inhibits docking (acts on docking phase of fusion curve, 0–30 min; Figure 4A). Numerous parallel fusion reactions were started. The inhibitors were added at different times after the onset of a fusion reaction. After addition of the inhibitor, the incubation was continued at 27°C until the end of the normal fusion period, and finally fusion activity was assayed. Control samples received only buffer before being re-transferred to 27°C, or they were set on ice in order to stop the reaction at this time point. The fusion reactions became resistant to excess rSec18/NSF after 15 min, with the same time course as to anti-Sec17/α-SNAP. Resistance to anti-Sec17/-αSNAP is a marker for the completion of priming. Resistance to anti-Ypt7 or GDI as markers for the completion of docking was attained after 30 min, the time at which the docking reaction is completed. This suggests that excess rSec18/NSF affects the priming phase of vacuole fusion but is not inhibiting docked vacuoles that have passed this stage.
Based on this observation, we investigated its influence on the stability of cis-SNARE complexes, whose existence locate to the same time period. The rationale of this experiment is that we tried to force a disassembly of reduced cis-SNARE complexes by adding an excess of purified rSec18/NSF to ATP-containing fusion reactions (Figure 4B), thereby gaining evidence for a fusion relevant role for these complexes. Indeed, increasing concentrations of rSec18/NSF gradually destabilized the association of Vti1-Qb and Vam7-Qc with Nyv1-R (QbcR). This destabilization was not observed, even with the highest concentration of rSec18/NSF, when ATP was omitted from the incubation (unpublished data).
To monitor proper rSec18 activity, we subjected each Sec18 preparation to a quality control employing wildtype and vtc4Δ vacuole fusion reactions (Figure S8). The vacuolar Vtc-complex comprises multiple subunits and displays a polyphosphate synthase activity [37], which is for yet unknown reasons linked to Sec18 activity. Vacuoles purified from vtc4Δ strains strictly depend on the addition of functional rSec18 for their fusion activity since endogenous Sec18 function is impaired on these vacuoles [38]. Therefore, addition of rSec18 to vacuoles derived from vtc4Δ strains leads to stimulation of fusion at lower concentrations, but to inhibition of fusion at higher concentrations as observed for wildtype vacuoles (Figure S8).
As the addition of GDI led to the inhibition of trans-SNARE formation and destabilized cis-QbcR complexes (Figure 1 and Figure 3), we speculated whether excess rSec18/NSF might influence the interaction of the QbcR complex with the Ypt7-effector HOPS. HOPS is the tethering complex of vacuolar system composed of six different subunits, one of which is termed as Vps39 [20]. If the physical presence of HOPS is needed for stabilizing the post-priming QbcR complex, excess Sec18/NSF might compete for or prevent the binding of the QbcR complex to HOPS.
We therefore probed for the presence of Vps39 in the Nyv1-R-precipitations in the presence of increasing amounts of rSec18/NSF (Figure 4C). The concentration range in which Sec18/NSF destabilized the cis-SNARE associations led to a corresponding decrease in association with HOPS, indicating that HOPS and Sec18/NSF compete for binding to the QbcR complex. Concomitantly, fusion activity of the vacuoles decreased with increasing concentrations of Sec18/NSF (Figure 4B). While this decrease of fusion activity correlates to the disassembly of the cis-SNARE interactions, it could also reflect the disassembling activity of Sec18/NSF on trans-SNAREs. This appears unlikely, since the kinetic analysis displayed in Figure 4A excludes a direct effect of Sec18/NSF on trans-SNARE complexes, suggesting that they are resistant to disassembly, consistent with the increased NSF resistance of trans-SNARE complexes observed in a liposome system [10],[39].
Second, we deliberately oxidized vacuoles and probed the stability of cis-SNARE complexes under this condition in order to investigate the consequence of unstable QbcR -complexes for the following trans-SNARE establishment (Figure 5 and Figure S9). We tested this by mixing Nyv1-HA(R) vacuoles with Vam3-VSV(Qa) vacuoles. Mixing these two populations allows us to identify trans-interaction (Nyv1-HA/Vam3-VSV, R-Qa). After 30 min of incubation in the presence of ATP, trans-interactions increased significantly. These trans-interactions were sensitive to GDI, which inhibits the vacuolar Rab-GTPase, Ypt7p (Figure 5A), and thereby prevents tethering and docking [20]. In contrast, oxidized vacuoles did not form ATP-dependent trans-SNARE interactions (Figure 5B) even though the priming of the cis-SNARE complexes occurs normally, as evident from the ATP-dependent destabilization of the Nyv1-HA/Vam7 (R-Qc) interaction (Figure 5B).
Third, we inactivated Ypt7 by addition of GDI and asked whether this might influence cis-SNARE-stability and give evidence for the involvement of the tethering machinery in cis-SNARE complex stabilization. The fact that GDI is an effective inhibitor of trans-SNARE formation (Figures 1 and 5A) led us to speculate about a possible influence of this inhibitor on cis-SNARE stability. This is not evident from the kinetic analysis displayed in Figure 4A, as the inhibitory effect of GDI is clearly located on the docking curve. But this does not exclude that GDI might affect fusion components at an earlier stage of membrane fusion, since the architecture of the kinetic experiment shown in Figure 4A only resolves the latest fusion inhibitory effect of GDI. Moreover, the observation that excess rSec18 already inhibits the interaction of HOPS with the QbcR-complex in the priming reaction (Figure 4C) points to a possible role of Ypt7 during an earlier phase of vacuolar fusion. Indeed, addition of GDI destabilizes the QbcR–complex, indicating that HOPS and Ypt7 are required for the persistence of the QbcR-complex during the priming phase of vacuolar fusion (Figure 3B).
Taken together, these findings suggest that destruction of this cis-SNARE association by excess Sec18/NSF, or by oxidation of the vacuoles, or by Rab-inactivation leads to inefficient trans-SNARE pairing and fusion deficiency.
To further confirm that members of the tethering machinery are indispensable for stabilizing a post-priming QbcR-complex, we tested the dependence of cis-SNARE pairing on the Rab-GTPase, Ypt7, and its GEF, the Ccz1/Mon1 complex ([40]–[42]; Figure 6), and on the Ypt7 effector complex subunit Vps41p [43],[44]. We assayed the existence of the Vam7/Vti1/Nyv1 (QbcR) association in Δvps41 vacuoles, in ccz1Δ vacuoles, and in ypt7 vacuoles expressing the T22N allele of Ypt7, which produces Ypt7 protein mimicking the GDP-bound state [45]. Since the vps41Δ mutant showed significantly reduced Vam7-Qc levels on the vacuoles (Figure 6A), but not in the whole cell (Figure S10), we corrected this deficiency by over-expressing Vam7-Qc, yielding mutant vacuoles that were similar to wildtype vacuoles (Figure 6B). Despite almost normal wildtype expression levels of Vam7-Qc, mutant vacuoles showed strongly reduced cis-SNARE association of Nyv1-R, Vam7-Qc, and Vti1-Qb (Figure 6A and Figure 6B). Upon addition of ATP, these strongly reduced levels did not decrease further, suggesting that they represent the background of the assay. The absence of functional cis-SNARE complexes in all mutants leads to fusion-incompetent vacuoles (Figure 6C and Figure 6D). These results might explain earlier findings [42],[43] and support the notion that the stabilization of a cis-SNARE complex of Nyv1-R, Vam7-Qc, and Vti1-Qb depends on the GTP-bound form of the Rab-GTPase Ypt7 and on the presence of a functional HOPS complex.
The involvement of the tethering factors Ypt7 and HOPS in post-priming SNARE-complex stabilization presents an implication that trans-SNARE interactions may help to generate these SNARE complexes. In order to exclude this possibility, and to clearly demonstrate that Ypt7 and HOPS act in cis to stabilize post-priming SNARE complexes, we performed dilution experiments. Prior to the addition of salt, isolated vacuoles were diluted up to a density that does not support fusion. This is an indication that contact is lacking between vacuoles and, therefore, establishment of trans-SNARE interactions is not to be expected. We found no difference in QbcR complex stability between vacuoles fused under standard conditions and diluted vacuoles (Figure 7), clearly demonstrating that Ypt7 and HOPS act in cis to stabilize the post-priming QbcR-complex.
The assembly pathway for trans-SNARE complexes and the resulting topology are of fundamental importance for the control of fusion reactions. The fact that Rab-GTPases and tether proteins must stabilize subcomplexes of SNAREs will determine whether these proteins must act on vesicles or target membranes and will determine the possibilities for control of fusion reactions by signaling cascades. Control by external signals can only be studied once the assembly process has been elucidated. The final topology of the assembled trans-SNARE complex, i.e., the distribution of its subunits over the two membranes, should also influence its activity. SNARE subunits are membrane-anchored, in most cases by transmembrane helices and in few others by lipidation or by lipid binding domains [46]. It has been proposed that the orientation of the SNARE complex should influence its capacity to exert stress on the membrane, disturb the bilayer structure, and induce fusion [47]. Whether a given subunit of the trans-SNARE complex is anchored in one fusion partner or the other must determine the rotational orientation of the complex between the two membranes (see model in Figure 8). Since the SNARE complex itself is of considerable size—and hence an obstacle to direct contact between the lipid bilayers [48], its twisting could induce strong local strain on the bilayer, using the large hydrophilic part of the complex as a lever. Therefore, it appears likely that the assembly and situation of the trans complex are restricted and controlled by cells.
Studies with purified SNAREs, both in soluble or liposome-associated form, indicated that the Qa, Qb, and Qc helices spontaneously preassemble in the target membrane in order to form a Qabc acceptor complex for an R-SNARE from the other fusion partner [49]–[50]. Our studies and the results from ER-Golgi transport and regulated exocytosis suggest, however, that in intact membranes, SNARE complex assembly occurs via a Rab- and tether-stabilized QbcR subcomplex. These discrepancies probably reflect the absence of constraints for SNARE assembly in the liposome systems, constraints that are imposed in the intact membrane system by Rab-GTPases and tether factors. Purified single SNAREs are largely unstructured [49]–[51]. Their rearrangement into a coiled-coil conformation can be kinetically limiting for trans-SNARE complex formation and fusion of proteoliposomes. In the absence of other factors, Qa, Qb, and Qc helices can form stable subcomplexes that can subsequently integrate an R helix. Therefore, co-reconstitution of a SNARE combination that allows slow pre-structuring of a cis-SNARE subcomplex in one membrane (e.g., during production and purification of the proteoliposomes) can render the integration of the remaining SNARE helix from the other fusion partner much faster and strongly enhance the rate of fusion. This explains why in some studies, proteoliposomes in which Qa, Qb, and Qc SNAREs were co-reconstituted into one vesicle and the R-SNARE in the other yielded higher fusion activities [4]. Depending upon the experimental condition chosen, however, other distributions of the four SNAREs over the two membranes can become fusogenic [11]. This important result illustrates that SNAREs can assemble into trans complexes in various topologies. In a physiological membrane, by contrast, SNAREs are associated with Rab and tether proteins that may restrict the assembly pathway. These control factors recently were shown to further stimulate fusion of SNARE-containing liposomes [52],[53], but how they influence the assembly pathway and topology of trans-SNARE complexes in these reactions has not yet been resolved.
Our results suggest that Sec18/NSF selectively removes the Qa SNARE from vacuolar cis-SNARE complexes, generating a QbcR subcomplex that is stabilized by the Rab-GTPase Ypt7 and the associated HOPS complex. This QbcR subcomplex serves as a template for integrating a Qa SNARE from the other fusion partner.
Two possibilities of HOPS mediated cis-SNARE stabilization are conceivable. Either HOPS stabilizes a partially zippered up QbcR complex or single SNAREs are separately coordinated on HOPS subunits. Although we cannot say how exactly HOPS stabilizes an intermediate QbcR complex, it is evident that Ypt7 in the GDP-bound state does not scaffold a QbcR cis complex, suggesting that active control via Ypt7 occurs. The functional relevance of this topology is supported by the observation that vacuoles retain their fusion competence only if inactivating SNARE mutations are distributed over the two fusion partners in combinations permitting formation of a QbcR-Qa trans-complex. We noted that adding an excess of the Qc-SNARE Vam7 permits the production of large quantities of Qabc-R complexes also in the vacuole system (C. Peters and A. Mayer, unpublished). Excess Vam7-Qc provides a fusion activity with reduced sensitivity to the Ypt7 inhibitor GDI [54], suggesting that excess Vam7-Qc partially compensates for the lack of Ypt7 activity. Trans-SNARE complexes accumulated 3–5 times higher amounts than normal, but now mainly in a Qabc-R topology (unpublished data). This underscores the potential for forming Qabc-R trans-complexes—consistent with the liposome studies that used these combinations—but demonstrates that active Ypt7 and HOPS channel trans-complex assembly mainly into the QbcR-Qa arrangement by restricting complete cis-SNARE disassembly.
Our approach and experimental system impose two limitations that raise caveats for this interpretation. First, only a small percentage of the SNAREs enter trans-SNARE complexes, which renders it impossible to firmly exclude the existence of trans-SNARE complexes in topologies other than QbcR-Qa. Second, the persistence of cis-SNARE associations and the preference for the QbcR-Qa trans-associations are not absolute. Our interpretation of the interactions in the trans-complex, therefore, relies mainly on the observed trends and the increase in the abundance of these trans complexes. However, the significance of the observed QbcR-Qa interaction is supported by the functional effects of SNARE mutations distributed over the two fusion partners in combinations that allow one to distinguish between QbcR-Qa and Qabc-R trans-complexes. We consider this correlation to be a strong argument for the validity and relevance of the observed interactions.
Could the Rab-controlled assembly of trans-SNARE complexes in a QbcR-Qa topology also apply to other SNARE-dependent fusion systems? Several published observations suggest that this may be the case. For fusion of ER-derived COPII vesicles with the Golgi, the use of different combinations of temperature-sensitive SNARE proteins showed that the SNAREs Bos1-Qb and Bet1-Qc act on the vesicles, and only the SNARE Sed5-Qa acts on the acceptor membrane [27],[55],[56]. Chemical depletion of the vesicular Sec22-R pool reduced fusion [57]. Bos1-Qb interacts with Bet1-Qc and also with the R-SNARE Sec22, and this latter interaction depends on the Rab-protein Ypt1 [58]–[60]. It is unknown whether these interactions represent pre- or post-fusion states, whether they change in the course of fusion, or whether they might have occurred after solubilization of the membranes. However, these findings fit seamlessly with the sequence of events that we resolved in vacuole fusion.
Furthermore, functional studies on regulated exocytosis in cracked PC12 cells favored a QbcR receptor complex as a post-priming intermediate rather than a Qabc complex. Scheller and colleagues showed that SNARE priming sensitizes exocytosis selectively to competition by soluble syntaxin (Qa) peptides, but not to VAMP2 (R) peptides [28]. Analyses of neurotoxin sensitivity at different stages of exocytosis further supported the hypothesis that priming might create a SNAP25/VAMP2 complex (QbcR) that subsequently incorporates syntaxin (Qa) from the plasma membrane. Although cis- and trans-SNARE complex formations were not directly assayed in these studies, their results are compatible with our observations.
While functional studies on exocytosis in PC12 cells are consistent with a Qa binding site being created by SNAP-25 (providing a Qb and a Qc helix) and VAMP2 (R), the major pool of SNAP-25 is located on the plasma membrane, whereas VAMP2 is mainly on the vesicles [61],[62]. In order to resolve this contradiction we can invoke two scenarios for formation of a VAMP2/SNAP-25 receptor complex [28]. One scenario is that SNAP-25 from the plasma membrane would first assemble with VAMP2 from the vesicle, creating a QbcR complex that connects the two membranes. Alternatively, SNAP-25 on the vesicle could associate with VAMP2, creating a QbcR cis complex. This latter scenario is supported by several studies that detected microscopically localized SNAP-25 on various types of secretory vesicles [23],[63] and provided convincing biochemical evidence for the presence of SNAP-25 in SNARE complexes on highly purified synaptic vesicles [64],[65]. Despite all these observations, we cannot rule out the existence of an alternative SNARE topology mediating neurotransmitter release. However, the combined evidence strongly favors the view that in physiological membranes, the Rab-GTPase and its associated tether factors bias trans-SNARE assembly by stabilizing a QbcR receptor complex that integrates Qa SNARE in a second step.
BJ3505 strains carrying tagged SNAREs were grown in YPD at 30°C at 225 rpm to OD600 = 2 and harvested (3 min, 5,000× g). Vacuoles were isolated as described [47], but all solutions contained 2 mM DTT and cell walls were hydrolyzed by lyticase [66], recombinantly expressed in E. coli RSB805 (provided Dr. Randy Schekman, Berkeley), and prepared from a periplasmic supernatant. Harvested cells were resuspended in reduction buffer (30 mM Tris/Cl pH 8.9, 10 mM DTT) and incubated for 5 min at 30°C. After harvesting as described above cells were resuspended in 15 ml digestion buffer (600 mM sorbitol, 50 mM K-phosphate pH 7.5 in YP medium with 0.2% glucose and 0.1 mg/ml lyticase preparation). After 20 min at 30°C, cells were centrifuged (1 min 5,800 rpm in JLA25.5 rotor). The spheroplasts were resuspended in 2.5 ml 15% Ficoll-400 in PS buffer (10 mM PIPES/KOH pH 6.8, 200 mM sorbitol) and 200 µl DEAE-Dextran (0.4 mg/ml in PS). After 90 s of incubation at 30°C, the cells were transferred to SW41 tubes and overlaid with steps of 8%, 4%, and 0% Ficoll-400 in PS. Cells were centrifuged for 60–75 min at 2°C and 30,000 rpm in a SW41 rotor. Cytosol was prepared as described [67].
Nyv1 and Vam3 were precipitated from samples containing 1 ml of vacuoles at a concentration of 500 µg/ml. Vacuoles were primed for 5 min in PS buffer with 125 mM KCl, 0.5 mM MnCl2, 1 mM DTT, and ATP-regenerating system. GDI was added at a concentration of 5 µM. Prior to centrifugation, 3 mM EDTA was added and incubated for 15 min at 27°C. Vacuoles were centrifuged for 2 min at 20,000 g and solubilized in PS buffer supplemented with 50 mM KCl, 3 mM EDTA, 0.5% Triton X-100, and 3 mM DTT. After centrifugation (4 min, 20,000 g at 4°C), 15 µg of polyclonal antibodies and 50 µl of a 1∶1 slurry of protein A were added and gently rotated for 1 h at 4°C. The beads were subsequently washed three times with extraction buffer diluted 1∶1 with PS-buffer and subjected to SDS-PAGE and Western blotting. For dilution experiments vacuoles were harvested and adjusted to a density of 100 µg/ml prior to the addition of KCl. After incubation for 5 min at 27°C, 3 mM EDTA was added and further incubated for 15 min at 27°C. Subsequently, vacuoles were directly detergent extracted by adding Triton X-100 to a final concentration of 0.5%. After centrifugation the samples were processed as described above.
Vacuoles were adjusted to a protein concentration of 500 µg/ml. The total volume of one assay was 1 ml containing equal amounts of the two fusion partners in PS buffer with 125 mM KCl, 0.5 mM MnCl2, and 1 mM DTT. Mixed vacuoles were incubated for 5 min at 27°C in the absence of ATP. The fusion reaction was started by adding ATP-regenerating system (0.25 mg/ml creatine kinase, 20 mM creatine phosphate, 500 µM ATP, 500 µM MgCl2). After 5 min at 27°C, the vacuoles were cooled down to 7°C and incubated further for 30 min at this temperature. Thereafter, 3 mM EDTA was added and vacuoles were centrifuged for 2 min at 4°C at 20,000 g. The pellet was resuspended in 1.5 ml solubilization buffer (0.5% Triton, 50 mM KCl, 3 mM EDTA, 3 mM DTT in PS). After centrifugation for 4 min at 4°C (20,000 g), the supernatant was incubated with 30 µl Protein G-beads (Roche) and 15 µg HA-antibodies (Covance, mouse monoclonal) for 1 h at 4°C with gentle shaking. The Protein G-beads were washed three times with 50 mM KCl, 0.25% Triton, 3 mM DTT, and 3 mM EDTA in PS buffer, and incubated for 5 min at 60°C in 2– concentrated reducing SDS sample buffer.
DKY6281 and BJ3505 vacuoles were adjusted to a protein concentration of 500 µg/ml and incubated in a volume of 30 µl PS buffer (10 mM PIPES/KOH pH 6.8, 200 mM sorbitol) with 125 mM KCl, 0.5 mM MnCl2, 1 mM DTT. Inhibitors were added before starting the fusion by addition of the ATP-regenerating system (0.25 mg/ml creatine kinase, 20 mM creatine phosphate, 500 µM ATP, 500 µM MgCl2). After 60 min at 27°C, or on ice, 1 ml of PS buffer was added, vacuoles were centrifuged (2 min, 20,000× g, 4°C) and resuspended in 500 µl developing buffer (10 mM MgCl2, 0.2% TX-100, 250 mM TrisHCl pH 8.9, 1 mM p-nitrophenylphosphate). After 5 min at 27°C, the reactions were stopped with 500 µl 1 M Glycin pH 11.5 and the OD was measured at 400 nm.
For experiments implicating temperature-sensitive mutants, cells were grown in YPD at 25° and 225 rpm. Cells were harvested at OD600 = 2 and vacuoles were prepared essentially as described above, with the following modifications: Cells were incubated for 7.5 min with reduction buffer at 25°C. After centrifugation, pep4 cells were spheroplasted (25 min at 25°C or 40°C) in spheroplasting buffer containing 2 mM DTT. For pho8Δ cells, the spheroplasting step was performed for 12.5 min at 40°C or for 25 min at 25°C. If spheroplasting was performed at 40°C, the amount of lyticase was reduced by 50% compared to spheroplasting at 25°C. All further steps were as described above in solutions containing 2 mM DTT. Vacuoles from BJ3505 expressing a vam3tsf allele contained 50% less of the reporter enzyme pro-alkaline phosphatase. This was taken into account and corrected in calculating the fusion activities.
|
10.1371/journal.ppat.1000291 | HCV Induces Oxidative and ER Stress, and Sensitizes Infected Cells to Apoptosis in SCID/Alb-uPA Mice | Hepatitis C virus (HCV) is a blood-borne pathogen and a major cause of liver disease worldwide. Gene expression profiling was used to characterize the transcriptional response to HCV H77c infection. Evidence is presented for activation of innate antiviral signaling pathways as well as induction of lipid metabolism genes, which may contribute to oxidative stress. We also found that infection of chimeric SCID/Alb-uPA mice by HCV led to signs of hepatocyte damage and apoptosis, which in patients plays a role in activation of stellate cells, recruitment of macrophages, and the subsequent development of fibrosis. Infection of chimeric mice with HCV H77c also led an inflammatory response characterized by infiltration of monocytes and macrophages. There was increased apoptosis in HCV-infected human hepatocytes in H77c-infected mice but not in mice inoculated with a replication incompetent H77c mutant. Moreover, TUNEL reactivity was restricted to HCV-infected hepatocytes, but an increase in FAS expression was not. To gain insight into the factors contributing specific apoptosis of HCV infected cells, immunohistological and confocal microscopy using antibodies for key apoptotic mediators was done. We found that the ER chaperone BiP/GRP78 was increased in HCV-infected cells as was activated BAX, but the activator of ER stress–mediated apoptosis CHOP was not. We found that overall levels of NF-κB and BCL-xL were increased by infection; however, within an infected liver, comparison of infected cells to uninfected cells indicated both NF-κB and BCL-xL were decreased in HCV-infected cells. We conclude that HCV contributes to hepatocyte damage and apoptosis by inducing stress and pro-apoptotic BAX while preventing the induction of anti-apoptotic NF-κB and BCL-xL, thus sensitizing hepatocytes to apoptosis.
| Hepatitis C virus is a common cause of liver disease worldwide. The details of how HCV causes liver disease are not well understood. It has been thought that HCV infection does not kill liver cells directly, but indirectly by stimulating the immune system to kill HCV-infected liver cells. In this study we have used a mouse model that supports HCV infection and replication. These mice do not have an adaptive immune system. Despite the lack of an adaptive immune system, we have shown that HCV infection still leads to the death of infected liver cells. This study provides new insight into how HCV damages the liver in chronic HCV carriers.
| Hepatitis C virus (HCV) is a positive strand RNA virus that belongs to the family Flaviviradae. HCV is a blood borne pathogen which is a major cause of liver disease worldwide, with an estimated 200 million people infected. It is estimated that 30% of chronically infected patients eventually develop progressive liver disease including cirrhosis and end stage liver disease [1]. HCV is now the leading indication for liver transplantation in North America [2]. Due the absence of a proofreading activity in the viral RNA polymerase, HCV has a high mutation rate that contributes to the genetic heterogeneity of the virus. Six different genotypes, and at least 52 subtypes have been described [3].
Chronic infection by HCV results in a highly variable disease course, and despite advances in the molecular virology of HCV, the factors involved in hepatocyte injury and the progression of liver disease remain unclear. The complexity of the host response has been examined by transcriptional profiling of liver biopsy samples from both chimpanzees and humans. These studies show that HCV induces genes involved in the interferon response, lipid metabolism, oxidative stress, and chemokines, as well as markers of inflammation [4],[5]. Such studies are hampered by the requirement that infected and uninfected hepatocytes come from different patients and also by the presence of an adaptive immune response in the patients. Gene expression profiles of human hepatocytes from SCID/Alb-uPA chimeric mice infected with HCV patient serum, control these variables [6]. Until recently it has been thought that HCV is a non-cytopathic virus and that hepatocyte damage in chronic HCV infection is due to HCV-specific adaptive immune responses [7],[8]. However, in SCID mice lacking an adaptive immune system, we observed induction of apoptosis in HCV infected mouse livers, similar to that seen in liver biopsies from HCV infected patients [6].
During HCV infection, hepatocyte apoptosis could be induced by immune attack on infected cells or directly by viral infection. It has been shown that hepatocyte damage can lead to apoptosis, which plays a role in the recruitment and activation of stellate cells and macrophages and the subsequent development of fibrosis [9],[10]. HCV infected patients have higher levels of immune related death ligands; TRAIL, TNF-α, FAS, and FASL are all elevated in HCV infected patients [11]–[13]. Increased expression of stimulators of apoptosis in HCV infected patients is tempered by hepatocyte insensitivity to death ligand mediated apoptosis. In hepatocytes death receptor mediated caspase-8 activation is weak, and thus they are inherently resistant to TNF-α and TRAIL killing [14],[15]. Hepatocytes are likely type II cells and can be sensitized to death ligand mediated apoptosis by caspase-8 induction with IFN-α (interferon) or toxins [16]–[18]. In addition, hepatocytes can be sensitized to both TRAIL and TNF-α induced apoptosis by inhibition of NF-κB activity [19]–[21]. Conversely, induction of NF-κB has been shown to inhibit TRAIL, TNF-α, and FAS mediated killing [22]–[24].
The induction of apoptosis directly by HCV remains controversial. Several HCV proteins have been proposed to have both pro- and anti-apoptotic effects [25]–[28]. It has been shown that expression of either the HCV genome or individual HCV structural proteins induces endoplasmic reticulum (ER) stress [27],[29] and the unfolded protein response (UPR), which can lead to apoptosis. However, HCV proteins have also been shown to modulate the UPR [30],[31]. It has also been proposed that HCV infection induces oxidative stress, which can enhance apoptosis [6],[32]. Expression of HCV core induces oxidative stress and expression of antioxidant genes [33],[34]. In addition, HCV patients have more DNA lesions produced by oxidative damage [35]. Oxidative stress leads indirectly from DNA damage to p53 induction, which can lead to activation of BAX and apoptosis [36],[37]. However, there are also reports of inhibition of p53 induced apoptosis by NS5A [38]–[41].
In this study, we used the mouse model for HCV infection in which severe combined immunodeficiency disorder (SCID) mice transgenic for an array of urokinase plasminogen activator (uPA) under control of the albumin (Alb) promoter are transplanted with human hepatocytes and then infected with HCV [6], [42]–[44]. We have previously compared HCV induced gene expression in chimeric mice infected with genotype 1a patient serum to uninfected controls containing human hepatocytes from the same donor [6]. There was evidence of activation of innate antiviral signaling pathways, induction of lipid metabolism genes, as well as signs of hepatocyte damage and an inflammatory response. To further reduce variation from HCV quasispecies present in the mice, in this study we have infected mice with the infectious clone HCV H77c [45],[46]. We confirmed the results of the previous study that used infectious patient serum, and to determine the cause of hepatocyte damage, we examined the expression HCV antigens and key proteins involved in the stress response and apoptosis using immunohistochemical and fluorescent confocal microscopy. We found that HCV infection correlated with increased levels of the ER chaperone GPR78/BiP, a key regulator of the unfolded protein response. In addition, levels of pro-apoptotic BAX were increased, while anti-apoptotic NF-κB and BCL-xL were decreased in HCV infected cells. Taken together these results indicate that ER stress induced by HCV combined with lower NF-κB and BCL-xL levels sensitizes hepatocytes to apoptosis.
Previous studies indicated HCV infection in chimeric mouse livers was restricted to the human hepatocytes [44]. We confirmed this by performing immunofluorescent confocal microscopy on uninfected and HCV H77c infected mouse livers with antibodies specific for the HCV NS3 protease and human albumin (Figure 1 and S1A–B). Only liver sections infected with HCV H77c stained with HCV specific antibodies, and this was restricted to hepatocytes that also were also stained with antibodies specific for human albumin. Since liver consists of a mix a hepatocytes and adventitial cells, and albumin only stains hepatocytes, there was the possibility that some human cells other than hepatocytes also colonized the mouse liver. Therefore we wished to examine whether any of the adventitial cells were also human. We performed in situ hybridization using probes specific for human Alu repeats on chimeric mouse livers (Figure 2). Only the hepatocytes were stained, indicating all of the adventitial cells in chimeric mouse livers were of mouse origin. This, and the elimination of a small percentage of mouse sequences that cross-hybridized to the human arrays, ensured that the transcriptional profiling reflects only the processes occurring in human hepatocytes.
Transcriptional profiling was performed on mRNA samples isolated from three HCV-infected animals and from uninfected controls. All animals contained hepatocytes from a single donor. The serum HCV titers, liver viral loads, and the length of time infected are given in Table 1. The experiments included three liver samples from an animal infected with H77c (+) serum (990), two samples from an animal inoculated intrahepatically with H77c RNA (975), and a single experiment with liver tissue from an animal inoculated intrahepatically with H77c RNA containing a mutation in the active site of the NS5B polymerase (986). Four samples from three separate animals were pooled to serve as the uninfected control. Because each pair of mice contained hepatocytes from the same donor, changes in gene expression should mainly be induced by the HCV infection and be independent of host variation. Consistent with previous studies, the effect on host gene expression by HCV infection was not extreme, 766 genes showed a 2-fold or higher change in expression (P value≤0.05) in at least one experiment (Figure 3). The grouping of experiments by the clustering algorithm suggested that the effect on host gene expression was very similar among individual pieces of liver from the same animal. Importantly, the global gene expression profiles in the animals infected with H77c (+) serum (990) and H77c RNA (975) were also very similar. This suggests that the source of HCV inoculum does not significantly impact the host transcriptional response to infection. Interestingly, the animal inoculated with H77c RNA encoding an inactive NS5B polymerase also showed a similar host response. While it was expected that the mouse inoculated with the replication defective HCV RNA might show activation of dsRNA and RIG-I signaling pathways similar to replicating virus [47],[48], we expected substantial differences in the overall host response 47 days after RNA administration. This mouse showed no detectable HCV RNA in the serum at the time of sacrifice and no HCV RNA was detected in the sample used for microarray analysis.
Infection with HCV H77c activated innate antiviral signaling pathways, as indicated by the induction of interferon-stimulated genes (ISGs) (Figure 4A). In general, the induction was similar among all three infected animals. However, there does seem to be a slightly higher induction of ISGs in the animal (975) inoculated intrahepatically with wild-type H77c RNA relative to the animal (990) inoculated with H77c (+) serum, which is likely due to the high level of naked RNA injected directly into the liver of this animal (100 µg, 2×1013 copies). This increased response relative the animal inoculated with serum (990) was not observed in the animal (986) injected with H77c RNA containing inactive NS5B, indicating that part of the increased response in animal 975 might be due to replication of the inoculated HCV RNA. In our previous study of mice infected with HCV-positive patient sera the magnitude of induction of ISGs varies among mice containing hepatocytes from different donors. Comparison of the gene expression data from HCV H77c-infected mice with that from the initial study indicate the induction of ISGs in the H77c-infected mice is relatively weak. Consistent with what was observed in animals with a weak IFN response in the initial study, regulation of numerous genes associated with lipid metabolism were observed in the HCV H77c-infected mice (Figure 4B). These included genes involved in cholesterol and fatty acid biosynthesis, β-oxidation and peroxisome proliferation. There did not appear to be any significant differences due to inoculum source. Interestingly, the animal that received replication incompetent H77c RNA also showed some regulation of these genes, although at a lower magnitude. The fact that this mouse shows any changes at all, in the absence of viral replication, may be because injection of viral RNA from a positive-sense RNA virus likely results in the synthesis of viral proteins.
Consistent with the up regulation of genes involved in oxidative stress seen in this and previous expression array studies, histological analysis revealed signs of hepatocyte damage in the human hepatocytes of HCV infected chimeric mice (Table 2). Steatosis was apparent in the majority of the human hepatocytes regardless of infection. However, there were significant differences in the histology between the animals inoculated with HCV RNA and naive animals. Increased hepatocyte ballooning and lobular inflammation were associated with HCV-infection. Staining of sections with the antibody F4-80 (anti CD68) revealed that the lobular inflammation was due to infiltration of monocytes and macrophages (not shown). A particularly intriguing observation was the presence of apoptotic hepatocytes in HCV-infected animals originally detected as caspase-3 activation and quantitation of apoptotic bodies [6]. We also observed apoptosis in H77c-infected mice by TUNNEL assay (Figure 5D–F). Apoptosis was absent in the animal inoculated with the replication defective H77c-AAA mutant. This suggests that apoptosis observed in HCV-infected mice is dependent upon active HCV replication. The expression of genes associated with cell death was analyzed to gain further insight into possible mechanisms of apoptosis. While there was regulation of cell-death related genes in HCV H77c infected animals, the number of genes affected is small (data not shown). This is perhaps not surprising given the low percentage of hepatocytes that are actually undergoing apoptosis. Quantitation of the TUNEL data in Table 2 (average 723 cells/field) revealed on average 5% of cells undergoing apoptosis in infected mice.
To further investigate the mechanism of increased apoptosis associated with HCV infection, we examined FAS expression on liver sections from infected and uninfected mice and compared this to liver sections subjected to TUNEL analysis. Similar to what has been seen in mice infected with patient serum and in patient biopsies [6], there was increased FAS staining in infected compared to donor matched uninfected mice (Figure 5A–C). As can be seen in Figure 5 and at higher magnification in Figure S2, there was strong FAS reactivity in a majority of the human cells in infected mice, however TUNEL positive nuclei were seen only in a small proportion of human cells (Figure 5D–F). Interestingly, although staining was not as intense, there was an increase in FAS staining in the mouse inoculated with the H77c-AAA mutant, without a correlative increase in the TUNEL positive nuclei (Table 2). This suggests that increased FAS expression is not the only factor required for induction of apoptosis and TUNEL reactivity. We next investigated the correlation between HCV infection, and either FAS expression, or TUNEL reactivity. When we stained liver sections with FAS- and HCV-specific antibodies (Figure 6A), we found that expression of FAS does not depend on the presence of HCV in the cells, but is a host reaction to infection of neighbouring cells. However, we cannot rule out the possibility that only a portion of cells in the liver express enough HCV antigen to be detected using this antibody. When we subjected liver sections to a fluorescent TUNEL assay and then stained them with HCV specific antibodies, we found that all of the TUNEL positive cells in areas populated by human hepatocytes also stained with HCV specific antibodies (Figure 6B). The exception was that some murine Kupffer cells which also contained multiple TUNEL positive nuclei. Thus, HCV replication seems to be required for hepatocyte apoptosis. This is unlike FAS expression, which could be induced by HCV replication in neighbouring cells.
To further investigate the relationship between HCV infected cells and apoptosis we performed immunohistochemistry and fluorescent confocal microscopy using antibodies for key proteins involved in apoptosis. It has been proposed that HCV induces oxidative stress [6],[32]. Oxidative stress can lead to the induction of p53 and BAX, both of which can translocate to the mitochondria and induce apoptosis [36],[37]. We performed immunohistological staining for p53 on infected and uninfected liver sections and found no evidence for p53 induction in either the cytoplasm or nucleus or its translocation to the mitochondria except in one infected animal where very few cells appeared to have up-regulated p53 (data not shown).
Expression of HCV structural proteins has been shown to induce ER stress [27],[49], which can induce the oligomerization and translocation of BAX/BAK to the mitochondria. To examine the role of ER stress in HCV induced apoptosis, we compared the expression of the ER chaperone GRP78 (BiP) in uninfected and H77c infected mice by immunohistochemistry (Table 3 and Figure 7A–C) and by fluorescent confocal microscopy (Figure 7D, E). We found higher levels of BiP in infected mice, and that BiP expression correlated with HCV infection. Additionally, BiP and HCV seemed to co-localize, consistent with replication of HCV on the ER. The ER chaperone BiP is a key sensor in the unfolded protein response (UPR); it maintains ER membrane signal proteins in inactive states. It has been shown that BAX and BAK interact directly with one of these membrane signal proteins, IRE1, and are essential for IRE1 activation [50]. Extensive or prolonged ER stress initiates apoptosis through activation of BAX/BAK. We therefore examined the expression of BAX in liver sections (Table 3 and Figure 8) and found BAX was also overexpressed in HCV infected livers. BAX is normally diffusely expressed throughout the cell, however when activated it translocates to the mitochondria and appears as a granular staining pattern. We found that both patterns of staining were elevated in HCV infected livers (Figure 8A–D). Figure 8D shows both the intense granular staining pattern as indicated by the black arrows and the less intense cytoplasmic staining indicated by the red arrows. The large granular staining pattern correlated with HCV infection (Figure 8F). It is worth noting that not all cells that stained positive for HCV also showed activated BAX. The number of cells staining for activated BAX approximately correlated with the number of TUNEL positive nuclei. The elevation of both BiP and BAX in the absence of increased levels of p53 suggests that ER rather than oxidative stress leads to BAX activation, however additional mechanisms of BAX activation by oxidative stress cannot be eliminated.
In cells under ER stress, BiP preferentially binds to malfolded proteins, releasing IRE1, PERK, and ATF6, activating downstream effectors which induce transcription of alarm and adaptation genes including BiP itself and GADD153 (CHOP) [51]. When the UPR is overwhelmed, apoptosis is induced by a number of molecules including CHOP, which translocates to the nucleus and blocks transcription of BCL-2, [52] an inhibitor of BAX/BAK. To examine whether the ER stress found in infected livers overwhelmed the unfolded protein response, indicated by translocation of CHOP to the nucleus, we performed immunofluorescent confocal microscopy with anti-CHOP and anti-HCV antibodies (Figure 9). Because we found very few nuclei that stained positively for CHOP and these did not correlate with the staining by HCV specific antibodies, in Figure 9 both panels are from infected mice; panel A shows a predominately infected area, and panel B shows a predominantly uninfected area with only a few infected cells. CHOP can be elevated in both infected and uninfected cells indicating that HCV infection does not overwhelm the UPR. This may explain why we do not see expression of ER stress genes in the microarray analysis.
In addition to pro-apoptotic proteins, we also examined key inhibitors of apoptosis. Both CHOP and NF-κB are both activated by ER stress [53],[54], but have opposing effects; CHOP is pro-apoptotic while NF-κB is anti-apoptotic. NF-κB is activated in response to a myriad of other stimuli, at least one of which is inhibited by HCV [55]. BCL-xL inhibits the apoptosis induced by BAX, and its transcription is activated by NF-κB [56]. Overall, when HCV infected livers were compared with uninfected livers, the levels of both NF-κB and BCL-xL appeared to be elevated in the infected liver, consistent with expression analysis, which indicated that NF-κB levels are elevated by HCV infection. However, when HCV infected cells are compared to uninfected cells within HCV infected livers, we found that total levels of NF-κB p65 expression was lower in HCV infected cells than in surrounding uninfected cells (Figure 10A–B and E and S4). Quantitation of total p65 fluorescence from uninfected and infected cells in 6 fields from infected livers revealed that p65 levels in infected cells were approximately half that in uninfected cells. The average fluorescence of uninfected cells in a field was arbitrarily set to 1. Consistent with reduced expression of NF-κB, the expression of BCL-xL was also lower in cells that stained with HCV specific antibodies (Figure 10C–D and F). Quantitation of total BCL-xL levels in infected livers also revealed that levels of BCL-xL in infected cells were approximately half of that in infected cells.
The course of HCV pathophysiology is extremely variable, as a result of complex interactions between viral variants and the host's innate and adaptive immune systems. We have used SCID/Alb-uPA mice that have chimeric human and mouse livers to examine the processes that occur in hepatocytes in response to HCV infection. Previous studies have shown that the transcriptional response HCV infection in mice is similar to that of humans and chimpanzees, with the exception of immune cell markers, which are absent in SCID mice. To further simplify the host response to infection, we infected mice with RNA from the HCV clone H77c [45],[46]. We found the same induction of interferon response genes and changes in expression of genes involved in lipid metabolism seen in earlier studies. This has been postulated to lead to the generation of oxidative stress [6],[57], and it has been shown that both oxidative stress and ER stress can lead to apoptosis [27],[29],[58],[59]. As well, apoptosis is one of the factors in the induction of fibrosis, which can culminate in cirrhosis [9],[10],[60],[61]. Interestingly, in the absence of an adaptive immune system, there was evidence for the induction of apoptosis in HCV infected mice. This was restricted to HCV infected cells despite increased FAS expression on both infected and uninfected human hepatocytes of infected animals. This generalized FAS expression may be a consequence of the interferon response occurring throughout the liver [6] since FAS/FASL are among those mediators of apoptosis that are also interferon response genes [62].
We therefore examined processes leading to apoptosis that we thought were likely to be affected by HCV in infected cells. Oxidative stress generated by HCV induced lipid metabolism, and ER stress generated by HCV replication and protein translation in and on the ER were both potential candidates. Hepatocytes are type II cells that contain low levels of caspase 8 and therefore require activation of the mitochondrial apoptosis amplification pathway to initiate apoptosis. This can be blocked by over expression of BCL-2 or BCL-xL. Mitochondria seem to be the site where the antiviral interferon response and apoptotic signals are integrated; recently it has been shown that the mitochondrial signaling molecule interferon promoter stimulating factor-1 (IPS-1) is cleaved during apoptosis and cleavage can be blocked by overexpression of BCL-xL [63]. In addition, both the response to oxidative stress and the response to ER stress converge at the mitochondrion; p53 activated by oxidative stress stimulates the oligomerization of BAX [36], and BiP responding to ER stress releases IRE-1 to which BAX is bound. BAX has been shown to be required for both IRE-1 activation and for apoptosis initiated by ER stress [50],[51]. Since oxidative stress can lead to p53 induction, Bax activation and apoptosis [36],[37], we examined p53 localization and levels and found that they were not affected by HCV infection. This may be due to NS5A mediated inhibition of the mitochondrial translocation, apoptosis inducing, and DNA binding activities of p53 [38]–[41]. ER stress can lead to induction of the UPR, and activation of BiP, CHOP, BAX and apoptosis. Consistent with the generation of ER stress by HCV we found that induction of the ER chaperone BiP and pro-apoptotic BAX correlated with HCV expression, but there was little translocation of CHOP/GADD153 to the nucleus, which indicated that the UPR was not overwhelmed.
In hepatocytes, it appears that NF-κB is one of the key determinants of whether apoptosis is induced in response to death ligands [19]–[24],[64]. In evading of the interferon response, HCV inhibits the activation of NF-κB; inhibition of the TLR-3 and RIG I pathways by cleavage of TRIF and IPS-1/MAVS/VISA/Cardif by the HCV NS3/4A protease, inhibits NF-κB and IRF-3 phosphorylation preventing nuclear translocation in response to RIG-1 activation by viral RNA [65]–[68]. In addition, there are a number of other reports that HCV modulates NF-κB activity [69]–[72]. Consistent with the reports of inhibition of NF-κB we found that total levels of NF-κB p65 were lower in HCV infected cells. Furthermore consistent with the transcriptional regulation of BCL-xL by NF-κB, we found that total levels of BCL-xL were lower in HCV infected cells. In conclusion, we propose a model (Figure 11) where HCV induces both ER stress and oxidative stress in infected cells, and activates pro-apoptotic Bax while it prevents induction of anti-apoptotic BCL-xL thus sensitizing HCV infected cells to apoptosis which may be mediated by death receptors and ligands, for example FAS and TRAIL (TNFSF10-Figure 4B). A combination of induction of pro-inflammatory chemokines (Figure 4A) and cross talk between human and mouse chemokines and their receptors may lead to a situation similar to that in patients; inflammation, which in turn stimulates release of pro-inflammatory cytokines and effector molecules such as TNF-α and FasL (which in these mice may be released by macrophages and NK cells), creating the circle of hepatocyte damage and repair that is a hallmark of HCV infection.
All mice were housed VAF and treated according to Canadian Council on Animal Care guidelines. Experimental approval came from the University of Alberta Animal Welfare Committee, and human hepatocytes were obtained following informed consent of all donors with ethics approval from the University of Alberta Faculty of Medicine Research Ethics Board. Animals were transplanted with freshly isolated human hepatocytes [42],[44],[73]. Eight weeks after transplantation mice with human α-1antitrypsin (hAAT) levels [42] greater than 100 µg/mL were injected intrahepatically (ih), with 50 µg of in vitro transcribed H77c RNA [45] into each of 2 red liver nodules (presumed to be human hepatocytes). As a negative control, mice were injected ih with non-replicative H77c RNA in which NS5B polymerase active site residues GDD (amino acids 2736–2738) have been changed to AAA (H77c-AAA). Passage of H77c virus was done by ih inoculation of naive mice with 50 µL of serum obtained from mice infected by H77c RNA. One mouse was infected by ih inoculation with patient serum for histochemical comparison. Serum samples were taken at various time points after inoculation and HCV RNA was quantified. Animals were infected for 25 or 47 days and dissection of mouse livers, isolation of RNA, genomic DNA, and ratio of human to mouse cells in each sample was performed as previously described [43]. The serum HCV titers, liver viral loads and the length of time infected are given in Table 1. The plasmid for in vitro transcription was pCV H77c and was a gift from Dr. Jens Bukh.
The purity of human hepatocytes was greater than 70% in all samples used for microarrays. Microarray format, protocols for probe labeling, and array hybridization are described at http://expression.microslu.washington.edu. Briefly, a single experiment comparing two mRNA samples was done with four replicate Human 1A (V2) 22K oligonucleotide expression arrays (Agilent Technologies) using the dye label reverse technique. This allows for the calculation of mean ratios between expression levels of each gene in the analyzed sample pair, standard deviation and P values for each experiment. Spot quantitation, normalization and application of a platform-specific error model was performed using Agilent's Feature Extractor software and all data was then entered into a custom-designed database, Expression Array Manager, and then uploaded into Rosetta Resolver System 4.0.1.0.10 (Rosetta Biosoftware, Kirkland, WA) and Spotfire Decision Suite 7.1.1 (Spotfire, Somerville, MA). Data normalization and the Resolver Error Model are described on the website http://expression.microslu.washington.edu. This website is also used to publish all primary data in accordance with the proposed MIAME standards. Selection of genes for data analysis was based on a greater than 95% probability of being differentially expressed (P≤0.05) and a fold change of 2 or greater. The resultant false positive discovery rate was estimated to be less than 0.1% (Walters, unpublished data). We have previously assessed the degree of cross hybridization in chimeric samples and eliminated the small percentage of genes that did cross react from subsequent analysis [43].
In situ hybridization using FITC labeled Alu DNA probes (InnoGenex, San Ramon, CA, USA) was performed according to the manufacturer's specifications, and developed using the supersensitive polymer HRP-ISH system (BioGenex). TUNEL was performed using the Apoptag Plus Peroxidase In Situ Apoptosis Detection kit (Chemicon International, Temecula, CA, USA) according to the manufacturer's specifications. The number of Tunel positive nuclei is an average of 15 fields at 200× magnification.
Haematoxylin and eosin, reticulin, Mason's trichrome, and periodic acid/Shiffs staining were performed according to standard procedures [74]. The sections were graded for inflammatory activity and staged for fibrosis according to the modified Batts and Ludwig scoring system [75]. The degree of fatty change was scored as 0 (<5%), 1 (6–33%), 2 (34–66%) or 3 (>66%). Hepatocyte ballooning and macrophages were scored on a scale of 0–4 where 0 is none and 4 is many. The lobular apoptotic body count is an average of 5 fields counted at 100× magnification.
Immunohistochemical and immunofluorescent analysis was performed on 4 µm formaldehyde fixed paraffin embedded sections that were deparaffinized by incubation in xylene for 5 min, followed by sequential rehydration by incubating twice for 3 min in each of 100%, 95%, and 70% ethanol, followed by a 5 min incubation in distilled water. Antigen retrieval was then performed by boiling in pH 6.0 10 mM citrate buffer for 15 min followed by cooling for an additional 15 min.
For immunohistochemical staining with rabbit anti-FAS antibodies (1∶50, Santa Cruz), or rabbit anti-BAX (1∶50, Cell Signaling Technologies), or purified rabbit IgG isotype control, slides were blocked in normal goat serum, washed, incubated with the primary antibodies, washed, incubated with 3% peroxide, and incubated with secondary goat anti rabbit poly-HRP antibodies (Dako Cytomation). The peroxidase was developed using the DAB Plus liquid substrate chromogen system (Dako Cytomation). For staining with goat anti-GRP78/BiP antibodies (1∶50 Santa Cruz) or purified goat IgG isotype control, slides were blocked with normal donkey serum, incubated with primary antibody, endogenous biotin was blocked using the avidin/biotin blocking kit (Vector laboratories), and the signal was amplified using the ABC method (Vector laboratories). The peroxidase was developed as before. Caspase staining was performed as previously described [6]. FAS staining was scored semi-quantitatively where 0 is no staining, 1 (1–25%), 2 (26–50%), 3 (51–75%) and 4 (76–100%). Caspase staining was scored semi-quantitatively as follows: 0 = none, 1 = focal weak, 2 = diffuse weak; 3 = diffuse weak and focal strong; 4 = diffuse strong. GRP78/BiP staining was scored semi-quantitatively where 0 is no staining, 1 (1–15%), 2 (16–30%), and 3 (31–50%). The intensity of the stain was also scored on a scale of 0–3, where 0 is no staining and 3 is intense staining. Since activated activated BAX has a distinct punctate staining that can be easily distinguished form inactive BAX, active and inactive BAX was scored on separate semi-quantitative scales. Inactive Bax was scored in the same manner as BiP, and the scale for activated Bax was 0 is no staining, 1(1–5%), 2 (5–10%), and 3 (10–15%).
For immunofluorescent confocal microscopy, the slides were deparaffinized, the antigens retrieved as before, and blocked as before. Additional blocking using mouse IgG (0.1 mg/ml) for 1 hour, followed by incubation with goat anti mouse-IgG (1 mg/ml) overnight at 4°C was done prior to incubation with mouse anti HCV NS3/4 diluted 1∶50 (TORDJI-22, Abcam), or its isotype control, and one of rabbit anti-FAS, BAX, GADD (Santa Cruz), NF-κB p65 (C-20 Santa Cruz), BCL-xL (Cell Signaling Technologies), or rabbit IgG all diluted 1∶50, or rabbit anti human Albumin diluted 1∶1000 (Dako Cytomation). Slides were blocked with 3% peroxide prior to incubation with goat anti mouse poly HRP and goat anti rabbit Alexa 488 (diluted 1∶100, Molecular Probes, Eugene, OR, USA). The peroxidase was developed using the TSA Plus fluorescence system with tyramide-tetramethyl red (Perkin Elmer). Mounting media (Vectastain-Vector laboratories) contained 1 µg/ml 4,6-diamidino-2-phenylindole (DAPI). For BiP/GRP78, the staining procedure was essentially the same, except slides were blocked with normal donkey serum and avidin/biotin block (Vector laboratories), the primary antibodies were goat anti-GRP78/Bip with the mouse anti-HCV, and the secondary antibodies were donkey anti-goat alexa 488 (Molecular probes), and donkey anti-mouse biotin, followed by avidin-HRP (Vector laboratories). The peroxidase was developed as before.
For co-localization of HCV antigens and TUNEL reactivity, the In Situ Cell Death Detection kit (fluorescein) (Roche) was used, according to the manufacturer's specifications. The incubation with terminal deoxynucleotidyl transferase was carried out prior to incubation with the primary TORDJI-22 antibody. All subsequent steps were carried out as before. Nuclei were stained using DAPI. Confocal microscopy was carried out using a Zeiss scanning LSM510 microscope with the 351 nm, 488, and 543 nm excitation lines, and digital images were collected with a 1 µm optical slice.
The accession numbers for the genes/proteins discussed in this manuscript are the following: HCV-H77c AF011751, TNF-α X20910, FAS M67454, FASL U11821, BiP/GRP78 NM_005347, p53 AF307851, CHOP/GADD153 BC003637, BAX NM138763, BCL-Xl Z23115, BCL-2 M14745, NF-κB p65 Z22751, Caspase-8 U60520, Caspase-3 BC016926, CD68 S57235, RIG-I AF038963, IPS-1-Q7Z434, TLR-3 U88879, TRIF AB086380, IL28RA AY129153, CMKOR1 BC008459, IFITM1 J04164, HLA-DRB5 NM_002125, IFIT1 M24594, IFIT2 M14660, B2M AB021288, HLA-A D3219, HLA-B M15470, HLA-F AY253269, HLA-G NM_002127, GBP1 BC002666, BIRC4BP X99699, CXCL11 U66096, CXCL10 X02530, CXCL9 X72755, PSMB9 NM_002800, OAS3 AF063613, OAS1 X04371, OASL AF063611, STAT1 NM_007315, G1P3 BC15603, G1P2 BC009507, IFI44 D28915, IFI27 X67325, ANGPTL4 AF202636, NR4A1 L13740, BDKRB2 S56772, EPO X02157, PPARGC1A AF106698, PCK1 NM_002591, ARG2 D86724, APOA5 AF202889, AVP M25647, CPT1A L39211, MT1A BC029475, FABP5 M94856, RXRA X52773, SREBF1 BC057388, PLIN AB005293, C11orf11 AB014559, APOF L27050, HPX J03048, CYP1A1 BC023019, SAA2 M26152, THRSP Y0809, SCD AF097514, PBP NM_04139, ACSS2 AF263614, GCK AF041014, FDPS J05262, HMGCR NM_000859, CYP51A1 U51685, SQLE D78130, HSD17B6 AF016509, C5 M57729, FTFD1 X69141, RDH16 NM_003708, TNFSF10 U37518, SC4MOL U93162, HMGCS1 NM_002130, IRE1 AF059198, PERK AF110146, ATF6 AB015856, hAAT X01683.
|
10.1371/journal.pcbi.1000383 | A Structure-Based Approach for Detection of Thiol Oxidoreductases and
Their Catalytic Redox-Active Cysteine Residues | Cysteine (Cys) residues often play critical roles in proteins, for example, in
the formation of structural disulfide bonds, metal binding, targeting proteins
to the membranes, and various catalytic functions. However, the structural
determinants for various Cys functions are not clear. Thiol oxidoreductases,
which are enzymes containing catalytic redox-active Cys residues, have been
extensively studied, but even for these proteins there is little understanding
of what distinguishes their catalytic redox Cys from other Cys functions.
Herein, we characterized thiol oxidoreductases at a structural level and
developed an algorithm that can recognize these enzymes by (i) analyzing amino
acid and secondary structure composition of the active site and its similarity
to known active sites containing redox Cys and (ii) calculating accessibility,
active site location, and reactivity of Cys. For proteins with known or modeled
structures, this method can identify proteins with catalytic Cys residues and
distinguish thiol oxidoreductases from the enzymes containing other catalytic
Cys types. Furthermore, by applying this procedure to Saccharomyces
cerevisiae proteins containing conserved Cys, we could identify the
majority of known yeast thiol oxidoreductases. This study provides insights into
the structural properties of catalytic redox-active Cys and should further help
to recognize thiol oxidoreductases in protein sequence and structure
databases.
| Among the 20 amino acids commonly found in proteins, cysteine (Cys) is special in
that it is present more often than other residues in functionally important
locations within proteins. Some of these functions include metal binding,
catalysis, structural stability, and posttranslational modifications.
Identifying these functions in proteins of unknown function is difficult, in
part because it is unclear which features distinguish one Cys function from the
other. Among proteins with functionally important Cys, a large group of proteins
utilizes this residue for redox catalysis. These proteins possess different
folds and are collectively known as thiol oxidoreductases. In this work, we
developed a procedure that allows recognition of these proteins by analyzing
their structures or structural models. The method is based on the analyses of
amino acid and secondary structure composition of Cys environment in proteins,
their similarity to known thiol oxidoreductases, and calculations of Cys
accessibility, reactivity, and location in predicted active sites. The procedure
performed well on a set of test proteins and also selectively recognized thiol
oxidoreductases by analyzing the Saccharomyces cerevisiae
protein set. Thus, this study generated new information about the structural
features of thiol oxidoreductases and may help to recognize these proteins in
protein structure databases.
| Compared to other amino acids in proteins, cysteine (Cys) residues are less frequent,
yet often more conserved and found in functionally important locations.
Protein-based Cys thiols can be divided into several broad categories wherein these
residues (i) are engaged in structural disulfide bonds, (ii) coordinate metals,
(iii) carry out catalysis, (iv) serve as sites of various posttranslational
modification, or (v) are simply dispensable for protein function.
Structural disulfide bonds are typically observed in oxidizing environments such as
periplasm in prokaryotes, and extracellular space and the endoplasmic reticulum (ER)
in eukaryotes. Structural disulfides are formed by designated systems for oxidative
protein folding, for example DsbA and DsbB in bacteria and protein disulfide
isomerase and Ero1 in the eukaryotic ER. In addition, disulfides as stabilizing or
regulatory elements may occur intracellularly. However, there are also situations
when the introduced intraprotein disulfide leads to a decreased protein stability
[1].
Structural stability may also be achieved when Cys residues are linked by metal
ions, such as zinc and iron. In addition, Cys-coordinated metal ions may serve
catalytic functions; for example, when the metal is zinc, copper, nickel, molybdenum
or iron. Metal-coordinating thiols are typically found intracellularly [2],[3], but
may also occur in the extracellular space.
Another important functional category of Cys residues involves catalytic Cys that act
as nucleophiles. This situation occurs, for example, in Cys proteases and tyrosine
phosphatases where Cys does not change redox state during catalysis, and in
thioredoxins and glutaredoxins where Cys undergoes reversible oxidation and
reduction. In the latter case, thiol oxidation may result in the formation of an
intermediate disulfide bond with another protein thiol. In the absence of nearby Cys
residue, thiol oxidation may lead to sulfenic acid (-SOH), sulfinic acid
(-SO2H), S-nitrosothiol (-SNO), or S-glutathionylation (-SSG). In the
majority of these intermediates (with the exception of sulfinic acids), the oxidized
forms of Cys can be reduced by thiol oxidoreductases, such as thioredoxin and
glutaredoxin, by glutathione, or by other protein and low molecular weight
reductants [4],[5]. Even sulfinic acids can be reduced in a select
class of proteins, for example, in peroxiredoxins by a protein known as sulfiredoxin
[6].
Since these oxidized thiol forms are often reversible, they constitute a facile
switch for modulating protein activity and function.
Reversible thiol oxidation has received considerable attention in recent years due to
its ability to regulate proteins, protect them against stress and influence
signaling. For example, sulfenic acid formation is often an intermediate step in
generating disulfides [7]. Recent work has analyzed Cys-SOH formation in a
set of test proteins by examining their functional sites and electrostatic
properties [8]. The authors characterized several features of these
proteins including significant underrepresentation of charged residues and
occurrence of polar uncharged residues in the vicinity of modified Cys.
Nevertheless, at present little is known about the sequence or structural features
that can be employed to predict these proteins in sequence or structure databases.
Much recent work has focused on S-glutathionylation [9], but common features
of these modification sites are also unclear, especially as tools to identify other
glutathionylation sites. Similarly, the determinants of S-nitrosylation are poorly
understood. In the latter case, previously reported features include the presence of
acid-base motifs flanking the modified Cys [10], and, in contrast to the
Cys-SOH-containing proteins, higher frequency of charged residues.
In addition, attempts have been made to examine sites of Cys oxidation at a
structural level. One study evaluated simple structural properties and aimed at
identifying common features of the environment in the vicinity of Cys residues that
undergo reversible redox changes [11]. Parameters that positively correlated with the
occurrence of these Cys included (i) proximity to another Cys residue; (ii) low pKa
(lower than ca. 9.06); and (iii) significant exposure (greater than 1.3
Å2) of the sulfur atom to solvent. Additional parameters
reported were spatial proximity of both proton donor and proton acceptor to the
redox Cys. However, this generic approach combined the analysis of catalytic and
regulatory Cys, which by nature, are different. In addition, with this approach,
almost all protein tyrosine phosphatases, ubiquitin-activating E1-like enzymes,
thymidylate synthases and other enzymes with catalytic non-redox Cys could be
detected, mainly because of their reactive (i.e., low pKa and high exposure)
catalytic Cys.
Although Cys residues often serve roles critical to protein function and regulation,
the presence of a Cys per se by no means implies any of these
features. Analyses of Cys conservation may help identify some catalytic and
functional Cys, but mostly for proteins with already known functions. Nevertheless,
at present, sequence-based methods provide the most straightforward approach to
analyze Cys function. For example, many catalytic redox Cys can be efficiently
identified by searching for Cys-selenocysteine (Sec) pairs in homologous sequences
[12].
This idea stems from the observations that known functions of Sec are limited to
redox functions and that most selenoproteins have homologs in which Sec is replaced
with a conserved Cys (implicating this Cys in redox catalysis).
We hypothesized that identification of Cys function may be assisted by examining
unique features of each Cys function in proteins. In this work, we analyzed general
features of catalytic redox-active Cys via functional profiles of active sites and
structural analyses of reaction centers. When integrated with the tools for enzyme
active site prediction and titration properties of active site residues, this
approach allowed efficient prediction of thiol oxidoreductases in protein structure
databases.
To examine common features of thiol oxidoreductase active sites, we first built a
protein dataset containing previously described thiol oxidoreductases. It
included representative members of protein families with known three-dimensional
structures. We paid particular attention to balance the representation of
thioredoxin fold (which is the most common fold found in thiol oxidoreductases)
and non-thioredoxin fold oxidoreductases.
The resulting dataset consisted of 75 structures in which none of the protein
domains, as defined by SCOP classification, was represented by more than 7
structures. Of these 75 proteins, 40 had thioredoxin-fold, including homologs of
glutathione peroxidase (10 representatives), thioredoxin (7),
glutaredoxin/thioltransferase (13), protein disulfide isomerase (3), DsbA (2),
C-terminal domain of DsbC/DsbG (2), selenoprotein W (2) and ArsC (1). The
non-thioredoxin fold proteins of our redox dataset included 35 proteins
organized in 10 structural folds (thirteen protein families), including
FAD/NAD-dependent reductase (9 representatives), Ohr/OsmC resistance protein
(6), methionine-S-sulfoxide reductase (3), reductase with the protein tyrosine
phosphatase fold (3), GAF-domain methionine sulfoxide reductase fRMsr (2),
FAD-dependent thiol oxidase (2), methionine-R-sulfoxide reductase (2),
antioxidant defense protein AhpD (2), Ero1 (1), and thiol-disulfide interchange
protein DsbD (1). The complete dataset is shown in Table S1.
Our initial analyses suggested that the most challenging problem in
characterizing the general features of redox-active Cys is distinguishing them
from other catalytic Cys residues. Clearly, these two Cys types share
active-site location and high reactivity (e.g., both redox and non-redox Cys are
often strong nucleophiles). To ascertain differences between these protein
classes, we built a separate control dataset of proteins containing catalytic
non-redox Cys (Table S2). This set was composed of 36 proteins (organized in the
form of 17 families/9 folds) including papain-like (9 representatives),
penta-EF-hand (2), ubiquitin carboxyl-terminal hydrolase UCH-13 (1), FMDV leader
protease (1), caspase catalytic domain (3), gingipain R (1), adenain-like (2),
pyrrolidone carboxyl peptidase (1), hedgehog C-terminal autoprocessing domain
(1), high molecular weight phosphotyrosine protein phosphatase (4), dual
specificity phosphatase-like (2), thymidylase synthase/dCMP hydroxymethylase
(2), low molecular weight phosphotyrosine protein phosphatase (1), calpain large
subunit, catalytic domain (domain II) (1), dipeptidyl peptidase I (cathepsin C)
domain (1), viral Cys protease of trypsin fold (2), Ulp1 protease family (1),
and ubiquitin-activating enzyme (1).
The method further presented in this work is divided into two parts (Figure 1A): the first employs
knowledge-based information for detection of thiol oxidoreductases by analyzing
structural and compositional similarity to the active sites of known thiol
oxidoreductases; and the second makes use of energy-based methods to assess
properties of the catalytic redox-active Cys. For simplicity we refer to the
first part as Active Site Similarity, and to the second as Cys Reactivity.
The Active Site Similarity analysis included three independent steps: (i) amino
acid composition of active sites at two distances from the catalytic Cys; (ii)
structural profiles of active sites; and (iii) secondary structure profiles.
Each of these steps contributed to the scoring function (SF).
To analyze amino acid composition of the region surrounding catalytic Cys in
known thiol oxidoreductases, we determined the occurrence of amino acids within
a sphere centered at the sulfur atom of the catalytic Cys with two radii, 6
Å and 8 Å (Figure 1B). For this, we separately examined thioredoxin-fold,
FAD-containing, and other non-thioredoxin fold thiol oxidoreductases.
For comparison, we analyzed two sets of randomly chosen Cys-containing proteins
(800 and 1000 proteins, respectively), from which any proteins present in the
thiol oxidoreductase and control datasets were excluded. Cys residues present in
randomly chosen proteins represented an average composition of amino acids in
the vicinity of Cys in protein structures. For each of the six so-defined groups
of proteins (i.e., three groups of thiol oxidoreductases, a group containing
catalytic non-redox Cys, and two groups of randomly chosen proteins) an average
amino acid composition was calculated for 6 Å and 8 Å
distances from the sulfur atom of Cys (Figure 2). Interestingly, each group of
proteins with catalytic Cys showed unique amino acid occurrence that was also
different from those of the two sets of randomly chosen proteins. This was
particularly evident in the 6 Å datasets.
However, statistical analysis of these data (standard deviations are given in
Figure 2 and the
p-values for frequency counts are listed in Figure 3) showed that some differences
observed were not significant. Thus, a complete definition of
thiol-oxidoreductases based only on amino acid frequency is not possible.
Nevertheless, these data can be used, in a multi-parameter approach like the one
presented here, to contribute to the description and predictability of these
enzymes. Thus, we proceeded in our analyses considering the average values as
shown in Figure 2.
We further employed the average occurrences of each amino acid in the vicinity of
Cys as profiles (or dictionaries, to avoid confusion with the structural
profiles described later on), specific for each set, in which every amino acid
had its protein function-specific occurrence. The use of these dictionaries as
predictive tool is straightforward: for a given protein, occurrences of amino
acids located within 6 Å and separately within 8 Å of each
Cys sulfur atom are calculated, compared with the dictionaries of each reference
protein class, and scored.
The occurrence that receives the highest score is assigned to the corresponding
protein class. For example, when a score is closest to those of thiol
oxidoreductase dictionaries, it is considered positive, and in all other cases
it is considered negative. In the former case, a positive value (0.375 from each
of the 6 Å and 8 Å distance calculations) is given to the
final SF while in other cases a null value is given. Thus, the dictionary
component of the Compositional analysis can give an overall contribution of up
to 0.75 to the SF.
These analyses detected differences in amino acid occurrence around catalytic Cys
between thiol oxidoreductases and proteins containing catalytic non-redox Cys
residues. In addition, within thiol oxidoreductases, the amino acid composition
of FAD-containing enzymes was unique. For example, thioredoxin-fold thiol
oxidoreductases showed an overall high representation of aromatic residues near
the catalytic Cys, whereas FAD-containing thiol oxidoreductases showed lower
occurrences of these residues. Thus, for this step of the procedure,
FAD-containing thiol oxidoreductases were not considered. It should be noted
that this did not affect the overall analysis as other steps of the method
performed well with these enzymes and they could still be identified by the
overall method. With this restriction, we found that several amino acids,
including Pro, Cys, Trp, Tyr and Phe, were overrepresented in thiol
oxidoreductases (Figure 2).
At the same time, Met, His, Gly, and Glu were found to be less frequent in these
proteins.
Based on this information, we empirically defined the following formula that
allowed separation of thiol oxidoreductases and other Cys-containing proteins:
(W+Y+F+1.5C+0.5P)/(G+H+Q+2M),
where letters correspond to abundances of amino acids (in single letter code)
and the numbers are coefficients. In developing this formula, we sampled
different coefficients and applied the formulae to true positive and control (S1
and S2) datasets. The coefficients most efficiently separating thiol
oxidoreductases from other proteins were kept.
The ratio in the formula reflected common features of thiol oxidoreductases,
distinguishing them from enzymes containing non-redox catalytic Cys. For
example, active sites of thiol oxidoreductases preferred non-polar aromatic
residues. While all aromatic amino acids were overrepresented (compared to their
average values in control sets, see Random Cys in Figure 2), histidine was less frequent (but
it had high frequency in non-redox proteins with catalytic Cys). Consequently,
all aromatic residues appeared in the numerator of the formula, but histidine
was placed in the denominator. Other features of catalytic Cys were also
included in the formula such as the well known preference for a second Cys
(often a resolving Cys) in the proximity of the catalytic Cys, while the enzymes
containing non-redox catalytic Cys showed a significant underrepresentation of
additional Cys in the active sites. Proline is also often observed in thiol
oxidoreductases, but is less frequently found in other enzymes (Figure 2).
Although the chemical basis for differences in the use of amino acid in the
vicinity of Cys is not fully clear, the application of this formula was found to
be quite effective. Generally, values higher than 1.0 corresponded to thiol
oxidoreductases. For example, 79% thiol oxidoreductases (Table S1)
had scores higher than 1.0, whereas in the control dataset (Table S2),
88% proteins had a score lower than 1.0.
When representatives of the Random Cys sets were screened with the formula, the
ratio of false positive prediction (i.e., non thiol oxidoreductases scoring
higher than 1.0) somewhat increased, e.g., among 100 analyzed proteins from the
Random Cys set 1, 22% scored above 1.0. Interestingly, many of these
scoring proteins contained metal-binding Cys. This was mainly because Cys
residues clustered in these proteins (e.g., in zinc finger or iron-sulfur
cluster-containing proteins). Thus, the contribution to the SF from this last
component of the Compositional analysis was lower than that of the dictionaries,
adding a value of up to 0.25 to the SF. Finally, when the three components of
the Compositional analysis (analysis of dictionaries for 6 Å and 8
Å and the application of the formula) were considered, the
contribution to the SF ranged from 0 to 1.0 (Figure 1).
A previous study assessed structural similarity of reaction centers by profiling
functional sites in proteins [13]. It built a signature sequence of amino acids
located in the active sites. In our work, segments of amino acids in the active
sites were extracted from the structure and combined into a single contiguous
sequence (called either structural profile or active site signature). A similar
approach was recently employed to examine proteins with Cys oxidized to sulfenic
acid [8], in which active sites were defined as an area
located within 10 Å from the oxidized Cys. This study [8]
proposed that pairwise alignments between signatures can be effective in
predicting protein function by analyzing an unknown profile against a set of
known profiles.
We used this idea and employed the 8 Å active site signatures derived
from each thiol oxidoreductase in our dataset (Table S1)
as the set of known profiles. It should be noted that, compared to the original
procedure [13], the parameters for weighting pairwise
alignments (i.e., relative weights for similarities, gaps and identities) were
empirically optimized to achieve the best separation of thiol oxidoreductases
and reference datasets (Figures S1 and S2). The
optimized parameters for equation 1 are described in detail in the Methods section.
The ability of this procedure to separate thiol oxidoreductases from other
proteins is remarkable; using an appropriate cut off for the output of equation
1 (for example, 0.4 in Figure S1) as described in the Methods section, no false positives were
detected. This feature (i.e., very low false positive rate) opened up an
opportunity, based on the structural profile analysis, to assign a wider range
of values as contributing to the SF. In particular, values higher than 1.0 could
be given to the SF when the output of equation 1 is sufficiently high. However,
values higher than 1.0 were appended to the SF only under the conditions where
the probability of false positive predictions was either null or very low. The
contribution of this procedure to the SF ranged from 0 to 2.5 with the latter
occurring only when the profile of a putative protein under examination was
almost identical to that of a known thiol oxidoreductase. Further details on
this part of the procedure are given in the Methods section.
We analyzed secondary structure composition within 6 Å from the
catalytic Cys for all proteins in our datasets (Figure S3).
A marked preference for alpha helical and loop geometries around the Cys was
found in thiol oxidoreductases. In turn, beta strands were infrequent (with
notable exception of MsrBs).
We implemented these observations with a simple function requiring helical
composition exceeding 35% and loops exceeding the composition of
strands. As alluded above, some thiol oxidoreductases (MsrBs, fRMsrs and
arsenate reductases) were missed at this step of the analysis. Since this
procedure could potentially miss other candidate thiol oxidoreductases, its
contribution to the SF ranged from 0 to 0.5.
When the three steps of the procedure (i.e., amino acid composition, structural
profile and secondary structure composition of the active sites) were applied
together to thiol oxidoreductase and control datasets, a nearly complete
separation of thiol oxidoreductases and other proteins was achieved (Figure S4).
Each thiol oxidoreductase (Table S1) received scores higher
(≥1.5) than any control protein (Table S2 and a representative subset of the
randomly chosen proteins), with a single exception: a low molecular weight
tyrosine phosphatase (PDB code 1D1P) scored as high as some of the low scoring
thiol oxidoreductases. However, this phosphatase showed marked analogy to thiol
oxidoreductases (e.g., some proteins annotated as low molecular weight tyrosine
phosphatases are in fact arsenate reductases). We discuss this feature in
greater detail later in the text (see results of the Yeast Analysis).
We hypothesized that properties of redox-active catalytic Cys could also be
suitable for distinguishing thiol oxidoreductases from proteins with other Cys
types. In addition, proteins with catalytic Cys could potentially be
distinguished from those with non-catalytic Cys by virtue of thiol
oxidoreductases being enzymes. Thus, we examined available active site
prediction programs with respect to recognition of Cys active sites in thiol
oxidoreductases. These programs included Q-site finder (http://www.modelling.leeds.ac.uk/qsitefinder/), Pocket finder
(http://www.modelling.leeds.ac.uk/pocketfinder/), THEMATICS
(http://pfweb.chem.neu.edu/thematics/submit.html), SARIG
(http://bioinfo2.weizmann.ac.il/̃pietro/SARIG/V3/index.html)
and FOD (http://bioinformatics.cm-uj.krakow.pl/activesite/). All of these
programs are freely accessible via web service, but some calculations could be
slow (e.g., THEMATICS).
For each program, we examined randomly chosen 15 thioredoxin fold and 15
non-thioredoxin fold thiol oxidoreductases (Figure 4). Two programs, FOD and SARIG, were
ineffective in predicting catalytic sites of thiol oxidoreductases. Pocket
Finder performed slightly better but still clearly missed many active sites with
catalytic redox-active Cys. The best methods for thiol oxidoreductase prediction
proved to be Q-site finder and THEMATICS. The use of THEMATICS is limited by its
speed. Thus, Q-site finder was further employed. Scoring of this method is
detailed in the Methods section. Briefly,
if a catalytic Cys ranked within the first 3 sites, a positive value (1.0) was
given to the SF, and a zero value was given if the sulfur atom of Cys was not
predicted in any of the 10 ranked sites. Intermediate situations resulted in the
contributions to the SF, declined in the range between 0 and 1, as detailed in
the Methods section.
The final step of our algorithm examined Cys titration curves. As discussed in
the Introduction, pKa and exposure have
recently been proposed as parameters that distinguish redox-regulated Cys from
other Cys types [11]. However, when applied to our dataset, they
proved to be ineffective in detecting differences between redox and non-redox
catalytic Cys residues (Figure S5). This is indeed not surprising, as
sulfur exposure and a reasonably low Cys pKa should be necessary features for
both thiol oxidoreductases and enzymes with other nucleophilic catalytic Cys.
Thus, we examined other properties and methods that could account for
accessibility and reactivity of catalytic redox Cys. While Q-site finder may
possibly account for effective accessibility of Cys to small molecular probes
[14], it provides no information on Cys chemistry. An
alternative was to directly employ theoretical titration curves of active site
Cys residues. Indeed, the main idea of THEMATICS is based on the observation
that theoretical titration curves and their deviation from standard
Henderson-Hasselbach (HH) behavior can inform on the location of active site
residues (if they are titrable). Analysis of the theoretical titration curve of
a titrable residue is often more informative than simple calculations of its pKa
[15]–[18].
In this work, we employed the web accessible H++ server [19] to
calculate Cys titration curves and developed in-house tools for analyzing the
output (details are in the Methods
section). Briefly, we examined theoretical titration curves of each candidate
Cys and compared them with the standard HH behavior. The two curves were
superimposed and numerically compared (Figure 5). Greater deviation between the two
curves (Figure 5A) implied a
higher probability of the Cys being part of the active site and was given a
positive contribution (up to 1.0) to the SF, whereas small deviation or no
deviation (Figure 5B) was
given a zero contribution. We combined the methods discussed above in a single
algorithm shown in Figure
1A.
For the initial test of the algorithm, we selected a set of randomly chosen
proteins (Test Case) not included in the datasets used to develop the method,
which consisted of 22 thiol oxidoreductases (13 thioredoxin-fold proteins, 4
FAD-binding proteins and 5 other non-thioredoxin fold enzymes), 13 proteins with
catalytic non-redox Cys and 21 proteins with non-catalytic Cys known to be
redox-regulated through nitrosylation or glutathionylation (Table S3).
Several Test Case proteins were homology models. We deliberately included them
as structural models ultimately represent application of the program to proteins
with unknown structures.
The Test Case was also used to analyze weight distribution for each parameter of
the algorithm; this process supported parameter weights shown in Figure 1A (values in
brackets). Details of these calculations are shown in Figure S6,
available as supporting information. We also assessed method performance upon
changes in weights, and this is shown in Figure 6 (details are given in the figure
legend).
The output of the algorithm with optimized parameter weights (Figure 1A), applied to the
Test Case, is shown in Figure
7. Complete separation of thiol oxidoreductases (shown by blue circles)
from proteins with other Cys functions (green circles) was achieved with a
cutoff value of 2.75. Details of the calculations for each protein in the Test
Case are shown in Table S3.
To validate the algorithm on a genome-wide level, without any bias in the
selection of proteins, we applied the method to the Saccharomyces
cerevisiae proteome. Initially, we selected a subset of yeast
proteins by including (i) all known thiol oxidoreductases found by literature
search and detected by PSI-BLAST searches using known thiol oxidoreductases as
queries; and (ii) all other proteins in the yeast proteome containing at least
one highly conserved Cys (conserved in ≥90% homologs). From
this set, proteins containing metal-binding Cys residues were filtered out using
Prosite patterns. The resulting set of 292 proteins was subjected to homology
modeling via Swiss Model (http://swissmodel.expasy.org/) or HOMER (http://protein.cribi.unipd.it/Homer/), which generated 149
structural models (Table S4).
Among these proteins, 42 were predicted by our algorithm as thiol oxidoreductases
(i.e., scored≥the cutoff value of 2.75) (Figure 8 and Table S5
and Table
S6). Interestingly, 33 of the 42 predicted proteins were indeed known
thiol oxidoreductases, and the remaining 9 proteins likely included candidate
thiol oxidoreductases and false positives. The correctly predicted thiol
oxidoreductases were (Table S6) 6 glutaredoxin/glutaredoxin-like
proteins (>gi|6320720, >gi|6323396, >gi|6320492,
>gi|6319814 >gi|6320193, >gi|6321022), 4
thioredoxins/thioredoxin-like (>gi|6319925, >gi|6321648,
>gi|6323072, >gi|6322186), 1 glutathione reductase
(>gi|6325166), 2 thioredoxin reductases (>gi|6321898,
>gi|6320560), 1 Ero1 (>gi|6323505), 1 Erv1 (>gi|6681846), 1
Erv2 (>gi|6325296), 5 peroxiredoxins/peroxiredoxin-like
(>gi|6323613, >gi|6320661, >gi|6320661, >gi|6322180,
>gi|6319407), 2 glutathione peroxidases (>gi|6322228,
>gi|6322826), 1 alkyl hydroperoxidase (>gi|6323138), 1
methionine-S-sulfoxide reductase (>gi|6320881), 1 methionine-R-sulfoxide
reductase (>gi|6319816), 4 protein disulfide isomerases
(>gi|6319806, >gi|6324484, >gi|6324862,
>gi|6320726), and 1 dihydrolipoamide dehydrogenase (>gi|14318501).
The results are further illustrated in Figure 8 where all 149 yeast proteins for
which models have been generated are represented (green circles correspond to
known thiol oxidoreductases).
One of the candidate thiol oxidoreductases was 6-O-methylguanine-DNA methylase.
Interestingly, in addition to this algorithm, this protein was predicted as
thiol oxidoreductase by a method based on Cys/Sec pairs in homologous sequences
[12]. As the structure of E. coli
6-O-methylguanine-DNA methylase is known (PDB code 1sfe), we separately
subjected this protein to our algorithm. For Cys135 of this protein, the score
was 3.75, a value above the cutoff. The same Cys was predicted by the Cys/Sec
method. Overall, the data suggest that yeast 6-O-methylguanine-DNA methylase is
a strong candidate for a novel thiol oxidoreductase.
Other predictions included (i) >gi|6325330| homologous to mammalian PTP
(LTP1), (ii) >gi|6321631| glyceraldehyde-3-phosphate dehydrogenase
1(GAPDH-1); (iii) >gi|6322409| glyceraldehyde-3-phosphate dehydrogenase 2
(GAPDH-2); (iv) >gi|6324268| similar to tRNA and rRNA
cytosine-C5-methylase (NOP2); (v) gi|14318558| ubiquinol-cytochrome c
oxidoreductase subunit 6 (QCR6); (vi) >gi|6322155|ref|NP_012230.1|
capping - addition of actin subunits (Cap2p); (vii)
>gi|6321388|ref|NP_011465.1| hypothetical ORF (Ygl050wp); and (viii)
>gi|6322921|ref|NP_012994.1| hydrophilic protein implied in targeting and
fusion of ER to Golgi transport vesicles (BET3). While the functions of some of
these proteins are not known, the first three are worth a comment. GAPDH
proteins are known to have a catalytic nucleophilic Cys in the active site which
is highly sensitive to redox regulation by both thiols and reactive oxygen
species [20]–[22]. Oxidized GAPDHs were
also found to recover full activity in the presence of thioredoxin [23] or
DTT [24]. It appears that these proteins share properties
with thiol oxidoreductases, and their active site Cys showed common features
with catalytic redox Cys in other enzymes.
Low molecular weight protein tyrosine phosphatases (lwPTP) share the
phosphotyrosine protein phosphatase I-like fold with arsenate reductase (ArsC)
of gram-positive bacteria and archaea [25], which are thiol
oxidoreductases. These enzymes (lwPTP and ArsC) belong to the same superfamily
(phosphotyrosine protein phosphatases I). In our original dataset, there were
two ArsC proteins (PDB coded 1LJL and 1Y1L), and recognition of their
nucleophilic catalytic Cys as redox-active residues may reflect such similarity.
With regard to other predictions, no strong evidence to support or reject them
as thiol oxidoreductases was obtained, so at least some of these proteins could
indeed be thiol oxidoreductases.
Finally, one known thiol oxidoreductase among the 149 modeled yeast proteins was
not correctly detected by our method. This protein was a monothiol glutaredoxin
(>gi|6319488, GRX7), which corresponds to the single green circle in
Figure 8 located
slightly below the cut off value. However, the only contribution to the score
for this protein came from the Active Site Similarity method, whereas the Cys
Reactivity contribution was zero: Q-site finder did not predict its catalytic
Cys in any one of the 10 ranked sites, and the Cys titration curve strictly
followed HH behavior. We also submitted the protein to the THEMATICS server, but
its catalytic Cys was not predicted as an active site residue. The fact that
these independent structure-based calculations, which proved to be quite
effective in other analyses, did not recognize the active site and its catalytic
Cys could potentially be explained by poor quality of the homology model.
It can be argued, that the Similarity part of our algorithm should work better
than the Cys reactivity part with scarcely refined (but still reasonable)
structural models, due to its lesser dependence (especially for secondary
structure and compositional analysis) on the accuracy of predicted atomic
positions; these, in turn, determine titration curves and all types of
docking-like calculations (e.g., Q-site finder predictions). Therefore, poorly
refined structural models should affect predictions of the energy-based
calculations of the Cys reactivity part of the method to a greater extent.
Finally, it should be noted that all other glutaredoxins and glutaredoxin-like
proteins could be confidently predicted, which is consistent with the idea that
the low score for Cys reactivity in GRX7 may be related to the quality of the
structural model rather than inability of the procedure to detect this specific
protein. Overall, the method presented in this study showed very good
selectivity and specificity. It should find applications in examining protein
structures and identifying new thiol oxidoreductases and catalytic redox-active
Cys residues in these proteins.
During the review of our study, another paper was published [26] that analyzed
performance of active site prediction and employed multiple and independent
parameters. The authors observed improved performance when the analyses included
theoretical titration curves, residue exposure and sequence alignment-based
conservation scores. This study and our work suggest that implementing different
chemical (e.g., titration curves), physical (e.g., solvent accessibility), and
biological (e.g., sequence alignment) parameters offers a promising approach to
develop efficient tools for protein structure-function predictions. Such
approaches may allow the user to achieve specific biologically meaningful
insights, a feature often missing in predictive bioinformatics tools. Finally,
we suggest that the use of similar approaches may address the challenging issue
of prediction of Cys-based modification sites in proteins.
A set of known thiol oxidoreductases present in Saccharomyces
cerevisiae was collected by searching literature, analyzing homology to
known thiol oxidoreductases from other organisms, and similarities to
Sec-containing proteins [12]. Sequence alignments were prepared with
PSI-BLAST against the NCBI nonredundant protein database with the following
search parameters: expectation value 1e-4, expectation value for multipass model
1e-3, and maximal number of output sequences 5,000. Cys conservation for yeast
Saccharomyces cerevisiae proteins was determined using an
in-house Perl-script by parsing the PSI-BLAST output.
Models were built via Swiss Model (http://swissmodel.expasy.org/) and HOMER (http://protein.cribi.unipd.it/Homer/). VegaZZ 2.2.0 molecular
modeling package was used to check for missing residues, and for minimization
runs (with CHARMM22 force field), fixing planarity problems, editing multiple
sidechain conformations, adjustment of incorrect geometries, and residue
renumbering. Most of these operations were required for successful submission to
a server, such as SARIG and H++. With HOMER analyses, the
selection of template for modeling was done using PDB Blast. Calculations of pKa
values for dataset proteins were made with H++ server and with
PropKa implementation in VegaZZ (only for calculations shown in Figure S5,
for consistency with the previously published procedure [11]). Calculations of
accessible surface area were performed with a standalone program, Surface 4.0,
downloaded from http://www.pharmacy.umich.edu/tsodikovlab/.
The overall procedure was based on observations of unique properties of active
sites and catalytic Cys in thiol oxidoreductases. Each parameter of the method
(Figures 1) was
optimized for the ability to separate thiol oxidoreductases from other proteins.
Optimization of the parameters was carried out on an empirical basis: separately
for each subpart of the method we tested different parameters and calculations
were then performed against the dataset. The parameters which permitted better
resolution of the dataset (i.e., better separation of thiol oxidoreductases from
set 1 against other reference proteins – set S2 and representatives of
the Random Cys set) were kept and used in the composite procedure. A
representative example is given in Supporting information, Figure S2.
To analyze the relative weight distribution for each parameter of the algorithm
and how the algorithm performance depends on them, we carried out calculations
of a set of proteins (Test Case, described in the Results section) not belonging to the dataset. This analysis
supported the arrangements of parameter weights shown in Figure 1A (values in brackets). Details of
these calculations are shown in Figure S6, Figure 6, and Figure 7. The analysis of the Test Case also
allowed us to identify a cut-off value for the scoring function (described later
on in this section) to efficiently discriminate thiol oxidoreductases form other
proteins. The final scoring function, SF, was made up of contributions from each
part of the method, as detailed further in this section. The overall method was
divided into 2 parts: the first, Active Site Similarity, analyzed structural
similarity of test proteins to known thiol oxidoreductases. The second, Cys
Reactivity, employed external software for energy-based calculations of Cys
properties. Both parts were further subdivided into subparts, as shown in the
scheme of the algorithm in Figure
1, and each is further discussed separately here in the Methods section.
Analysis of amino acid composition around Cys was carried out with in-house tools
written in Python (v2.4). Detection of amino acids within a cutoff distance (6
Å or 8 Å) from the catalytic Cys sulfur was made considering
all residues with one or more of their atoms at a distance equal or lower than
the cutoff. A simple graphical representation is shown in Figure 1B. We employed this procedure for all
proteins in the dataset (Table S1 and Table S2),
divided into 4 categories: (i) thioredoxin fold thiol oxidoreductases (Trx OxR);
(ii) non-thioredoxin fold thiol oxidoreductases (Non Trx OxR); (iii) FAD-binding
thiol oxidoreductases (FAD OxR); and (iv) proteins with catalytic non-redox Cys
(Non OxR). For each protein category, we computed an average amino acid
composition. This is shown in graphical form in Figure 2. Frequency of amino acid occurrence
was associated with each amino acid (the Y value in Figure 2). Consequently, four separate sets
of amino acid compositions were built for Trx OxR (blue bars in Figure 2), Non Trx OxR (pink
bars), FAD OxR (red bars), and Non OxR (green bars). We stored information for
each protein category in the form of specific dictionaries (after the Python 2.4
datatype actually employed), wherein each amino acid received a value of its
frequency.
In addition, two other sets of non-overlapping randomly selected proteins, one
made of 800 PDB structures and the other of 1000 structures, were built. These
sets were designated Random Cys set 1 and Random Cys set 2 and represented an
average composition of amino acids in the vicinity of Cys in protein structures.
Combined together these two sets made up the Random Cys set (bright yellow bars
in Figure 2). We required
that these sets have no overlap with datasets S1 and S2. Also for these two
sets, two specific dictionaries were built to store the set-specific amino
acidic composition. The use of the six dictionaries to carry out compositional
analysis is illustrated with the following example. Given the following short
structural profile, i.e., the amino acid sequence in the active site,
Cys-Ala-Val-Glu, and the following dictionaries,
When applying each dictionary separately to the profile, three different scores
are received, each obtained by appending the average set-specific frequency
value corresponding to an amino acid of the profile:
In this example, the highest score is obtained with the
“Set3” dictionary. If “Set3”
corresponds, for instance, to the Trx OxR dictionary, the putative sequence
resembles the composition of thiol oxidoreductases. In this case, a value of
+0.375 is added to the final scoring function, SF. The same happens if
the best scoring dictionary is that of Non Trx OxR. If instead the best scoring
dictionary is one of non-thiol oxidoreductases, then a zero contribution is
given to the SF. These dictionary-based calculations were done with 6
Å and 8 Å distance profiles, thus contributing a maximum
value of 0.75 to the SF (0.375 for the 6 Å profile and 0.375 with the
8 Å profile).
Another evaluation formula, limited in this case to the 6 Å distance
(because this distance shows the most significant difference among proteins in
the dataset, see Figure 2)
was based on the following ratio
(W+Y+F+1.5C+0.5P)/(G+H+Q+2M),
where letters correspond to the single letter code for amino acids and numbers
are coefficients. This empirical ratio was chosen as discussed in the Results session. We sampled different
coefficients (0.5, 1, 1.5, 2) for the amino acid composition in this formula: in
each case the same datasets (S1 and S2) were used. The coefficients most
efficiently separating thiol oxidoreductases from other proteins were kept. When
the formula was applied to a profile for a putative active site, the result (x)
was analyzed as follows. If x≥1.5, a value of 0.25 was given to the SF.
If it was between 1.5 and 1.1, a value of 0.125 was given. Otherwise a zero
value was given.
For this step in the procedure, we followed a previously published procedure of
functional site profiling [13]. Accordingly, we employed ClustalW (http://www.ebi.ac.uk/Tools/clustalw2/index.html) standalone
version 2.0.3 for pairwise alignment calculations between a putative profile and
each reference profile extracted from our dataset of known thiol oxidoreductases
(Table
S1). The evaluation function was carried out with Equation 1(1)where SI represents identities (n is the total number of
identities in the alignment), Ss strongly conserved residues (m is the total
number of Ss in the alignment), Sw weakly conserved residues (K is the total
number of Sw ), Sg gaps (j is the total number of Sg) and N is the number of
paired residues in the alignment.
Modified parameters were used for Equation 1 (in parenthesis are the original
values, also derived empirically): SI = 1.0
(1.0), Ss = 0.3 (0.2),
Sw = 0.1 (0.1),
Sg = 0 (−0.5). Starting from the
original parameters, we sampled different values to determine if it was possible
to improve the performance of Equation 1 against our datasets (S1, S2 and
representatives of the Random Cys sets). An example of the performance with
modified parameters is given in Figure S2. We found that an improvement can
be reached by underweighting the gaps, and we obtained the best results when the
gaps were treated like “non similar” paired residues. Our
parameters were more permissive than the original parameters, which were
developed and optimized to address a different biological question. It must be
stated that the original parameters performed better if the purpose was to
detect similarities between more related protein sets (for example, functional
families). For the analysis of distantly related proteins, a relaxation of
parameters was necessary, and we obtained the best results with our more
permissive ad hoc optimized parameters (Figure S2).
The flow of our structural profile analysis was as follows: given a putative
active site, pairwise alignments were made with ClustalW between the putative
profile and each of the profiles extracted from the known thiol oxidoreductases
in our dataset (Figure 1B).
Each pairwise alignment was evaluated with Equation 1. The highest scoring
alignment was selected and its score value (x) was kept for further analysis. If
the best result (x) of Equation 1 was lower than 0.35, a null value was given to
the SF. If 0.35≤x<0.4, a value of 0.5 was given. If
0.4≤x≤0.5, a value of 1 was given. If 0.5<x≤0.6, a
value of 1.5 was given. If 0.6<x≤0.75, a value of 2 was given.
Finally, if x>0.75, a value of 2.5 was given to the SF. In the latter
case, a x value higher than 0.75 actually meant that this profile was almost
identical to that of a known thiol oxidoreductase.
We analyzed the secondary structure content of active sites of each thiol
oxidoreductase in our dataset and then compared them with proteins in the
control sets. A three-state secondary structure classification (helix, strand,
or coil) was assigned to each amino acid within 6 Å from the Cys
sulfur atom. The evaluation was made as following: (i) if the helical content
was higher or equal to 35% and the coil content was higher than the
strand content, a value of +0.5 was given to the SF. (ii) If helical
content was equal to or higher than 10% and both the coil content and
the helix content were higher than the strand content, a value of +0.25
was given to the SF. In all other cases, this part of the method received a zero
contribution.
Thus, the overall contribution to the SF from the Active Site Similarity part of
the method ranged from 0 to 4.0; once again it must be clearly stated that the
latter value occurred only when a putative active site was nearly identical to
that of a known thiol oxidoreductase.
This part of the method was based on, but not limited to, calculations from two
publicly available external servers, Q-site finder (http://www.modelling.leeds.ac.uk/qsitefinder/) and
H++ (http://biophysics.cs.vt.edu/H++/index.php). We first discuss the
use of Q-site finder. For an overview of this program, we refer the reader to
the original paper [14]. To automate the analysis, the predictions of
Q-site finder were parsed in html format with an in-house Python tool. We
developed an ad hoc scoring of 10 differently ranked sites in
the Q-site finder output, derived on an empirical basis (i.e., by testing
against all dataset proteins).
A value of 1.0 was given to the SF if a Cys was predicted with its sulfur atom
among the first 3 sites, as ranked by Q-site finder. A value of 0.5 was given to
the SF if the sulfur atom was predicted in the 4th, 5th or
6th site. A value of 0.25 was given if the sulfur atom was
predicted in one of the remaining sites. If a residue was predicted in more than
one site, only the highest ranked site was considered.
H++ server calculations were performed by choosing the
following parameters: the interior dielectric constant (protein ε) was
set to 20 while the solution dielectric constant was set to 75. Salinity (sodium
chloride) of the medium was set to 150 mM. Of the H++ server
output files, we considered only the *.pkaout files, which contained a
list of all titrable residues with their pKa values. In addition, the files
contained two-dimensional coordinates of theoretical titration curves for each
residue. Parsing the H++ output file with an ad hoc Python
tool, the values for the residue (in our case, Cys) were extracted, as well as
its calculated pKa. We further considered the Henderson-Hasselbach (HH) equation:(2)Equation 2 can be rewritten to show the charge on the titrable residue(3)where C− indicates a negative charge on the
sulfur atom. Equation 3 is valid for acidic residues, which acquire negative
charge upon titration (e.g., Cys). Substituting H++
pKa-calculated value in Equation 3 and varying the pH between 0 and 18, the HH
behaving curve for an acidic residue was then obtained. Figure 5 shows two examples of
superimposition of theoretical titration curves obtained by the
H++ server (red curves) and the corresponding HH behavior
curves (blue curves). The HH behavior curves were viewed as standard behavior of
the residue if no perturbations due to other nearby titrable residues occurred
[15]. Thus, deviation of the red curve from the blue
curve in Figure 5 (in the
titrable range around the pKa) pinpointed the active site residue [16],[17]. Automatic
evaluation of the deviation between the two curve behaviors could be a challenge
[27].
In the present work, we were only interested in a simple way to perform a quick
quantification of the deviation between the two curves. Thus, point by point
subtraction (for each pH value) between the two curves was carried on. These
values were integrated over the entire pH range, resulting in the overall
difference absolute value (Σ Δ) for the deviation between the
two curves.
Σ Δ was next evaluated to give a contribution to the SF. Cutoff
values employed were as follows: if |Σ Δ|≥2.0, then a
value of 1.0 is given. If 2.0>|Σ Δ|≥1.5, a value
of 0.75 is given. If 1.5>|Σ Δ|≥1.0, a value of 0.5
is given. If 1.0>|Σ Δ|≥0.5, value of 0.25 is
given. Values below 0.5 correspond to a small or null deviation from the typical
HH titration behavior (Figure
5B), and consequently a zero value is given to SF. The overall
contribution of the Cys Reactivity method to the SF ranged from 0 to 2.0.
Finally, in the complete algorithm, the resulting value of the SF ranged from 0
to 6.0 (Figure 1A). We found
that a value of 2.75 was a minimum cutoff value that positively discriminated
catalytic redox-active Cys residues (Figure 7 and Figure 8).
|
10.1371/journal.pcbi.1006408 | The low spike density of HIV may have evolved because of the effects of T helper cell depletion on affinity maturation | The spikes on virus surfaces bind receptors on host cells to propagate infection. High spike densities (SDs) can promote infection, but spikes are also targets of antibody-mediated immune responses. Thus, diverse evolutionary pressures can influence virus SDs. HIV’s SD is about two orders of magnitude lower than that of other viruses, a surprising feature of unknown origin. By modeling antibody evolution through affinity maturation, we find that an intermediate SD maximizes the affinity of generated antibodies. We argue that this leads most viruses to evolve high SDs. T helper cells, which are depleted during early HIV infection, play a key role in antibody evolution. We find that T helper cell depletion results in high affinity antibodies when SD is high, but not if SD is low. This special feature of HIV infection may have led to the evolution of a low SD to avoid potent immune responses early in infection.
| The spike protein on the virus surface mediates its entry to the host cell and a high spike density promotes infection. HIV has a spike density that is almost two orders of magnitude lower than other viruses. This unique feature of HIV has defied explanation since it was first observed. By bringing together theory and computation, rooted in statistical mechanics, with immunology we suggest that the effects of dramatic depletion of T helper cells during HIV infection on antibody production provided an evolutionary driving force for HIV to evolve a low spike density in order to avoid potent immune responses. Additionally, we show that an intermediate spike density induces maximally potent antibody production, a result with implications for vaccine design.
| Viruses gain entry into their host cells by attaching to specific receptors on the host surface. The proteins that mediate entry comprise the viral spike. Since the host receptor does not mutate rapidly, spike proteins, while often being highly mutable, have conserved regions that bind to elements on the host receptor. For example, the HIV spike protein gp120 contains relatively conserved residues that bind to the CD4 co-receptor on T helper cells. In influenza, the spike is composed of a HA glycoprotein, that attaches to sialyl-oligosaccharide, which is a sugar found in many cell surface proteins [1].
From the standpoint of mediating cell entry and thus propagating infection, it is evolutionarily beneficial to exhibit a high concentration of spikes on the virus surface, thus increasing the probability of attaching to host cell receptors [2]. But, parts of the proteins that comprise the viral spikes are also the targets (epitopes) of antibodies produced by the humoral immune response. A lower spike density (SD) would hinder antibodies from binding to the same epitope on two spikes on the viral surface simultaneously with its two arms, thus taking advantage of cooperativity of binding by the two arms (avidity) [3]. Thus, there is also an evolutionary driving force for viruses to evolve a low SD. But, evasion of potent immune responses may not always favor a low SD. For example, in influenza, hypervariable features at the head of the spike have high immunogenicity. A high SD protects the more conserved domains near the stem from being targeted by antibodies [4]. In HIV, the conserved regions are partially protected from the action of antibodies by a shield of glycans or by their membrane-proximal location (a high SD would presumably better shield the latter epitopes). Furthermore, many immunogenic epitopes that do not include any conserved residues are also present on the HIV spike.
Available data indicate that most viruses express a very high number of spikes on their surface. For example, Influenza has around 450 spikes or a SD of 1 spike per 100 nm2 [5], HCV has 250 spikes or SD of 1.73 per 100 nm2 [6]. Table 1 summarizes much of the available information on SDs of common viruses, and this data shows that HIV is an extreme outlier, exhibiting between 7 and 14 spikes on its surface, resulting in a very low SD of 0.01 spikes per 100 nm2, which is 50–100 times lower than that for other viruses [7]. So, it appears that most viruses evolved high SDs (presumably enhancing infectivity and possibly distracting the immune system from targeting more conserved epitopes), while HIV has not. If HIV has evolved a low SD to avoid potent antibody responses, why have other viruses not employed the same strategy? If HIV spike proteins were significantly more vulnerable to antibody responses compared to other viruses, this could explain why HIV has evolved a significantly low SD to lower the avidity of antibody binding to epitopes on the spike proteins. But, no evidence exists suggesting that this is true. Another possibility would be that HIV as a retrovirus, has some fundamental architecture constraints on the maximum number of spikes it can display. However, MLV (another retrovirus) has a high surface density [8] with at least a hundred spikes on its surface [9]. Shedding of the spike (gp120) as an immune decoy could be another reason for the low spike number. However, high spike density viruses such as Ebola also shed their glycoprotein spikes [10]. Hence, this mechanism is not sufficient to explain HIV’s uniqueness of low spike density. Thus, an obvious long-standing question remains unanswered: why has HIV evolved to exhibit a significantly lower SD compared to other viruses?
Upon infection with pathogens, high affinity antibodies develop by a Darwinian evolutionary process called affinity maturation (AM). We inquired if the biology of HIV may influence the effects of SD on AM in a way that is not characteristic of other viruses, and whether this is the underlying reason for a low SD being favored by HIV. To explore this possibility, we developed a coarse-grained computational model of the dynamics of AM.
Results of our calculations show that antibody affinity to the epitopes on the viral spike is a function of its SD for all cases. In particular, highest affinity antibodies are produced for an intermediate SD. To the best of our knowledge, this effect of SD on the resulting Ab affinity has not been reported before. Importantly, we find that the decline in antibody affinity when SD exceeds the optimal value is very gradual, while the affinity declines sharply for SDs below the optimum density. These results suggest that a high SD (beyond the optimum defined above) allows viruses to exhibit high infectivity and evade potent responses directed toward mutationally vulnerable epitopes if they are located at the stem of the spike, while also reducing the affinity of antibodies directed toward the spike. Note, however, that this still allows the immune system to generate reasonably high affinity antibodies as the decline in affinity for SDs higher than the optimum is gradual.
Why is HIV different? Our calculations suggest that the answer lies in a key feature of HIV infection. HIV principally infects T helper cells (CD4 T cells). We find that if T helper cells become extraordinarily limiting during affinity maturation, as is the case immediately following HIV infection, then high spike densities will elicit even higher affinity antibodies—a bad outcome for the virus. We find that this is circumvented if the SD is low. Therefore, our results suggest that a key benefit of a lower SD for HIV is an avoidance of high affinity antibody responses that would otherwise be produced if the SD was high when T helper cell availability becomes much more limiting than usual. However, the tradeoff for having low SD is reduced infectivity [11], a long noted feature of this virus. The virus’s low infectivity has not prevented HIV transmission from reaching epidemic proportions; perhaps, because of its high replication rate in infected hosts.
Antibodies develop in domains within secondary lymphoid organs called germinal centers (GCs), which appear shortly after infection [12]. B cells with a moderate threshold affinity for the antigen (Ag) are activated and seed GCs. These B cells then undergo an evolutionary process of mutation and selection that results in B cells with higher affinity receptors as time ensues [13]. The AID gene introduces mutations into the B cell receptor (BCR) at a high rate in GCs. The mutated B cells undergo selection against Ag, which is displayed on Follicular dendritic cells (FDC) on immune complexes (IC) [14] (Fig 1a). The B cells attempt to capture Ag by forming transient synapses with the FDCs [15]. Captured Ag is processed and presented on their surface as peptide-MHC class II complexes. The B cells then compete with each other to bind to T follicular help cells (TfhCs) via interactions between these peptide-MHC complexes and the T cell receptor on the surface of TfhCs. Successful binding results in a survival/proliferation signal. B cells that display more peptide-MHC complexes have an advantage in this competition[16]. The majority of positively selected B cells undergo further rounds of mutation and selection [17–19]. Some of the positively selected B cells differentiate into antibody-secreting plasma cells and memory cells. As time progresses, antibodies with increasing affinity for the Ag are generated.
Affinity maturation has been studied extensively using theoretical models over the last decades, mostly using population dynamics approaches [19–24] and more recently, using detailed simulation of the dynamics of the immune cells invovled in the GCR [25]. To explore how B cell selection and the generation of high affinity Abs depend on Ag concentration and presentation, we constructed a simplified model of Ag capture from the FDCs, and a coarse-grained model of B cell selection in a GC. Thus, our model generalizes previous modeling approaches to include the interaction of the BCR with Ag (spike).
We are interested in ICs presenting a virus, or a virus fragment (Fig 1a). The Ags of interest are the spike proteins distributed on the virus surface. Assuming that the spikes are distributed on the surface of the virus of radius R, with average density nAg, it contains a total of MAg = 4πR2nAg spikes (Fig 1b). A virion particle with a radius of 120nm (Table 1) with spike density of 0.35 spikes/100nm2, has about MAg = 160 spikes on its surface. During the Ag capture process, BCRs scan for Ag over a region of the synapse and encounter an Ag molecule with probability p. We assume that the number NAg of Ag molecules in the scanning area is distributed randomly, according to the Binomial distribution (NAg ~ B(MAg, p)).
Consistent with data showing that B cell protrusions that extended toward FDCs retract if the associated BCRs do not find Ag [26], we assume that the BCR has a characteristic time to find the Ag (see Methods). If it does not bind to Ag during that time, no Ag is captured (see Fig 1ci and 1cii). Once one of the arms of the BCR binds to Ag, an Ag molecule may be pulled away by force[27]. At this stage, there is a tug-of-war over the Ag between the BCR and the IC (see Fig 1ciii). Characteristically, when the binding energy between the BCR and the Ag is much larger than the binding energy between the Ag and the virus/Fc receptor/FDC membrane (see Fig 1), the Ag will be captured by the BCR. We denote the potential interaction energy required to extract the Ag by EAg−mem, and the interaction energy of the BCR/Ab and Ag by EAg−Ab. A successful Ag capture event occurs when the bond between the Ag and the membrane ruptures before the bond between the BCR and the Ag (see Methods).
When the off-rate of the BCR arm from the Ag is smaller or of the order of the effective on-rate (qNAg), and if one arm is attached initially (see Fig 1cii), the second BCR arm can bind to another Ag molecule if one is available. Thus, a pulling attempt can have three possible outcomes, with zero, one or two Ag molecules captured (see Fig 1d), depending on the binding affinity and the number of available and accessible Ag molecules NAg.
In the first days following infection or immunization, M different clones (unique BCRs) of activated B cells proliferate with little competition, creating a pool of cells on which AM operates[28]. Few or no mutations are introduced to the BCR sequence at this early stage. We use a simple birth/death process to describe this initial growth stage.
B cells then start to mutate, and whether or not a B cell is subsequently positively selected depends on its ability to internalize Ag and compete with other B cells for survival signals by interacting with TfhCs. TfhCs have an important role in regulating the duration of the cell cycle in B cells during AM [17,18]. Following a TfhC signal, B cells divide (and mutate) multiple (4–6) times before going back for another round of selection [29–31].
Most theoretical models enforce selection by eliminating cells with low affinity BCR [23,24]. It has been shown [17] that TfhCs control the rate at which B cells go through the cell cycle. B cells that receive strong proliferation signal from the TfhCs divide multiple times in the dark zone before going back to the light zone. We model this behavior by using a birth-rate for B cells which is proportional to the amount of captured Ag.
The proliferation rate of B cells is a function of the number of captured Ag molecules, further modulated through competition for TfhCs. Indeed, it was shown [32], that TfhCs preferentially interact with B cells that have captured more Ag, presumably giving them a stronger proliferation signal. To mimic this competition we assume the birthrate of B cell i to be:
βi=β0C+AiC+〈A〉,
(1)
where 〈A〉 is the average amount of Ag consumed by all B cells that captured at least 1 Ag (B cells that do not capture Ag have zero birth rate); β0 is the basal birthrate, and C is a measure of the availability of TfhC help. When C>>〈A〉, all B cells get the same amount of metabolic boost and divide at the same rate β0. When C<<〈A〉 TfhC help is limiting and the cell birthrate is proportional to the amount of captured Ag. The GC has a limited capacity and grows during this competitive phase according to a stochastic logistic growth process. Thus, the selection process depends on the number of B cells in the system through the logistic death rate (Methods Eq (6)). This logistic term accounts for the increased competition between B cells to receive a survival signal from TfhCs when the number of the latter is limiting. TfhCs are essential for GC maturation of B cells, and without them, AM of B cell cannot occur [33]. We assume here that the number of TfhCs is not so small as to stop the formation of the GC. Thus, at the limit where C goes to zero, our model does not corresponds to the absence of TfhCs in the GC. However, we do use the reduction of C to model increased B cells competition in a GC when the number of TfhCs is gradually reduced, as occurs during the first stages of HIV infection. For these reasons, we do not consider here the limit of C→0. In a more complete proliferation model C should depend on the total number of B cell in the GC and thus change with time. However, the qualitative behavior of our model would not change if the parameter C would have such explicit dependency on the GC size as upon HIV infection we expect Tfh levels to be smaller at all times than upon infection with other pathogens.
During AM, B cells mutate the BCR encoding gene using the AID enzyme. This results in changing cytosine to tyrosine on one DNA strand [34]. We modeled the effect of mutation as a change in the on-rate q or the interaction energy EAg−Ab upon cell division (see Methods). We convert the interaction energy to the off-rate by r0=e−βEAg−Ab, and the affinity of the BCR is calculated as ω = q/r0. Our simplified model assumes that affinity improvements and reductions (through q or EAg−Ab) are equally likely following mutation and that all mutations change affinity. These choices are certainly not realistic, but making advantageous mutations rarer than deleterious ones, or making some of the mutations silent or lethal (as in reality) do not change the qualitative results.
To explore the role of SD on AM, we calculated the affinity of the most dominant clone at the end of the GC reaction (GCR), which is day 16 of the competitive phase. The affinity of this clone will play an important role in the resulting humoral immune response. An important observation is that the affinity of the dominant clone varies non-monotonically with the SD (nAg). Fig 2a demonstrates this finding for a TfhC level that was limiting across the range of nAg (SD) spanning 0.01 to 0.36. There is also a pronounced asymmetry in the variation of affinity with nAg. As nAg increases from very small values, there is a sharp rise in affinity, while its drop for larger nAg is gradual. The number of B cells in the GC is not reduced when SD is high. Since B cells have to capture at least 1 Ag molecule in order to attempt proliferation, the average proliferation rate increases with SD.
These qualitative results are robust to variations in the other parameters in the model (S1 Fig), and for mutations resulting in changes in q (S2 Fig) or EAg−Ab (Fig 2a). (The variation of affinity with SD is less pronounced when mutations change q for technical reasons described in S2 Fig.)
In the GC, memory and plasma B cells are produced throughout the GCR [12]. As a proxy for the affinities of memory and plasma B cells produced during the maturation process, we randomly selected 10% of the B cells at each intermediate time point. The resulting average, displayed in (S3 Fig) still exhibits a non-monotonous dependence on SD similar to that observed in Fig 2a.
The mechanistic reason underlying these results relates to the amount of antigen that is internalized by B cells as nAg changes (Fig 2b). Upon increasing nAg from a small value, the amount of Ag internalized by B cells grows significantly as it becomes more probable to bind one or two Ag molecules (see Fig 2c). However, once the SD is sufficiently high, most BCRs bind/internalize on average the same amount of Ag irrespectively of their affinity (if it is above a threshold value), and it is difficult to distinguish B cells based on the affinity of their receptor for the antigen. Competition between B cells to be positively selected is not severe in the latter regime, limiting the driving force to increase affinity by further mutations. Thus, the affinity of the antibodies produced begins to decline beyond an optimal SD. The gradual drop-off in affinity as nAg becomes too high, compared to the sharp decline when nAg becomes too low, is because the number of internalized antigens rises very sharply as nAg increases from a low value, but then slowly saturates to its maximum for sufficiently high SDs as depicted in Fig 2b. In the limit of high nAg, there is a finite probability for the BCR to capture 1 or 2 Ag mols. At this limit, the probability to extract one Ag mol is given by r/(r + λ) and the probability to extract two Ag mols is λ/(r + λ), where λ is the off-rate of an Ag mol from the membrane when one arm of the BCR is bound and r is the rupture rate of the BCR from an Ag mol when one arm of the BCR is bound.
Our model suggests that clonal diversity also varies with SD. When SD is low, antigen is scarce and the selection forces will be fiercer. That could result in a more rapid loss of diversity. We estimate diversity by computing the fraction of the GC comprised of the dominant clone (S2b Fig, S4a Fig), which we denote by fd. For very low SD a few dominant clones comprise most of the GC’s B cells at the end of the GCR (large fd). This is because for low SD, selection during AM is dominated by the large competitive advantage, compared to most B cells, of a few B cells that internalize more antigen due to stochastic fluctuations. As the SD increases, a greater number of B cells can internalize antigen successfully and the GCR produces a more diverse clonal population.
The fraction of dominant clones has a large variability across different GC realizations. Indeed, it has been shown [31] that while some GCs appear to have been taken over by a single clone at day 16, others show high clonal diversity. This is the direct result of stochastic selection forces at play in the GC [35]. A similar behavior is observed in our GC model (see S4 Fig). Here, the source of stochasticity is the random Ag capture process, as well as the stochastic proliferation and death of B cells.
While the fraction of the dominant clone mostly decreases with increasing SD, the higher affinity of the BCR at intermediate density contributes to a small increase in fd. Thus, fd is not monotonically decreasing with SD. Indeed, the improved selection at an intermediate density leads to greater loss of diversity in this range, which appears as a small bump in fd (approximately between nAg = [0.1, 0.18], see inset S4 Fig). This effect is more pronounced when the overall diversity loss is slower as a result of a smaller death rate (see S1 Fig).
To understand how cell selection in AM is related to Ag density, we estimated the probability Pweaker aff selection that, in any given round of mutation and selection, a B cell with a lower affinity for the antigen gets selected and expands in favor of another B cell that has a higher affinity. A B cell with a higher affinity receptor is likely to internalize more antigen, be more successful at obtaining a survival signal from T helper cells and proliferate more. However, because Ag capture is probabilistic, and the birthrate relates to the amount of captured Ag, there is a chance that a B cell with lower affinity will be selected in favor of one that expresses a higher affinity receptor.
To empirically estimate pweaker aff selection, we sampled B cell pairs from the affinity distribution computed in our simulations (see S5 Fig). We detail how the probability is estimated in the methods section. Interestingly, we find that the probability changes with time, and depends on Ag density (Fig 2d). Importantly, pweaker aff selection is minimal at intermediate Ag density, at a value similar to the maximum for the BCR affinities. Thus, very high or low SDs lead to poor differentiation between higher and lower affinity B cells. Because the B cells with a higher affinity are most likely to be chosen to proliferate at an intermediate SD, selection is strongest for such SD, and thus highest affinity antibodies evolve.
To further elucidate the origin of the maximum in antibody affinity at an intermediate SD, we studied the time evolution of the mean affinity of the B cell population. We aimed to find the dependency of the optimal density on the rates of different processes in our model, rather than attempt to recapitulate our simulation results. Assuming that the time τ to find the first Ag molecule before retracting the B cell protrusion is very long and that each B cell has only a single BCR, we find a mean-field equation (see SI) for the evolution of the mean on-rate q¯
τ0dq¯dt=cov(q,A)〈A〉+C1,
(2)
where 〈A〉 is the average amount of captured Ag by the B cell population (here 〈A〉 is between 0 and 2), cov(q, A) is the covariance between the number of captured Ag molecules A and q, C1 is the parameter that corresponds to TfhC availability parameter and τ0 ∝ (β0−μ0)-1 is the time scale of the process. We recall Fisher’s fundamental theorem of natural selection stating that “The rate of increase in fitness of any organism at any time is equal to its genetic variance in fitness at that time” [36]. Eq (2) is a generalization of Fisher’s fundamental theorem and is also reminiscent of Price’s equation [37] if we consider q to be a trait of the population. We further find that the variance σ2 of the on-rate distribution evolves as
τ0dσ2dt=cov((q−q¯)2,A)〈A〉+C1+2D.
(3)
Solving Eqs (2) and (3) numerically (S6 Fig), we find that q¯ exhibits the same non-monotonic behavior as in our simulation results S2a Fig. When TfhC level is scarce (small C1), the affinity increases faster due to the increased competition between B cells (S6 Fig). The position of the optimal density changes with time (S6 Fig) towards smaller Ag densities. For a given on-rate distribution (fixed mean and variance), the rate of increase in affinity is maximal at density n* given by
n*=(r+λ)(k+ξ)q¯ξλ+C1(r+λ)(1+C1)r+(2+C1)λ,
(4)
where λ and ξ are the off-rates of the Ag from the IC when one or two arms of the BCR are bound, respectively. r and k are the BCR arm rupture rate from the Ag when one or two arms of the BCR are bound to Ag, respectively (see SI).
When the interaction energy of the Ab and the Ag is equal to that of the Ag with the IC (EAg−Ab = EAg−mem), the optimal density is
n*=4λ0q¯eβxbF1+2C13+2C1,
(5)
where λ0 is the disassociation rate of the Ag from the IC. Thus, at the optimal selection density, the characteristic effective on-rate (n*q¯) is equal to λ0. At this density, B cells spilt into different populations, each capturing a different amount of Ag.
Interestingly, for very large values of q, the number of captured Ag molecules A is independent of q as the covariance between them goes to zero (see SI). In this limit, it is impossible to differentiate between B cells affinities and the mean affinity stops increasing, reaching an asymptotic value. Similarly, the variance of the distribution reaches an asymptotic value.
Finally, we speculated whether a scenario where the BCR has one arm (see S7 Fig) would still show the non-monotonic dependence on Ag density. We estimated the time evolution of q¯ for a hypothetical BCR with one arm (see SI and S7 Fig). Interestingly, when the time τ to find the first Ag molecule is of the same order as the effective on-rate (basal on-rate multiplied by the density), the mean affinity has a maximum at intermediate densities (see S7 Fig). However, τ is likely much larger than the on-rate, in which case the affinity is monotonically decreasing as a function of the Ag density (see S7 Fig). We conclude that the optimal SD observed in our simulations is a direct result of the cooperativity between the two arms of the BCR because the analytical model shows that an optimal SD emerges only when the BCR has two arms. The cooperativity allows a second arm to search and bind an Ag molecule while the first is bound to another Ag molecule on the surface of the IC.
During AM, Ag is depleted from IC as it is being consumed by the B cells. As shown in the previous sections, AM depends on the ability of the GCR to differentiate bewteen B cells of different affinities. Optimal selection is achieved when only B cells with the highest affinity are likely to capture 2 Ag molecules (see Eq (5)). Depletion of Ag from FDCs during AM will result in fewer immune complexes encountered by the BCRs. It will not change the local SD on the virus. Rather, it will reduce the first encounter probability of Ag molecules by a BCR (See Eq (7)). To study how the competative forces may be modulated in time, we studied a hypothetical scenario where SD exponentially decays during AM (nAg(t)=nAg(0)e−t/TAg−decay). Interestingly, when the SD decays, selection improves when the initial SD is high (S8 Fig). As a result, the affinity at day 16 of the GCR is higher compared to the fixed SD scenario (Fig 2a). At high SD, as the affinity of the B cell population improves, decreasing SD allows the GCR to remain at close to the optimal differention point. However, decreasing Ag density harms the development of GCs for which the initial SD is low. Since B cell have to capture at least one Ag mol to proliferate, most GCs do not succeed in reaching day 16 (S8 Fig) in this limit.
HIV dominantly infects CD4 T cells and significantly reduces their number immediately following infection during a time when the first antibody responses are developing in the host [38]. We therefore explored the effects of making TfhC activity even more limiting (than what is usual). To do so, we changed the parameter C (Eq (1)), which represents the availability of TfhCs during AM. Again, we use the affinity of the dominant clone at the end of the GCR as a metric of the affinity of antibodies generated. Previously (Fig 2a) we found that for high SDs, antibody affinity was lower than optimal. This was because, for sufficiently high SD, most BCRs internalized one or two antigens and so competition between B cells was restricted. But, when the level of TfhC (C) is even more limiting, even for high SDs, small differences between B cells with regard to the amount of internalized antigens are amplified by the intense competition between B cells for interacting with, and receiving a survival signal from TfhCs (Fig 3a, S9 Fig). Thus, if the SD is high, more potent antibodies are generated as TfhC levels decline. However, when the SD is low, decreasing C does not alter the affinity of the resulting antibodies significantly. Consistent with this result, the probability of selecting the lower affinity B cell decreases as C decreases (see Fig 3b). This is because selection is a stronger force when there are fewer TfhCs. In agreement with our result in Fig 3a, this enhancement in selection forces is more pronounced at intermediate and high SDs compared to low SDs. Thus, as TfhC levels are smaller, high affinity antibodies can be generated more readily for high SDs. Thus, perhaps, HIV, which is the only virus that principally targets T helper cells and depletes their numbers, evolved a low SD to prevent strong antibody responses from developing during the early stages of infection when T helper cell levels rapidly decline. (Note that Human T cell Leukemia viruses also infect T cells, but they do not result in a sharp decline in T helper cell numbers [39]. Their effect on the number and functioning of TfhCs is still unclear [40]).
Proteins that comprise the spikes on the surface of viruses bind to receptors on host cells to propagate infection. A high SD increases the probability of encounters with the host cell’s receptor, thus enhancing infectivity [41]. At the same time, the viral spike is the target of immune attack by antibodies [42] and a high SD may make the virus more susceptible to neutralization. For example, HIV’s low SD can inhibit effective neutralization [3]. The average inter-spike distance for HIV is about 23nm, while the average distance between the arms of the Ab is 15nm. Thus, the two Ab arms are unlikely to bind simultaneously to two Ag molecules, which would decrease avidity [3]. In some viruses like influenza, however, hypervariable features near the head of the spike have high immunogenicity, and a high SD can serve to protect the virus from antibodies that could target more conserved epitopes in the stem of the virus [4]. Thus, there appear to be evolutionary forces that favor both a high and low virus SD. Yet, most viruses have very high SD, and HIV appears to be unique in that its SD is about two orders of magnitude lower (see Table 1).
In order to shed light on the evolutionary forces that may have led to the evolution of a low SD for HIV, we studied a simplified computational model of AM. We found that the affinity of Abs produced by GC reactions is maximal for an intermediate SD (Fig 2a). For very low SD, most B cells internalize a single antigen molecule by binding via a single arm of the BCR or internalize no antigen at all. The occasional B cell that internalizes Ag by chance quickly evolves to become the dominant clone during AM. Thus, for very low SD, fluctuations prevent the system from being in the strong selection regime, thus resulting in lower affinity antibodies. For a high SD, BCRs on B cells are very likely to internalize one or two Ags, and so soon after AM ensues, most B cells internalize quite a bit of antigen. Therefore, there is reduced competition between B cells for obtaining a survival signal from TfhCs, and thus a low driving force to evolve higher affinity BCRs. An intermediate SD results in strong selection forces, and the highest affinity antibodies.
The basis for the optimal affinity at intermediate SD being related to the ability to differentiate between B cells with different affinities during the GC reaction is made clear by another aspect of our results. We show that the non-monotonic dependence of antibody affinity with SD is directly related to the binding cooperativity between the two arms of the BCR. Indeed, for a hypothetical single arm BCR, the affinity of the resulting Abs monotonically decreases with SD (S7 Fig). For normal BCRs with two arms, at the optimal density, the second arm serves to split the B cell population into those who manage to capture the additional Ag molecule while the first arm is still bound, and those who do not. Thus, resulting in strong selection for the B cells that internalize two antigens per BCR.
We thus hypothesize that having two arms may be beneficial for optimal B cell selection in the GC. Obviously, the two arms allow Abs to bind Ag with high avidity. However, the two arms also allow for a more precise differentiation between B cells. B cells usually capture Ag using BCR clusters [26] that can function as a “multiple arms” BCR for the purpose of differentiation. However, perhaps it is most beneficial for the BCR to have two arms (in the context of GC selection) when Ag density is very low.
The probability that a lower affinity B cell is stochastically selected to proliferate in favor of a higher affinity B cell is minimal for the intermediate densities. In other words, the ability of the GC reaction to produce highest affinity Abs at an intermediate SD is because this is the regime where selection effects are strongest. So, viruses could have evolved either a low or high SD to reduce the efficacy of antibodies directed against them. The results in Fig 4 suggest that evolving a low SD would be especially advantageous from this perspective. Furthermore, such a strategy reduces the avidity of antibodies, further inhibiting neutralization [43]. However, most viruses have evolved a high SD (Table 1). Our results suggest that this may be because the evolutionary driving forces of increasing infectivity and shielding conserved epitopes by maintaining a high packing density may be dominant. A very high SD (past the optimal density for AM) can lower antibody affinity somewhat (Fig 4) while favoring these two selection forces.
Why has HIV evolved SD that is roughly two orders of magnitude smaller than that observed for most viruses? HIV is unique in that it principally infects T helper cells and reduces their numbers significantly during the early stages of infection. We suggest that T helper cell depletion during HIV infection results in increased competition between B cells during AM. When T helper cells are limiting to the normal extent, once the SD is higher than the optimal value, selection forces are weaker. But, if T helper cells are significantly depleted, selection forces remain strong at high SDs, resulting in high affinity antibodies. As Fig 3b shows, depletion of TfhCs results in a lower probability of low-affinity B cells stochastically proliferating in favor of higher affinity B cells. However, our results show that, for low SDs, Ab affinity hardly increases upon depletion of TfhCs. This may be the reason that HIV, which is the only known virus that principally infects T helper cells and sharply depletes their numbers during early stages of infection, is the rare virus that has evolved a low SD. The low SD may aid HIV to avoid effective antibody responses in the early stages of infection in a way that would not be possible if it had a high SD. We note also that a low SD can decrease the ability to form diverse clonal lineages during the GC reaction, thus inhibiting AM from deeply exploring the antigenic space upon infection [44]. Because it is a chronic infection and replicates fast, the reduced infectivity of a low SD may be alleviated.
The Measles virus also infects T helper cells; could result in their depletion [45] and causes transient immunosuppression [46]. Unlike HIV, Measles causes acute disease and is rapidly cleared from the body. Notably, Measles has many spikes on its surface [43]. We hypothesize that its rapid mode of propagation resulted in its choice to have high infectivity, producing many virions in the short period during which the patient is infectious. An HIV carrier, on the other hand, is infectious for an extended period. Thus, the virus has to hide for much longer from the immune system and having a low spike density would allow it to do so. Finally, it seems that affinity maturation and antibody response is not the main way by which Measles is cleared. Rather, it induces a T cell response that is dominant in the first two weeks [47]. Only at a much later times (many weeks) affinity maturation starts to produce antibodies against Measles’s RNA and proteins [47]. Another virus that infects T helper cells is HTLV-1. Contrary to HIV, HTLV-1 infection does not appear to result in a sharp decline of the T-cell population but it may impair the function of TfhCs [48]. Our model would suggest that this effect should promote the evolution of a low SD. Another feature of HTLV suggests that a high SD for improving infectivity may be ameliorated; HTLV-1 predominantly infects new cells via cell-to-cell contact [49][50]. It seems that during cell-to-cell transfer, at the intersection between the two cells, the local density of env is high [49]. However, while we could not find precise quantification of their number on the free virions (which is the key variable during affinity maturation), spikes appear sparse on their surface when imaged with cryo-EM [51]. Given this paucity of quantitative data, at this stage, it is not clear how our hypothesis relating spike density and T cell number to the competitive force in the GC should be applied to HTLV-1.
The impact of SD on HIV and SIV propagation has been studied experimentally. A deletion in the tail of gp41 (part of the Env spike protein) has been suggested to increase the number of spikes [41] and their mobility [52]. These mutated virions have better infectivity in cell culture [53]. Surprisingly, the deletion is rarely seen in vivo and when macaques are infected with a short tail SIV mutant, the mutant reverts back to the long tail virus [53–55]. These results may suggest an evolutionary driving force, in the face of an immune response, to reduce SD in HIV.
Our results are reminiscent of the pioneering discovery by Herman Eisen showing that affinity increase upon AM was smaller for a very high Ag dose, suggesting that too high Ag concentration decreases competition in the GC. There is also experimental evidence that intermediate levels of antigen density lead to the highest Ab titers [56], related to optimal BCR activation. In this case, Ag molecules can mediate the formation of a BCR cluster when the density is sufficiently high, while for very high density, Ag sequesters the BCR molecules and the number of fully formed clusters is reduced.
Our results are relevant to vaccine design. It is possible to design liposomal nanoparticles that display varying density of Ag [11,57,58]. We have shown here that more is not necessarily better. We suggest that a precise design of the density of Ag can impact B cell selection in the GC, and as a result on the Ab affinity of the memory and plasma cells that are the product of vaccination. Thus, our results may guide the design of vaccination vectors that could optimize immune responses in the lymph node follicle.
Our arguments regarding the evolutionary driving forces that have led to HIV being unique in exhibiting an extraordinarily low SD may be difficult to examine experimentally (as is the case for most problems in evolutionary biology that are not contrived laboratory curiosities). However, by depleting TfhCs in mice to different degrees, the veracity of the underlying mechanism could be tested.
To account for the limited capacity of a GC during the competitive phase, we employed a variant of the stochastic logistic growth process [59], in which the death rate increases with the overall population size, from a basal rate of μ0, as
μ(n)=(μ0+(β0−μ0)∑i=1MniN).
(6)
Here, N is the population capacity taken to be 200; n = (n1, n2,…,nM) is the vector of cell numbers ni for the M clones such that ∑i=1Mni is the total number of B cells in the GC. The competitive phase lasts about 16 days in mice [31]. The total number of cells in the GC grows gradually until reaching the capacity N, where it remains approximately fixed.
We assume here that an arm of a BCR at the tip of a B cell protrusion has a characteristic time τ to find an Ag molecule, after which the protrusion retracts empty. The probability that an arm of a BCR at the tip of a B cell protrusion finds Ag before time τ is
Pbinding:=Probability{Onearmbindingbeforetimeτ}=(1−e−2qNAgτ),
(7)
where qNAg=qNAg is the on-rate for the BCR to find an Ag molecule given that there are NAg of them in its scanning radius, and the sequence-dependent on-rate with which a particular BCR binds to Ag is q. This leads to
p0=e−2qNAgτ,p1=1−e−2qNAgτ,
(8)
where p0 is the probability that no Ag is encountered, while p1 is the probability that one of the arms encounters an Ag molecule in this time period.
In order to capture the Ag, the BCR attached to a protrusion of the B cell applies a pulling force that works against the interactions of the BCR with Ag, and that between the Ag and the virus/Fc receptor/FDC membrane (see Fig 1). We denote the potential interaction energy required to extract the Ag by EAg−mem, and the interaction energy of the BCR/Ab and Ag by EAg−Ab.
The characteristic rupture time depends on the force applied by the B cell [60]. The force F does work xbF on the bond, reducing its free energy and increasing the rupture rate to rF=r0eβxbF, where r0 is the characteristic rate for bond disassociation when no force is present, xb is the distance at which the bond ruptures, and β = kBT. The intrinsic rupture rate with no force is estimated by Kramer’s escape from a potential barrier as
r0=Ke−βEAg−Ab,
(9)
where K is a coefficient that depends on the shape of the interaction potential [61]. We take K = 1, and note that in writing Eq (9) we have assumed that the activation barrier to form the bond is much smaller than the energy gained upon binding (EAg−Ab). F is typically of the order of a few pico-Newtons [27]. We assume that each B cell has 100 BCRs but the qualitative behavior of our results does not depend on the number of BCRs. As each BCR can extract 0/1/2 Ag molecules, a B cell can extract between 0 and 200 Ag molecules.
We modeled the effect of mutation as a change in the on-rate q or the interaction energy EAg−Ab upon cell division, with one daughter retaining the parent affinity, while for the other daughter:
EAg-Ab,daughter=EAg−Ab,parent+N(0,2D),qdaughter=qparent+N(0,2D)
(10)
Here, N is a normal distribution with zero mean and standard deviation of 2D, with D akin to an effective variability coefficient determining the magnitude of the change in q or EAg−Ab. Within this model, the energy, or q, can increase or decrease with equal probability at every division. We added a reflecting boundary condition at zero such that q is never negative. Since the off-rate scales as r0=e−βEAg−Ab, the affinity of the BCR is calculated as ω = q/r0.
To study the time evolution of the GC reaction, we performed Brownian dynamics simulations. At every time point B cells proliferate or die with a probability which is determined by the time step and the proliferation and death rates (Table 2). We choose the time step to be smaller than the average time for proliferation. The simulation proceeds according to the following algorithm:
|
10.1371/journal.pgen.1003106 | The Tumor Suppressor Gene Retinoblastoma-1 Is Required for Retinotectal Development and Visual Function in Zebrafish | Mutations in the retinoblastoma tumor suppressor gene (rb1) cause both sporadic and familial forms of childhood retinoblastoma. Despite its clinical relevance, the roles of rb1 during normal retinotectal development and function are not well understood. We have identified mutations in the zebrafish space cadet locus that lead to a premature truncation of the rb1 gene, identical to known mutations in sporadic and familial forms of retinoblastoma. In wild-type embryos, axons of early born retinal ganglion cells (RGC) pioneer the retinotectal tract to guide later born RGC axons. In rb1 deficient embryos, these early born RGCs show a delay in cell cycle exit, causing a transient deficit of differentiated RGCs. As a result, later born mutant RGC axons initially fail to exit the retina, resulting in optic nerve hypoplasia. A significant fraction of mutant RGC axons eventually exit the retina, but then frequently project to the incorrect optic tectum. Although rb1 mutants eventually establish basic retinotectal connectivity, behavioral analysis reveals that mutants exhibit deficits in distinct, visually guided behaviors. Thus, our analysis of zebrafish rb1 mutants reveals a previously unknown yet critical role for rb1 during retinotectal tract development and visual function.
| Before an organism can execute necessary behavioral responses to environmental stimuli, the underlying neural circuits that regulate these behaviors must be precisely wired during embryonic development. A properly wired neural circuit is the product of a sophisticated collaboration of multiple genetic pathways that orchestrate cell type specification, the extension and growth of the cell processes that connect each circuit component, and the refinement of these connections. In an unbiased genetic screen designed to identify the genes required for proper circuit formation in developing zebrafish embryos, we identified a human disease causing mutation in the retinoblastoma-1 (rb1) gene that disrupts the formation of the zebrafish visual circuit. rb1 canonically functions to regulate the cell cycle, and when mutated the loss of rb1-mediated cell cycle control elicits childhood ocular tumor formation. Genetic models of rb1 have been developed to study the developmental role of rb1 in the retina; however, ectopic cell proliferation and death within the retina have largely precluded the ability to evaluate the formation and integrity of neural circuits connecting the retina with the brain. In this study, through genetic and cellular analysis of a zebrafish rb1 mutant, we reveal a novel role for rb1 in regulating the establishment and functionality of the visual circuitry.
| Biallelic mutations in the retinoblastoma susceptibility gene rb1 are causal for intraocular childhood retinoblastomas. Rb1 is a member of a gene family that consists of three members, p105/Rb1, p107/Rb-like1, and p130/Rb-like2, collectively known as “pocket proteins” [1]. The activity of these proteins is controlled, in part, by cyclin/cyclin-dependant kinase complexes. Upon activation, Rb proteins bind to an array of proteins, including members of the E2F family of transcription factors to execute a range of cellular functions, including cell cycle exit, terminal differentiation, and cortical cell migration [2]. In humans, germline or somatic mutations occur throughout the 180 kb genomic region spanning the rb1 gene, including its promoter region, exons, and intronic essential splice sites, resulting in bilateral or unilateral retinoblastomas within the first 2 years of life [3], [4].
Given its clinical relevance, the role of Rb1 during embryonic development and during tumor suppression has been studied intensely, mostly using mouse models [5]. Rb1 is expressed ubiquitously during murine development, postnatally, and continues to be expressed in adults. Embryos harboring non-conditional Rb1 knockout alleles exhibit ectopic proliferation and apoptosis throughout the nervous system and die prenatally at embryonic day 14.5 [6], [7], [8]. Embryos with conditional loss of Rb1 in the retina display ectopic division and considerable apoptosis of retinal transition cells starting at E10 [9], [10], [11], [12]. Retinas in these animals contain reduced numbers of rods, bipolar cells, and RGCs, yielding a retina with a thin outer nuclear layer and a hypoplastic optic nerve. However, the etiology of optic nerve hypoplasia and if/how Rb1 functions in RGC axonal guidance has not been examined. Similarly, electroretinogram recordings from Rb1 deficient mouse retinas have revealed reduced photoreceptor to bipolar to amacrine signal transmission [10], yet the behavioral consequences have not been examined.
Here, we report that zebrafish space cadet mutants carry a rb1 mutation found in cases of sporadic and familial human retinoblastoma [13], [14], [15], [16]. In zebrafish rb1 mutants, RGC precursors show delayed exit from the cell cycle and hence a delay in the generation of early-born, postmitotic RGCs, whose axons are critical for pioneering the retinotectal tract. This delay leads to RGC intrinsic axon guidance defects, aberrant retinotectal connectivity, and deficits in phototactic behaviors. Together, this work describes a novel model for understanding the developmental role of rb1 and reveals a previously unknown role for rb1 in the formation of the retinotectal tract.
We previously identified two mutant space cadet alleles, based on abnormal startle response behavior to acoustic or tactile stimuli [17], [18]. Using recombination mapping, we mapped the space cadette226a allele to a 1.1 cM interval on chromosome 21 between single nucleotide polymorphic markers in the myo5b locus (20 recombinants/2688 meioses), and in the ncam1 locus (12 recombinants/2688 meioses), respectively (Figure 1A). This genomic interval contains several annotated genes, including rb1, lpar6, and cystlr2, which have retained syntenic positional conservation between humans, mice and zebrafish (Figure 1B). Sequencing of rb1 cDNAs isolated from spcte226a larvae revealed the presence of 4 nucleotides inserted between exon 19 and exon 20. Subsequent sequencing of genomic DNA isolated from spcte226a larvae confirmed a single nucleotide change in the splice donor sequence of intron 19 (nt1912+1: G to A; Figure 1C). This generates a cryptic splice site donor, resulting in the 4 base pair insertion into the rb1 mRNA. This 4 base pair insertion causes a premature stop codon in exon 20, predicted to truncate the protein at amino acid 677, thereby severely truncating the B domain and cyclin domain essential for Rb1 function (Figure 1D) [15]. Interestingly, identical mutations have been reported in human patients with familial and sporadic forms of retinoblastoma [13], [14], [15], [16].
The zebrafish rb1 gene is 67% similar (52% identical-based on amino acid sequence) to the mouse and human rb1. The Rb1 protein consists of an A and B domain forming Rb1's binding “pocket”, and a cyclin binding domain (Figure 1D), and shows 81% amino acid similarity (66% identical) between zebrafish and mammalian rb1 in these critical domains. Thus, sequence homology, syntenic conservation, and cDNA sequence analysis provide compelling evidence that space cadet phenotypes are caused by an rb1 gene mutation known to cause retinoblastoma. Sequence analysis of the second mutant space cadetty85d allele did not reveal any changes in the coding sequence or in any of the splice donor and acceptor sites, suggesting that this allele is caused by a regulatory mutation in the rb1 locus. Importantly, analyses of spcte226a and spcty85d mutants revealed no significant differences with regards to the strength of the phenotypes examined below (Table 1). From here on, we will refer to the space cadet gene as rb1 and describe anatomical and behavioral defects in the rb1te226a allele.
During zebrafish embryogenesis, the earliest born RGCs begin extending axons at 32 hpf, cross the ventral midline of the diencephalon to form the optic chiasm at 36 hpf, and project dorsally to the contralateral optic tectum by 48 h to form a retinotectal pathway critical for mediating visually guided behaviors by 120 hpf [19]. We previously reported that 120 hpf stage rb1 mutants display wild type like retinal lamination and expression of terminal RGC cell differentiation markers, but exhibit various RGC axonal pathfinding defects [18]. To determine the temporal onset and spatial site of RGC pathfinding errors in rb1 mutant embryos, we used the ath5:gfp transgenic line expressed in RGCs to examine the development of the retinotectal trajectory [20]. Importantly, between 28–96 hpf we did not detect a difference in the intensity of GFP fluorescence in the retinas of rb1 mutant compared to wild type retinas (Figure S1 and data not shown). At 36 hpf wild type RGCs have exited the retina and pioneered across the ventral midline to form an optic chiasm (n = 40, Figure 2A). In contrast, only 13% (n = 30) of rb1 mutant retinas had RGC axons that exited the retina (Figure 2B, 2C), suggesting that the loss of rb1 function causes a delay in the initial outgrowth of RGC axons from the retina. At 48 hpf the optic nerve in rb1 mutants was significantly thinner, with a mean diameter of 3.04 µm (n = 18), compared to the thicker optic nerves in wild type siblings (13.76 µm, n = 22; Figure 2D–2F). At 72 and 96 hpf, when wild type RGC axons have reached and innervated the optic tectum, rb1 mutant tecta show a significant reduction in RGC axon tectal innervation (Figure 2G–2L; see Methods for quantification details). Together, these results reveal that rb1 mutants exhibit a delay in RGC axonal outgrowth, leading to a delay in optic nerve development, and reduced innervation of the optic tectum.
To determine whether the identified mutation in rb1te226a is causative of the delay in retinotectal development, we injected wild type rb1 mRNA into one-cell stage rb1te226a mutants and examined optic nerve diameter at 48 hpf. Microinjection of wild type rb1 mRNA restored optic nerve diameter in rb1 deficient mutants in a dose dependent manner, demonstrating that mutations in zebrafish rb1 cause RGC outgrowth defects (Figure 3A–3C, 3E). To determine if and to which extent the mutant rb1te226a allele has retained biological activity, we examined the ability of rb1te226a mRNA to rescue retinotectal development in rb1te226a mutants (Figure 3D–3E). Injection of rb1te226a mRNA failed to significantly increase optic nerve diameter in rb1te226a mutants, suggesting that the rb1te226a protein product has very limited, if any, functionality. However, we cannot exclude the possibility that the mutant phenotype is ameliorated by maternal rb1 mRNA and/or protein deposition.
Finally, we asked whether Rb1 functions within RGCs for their axons to exit from the retina and enter the retinotectal path. Because zebrafish rb1 is expressed ubiquitously throughout development (Figure 4A–4B), we generated chimeric embryos by transplanting cells at the blastula stage between rb1 mutant and wild type embryos, and then examined their ability to exit from the retina (Figure 4C). A significant fraction of axons from genotypically mutant rb1 RGCs transplanted into wild type hosts failed to exit from the retina (19% of retinas showed failure of transplanted rb1 mutant RGC axons to exit, n = 31, Figure 4E–4F), consistent with the low but significant frequency of rb1 mutant retinas in which we observed a complete failure of RGCs to exit from the eye (11%; see below). Conversely, 100% of rb1 mutant retinas showed exit of axons from transplanted wild type RGCs (n = 69, Figure 4D). Thus, during zebrafish development rb1 acts RGC autonomously for axons to exit the retina and to form the optic nerve.
Given the RGC intrinsic defects observed in rb1 mutants, we next wanted to determine the primary defect leading to the delay of RGC axons to exit from the retina. Rb1 canonically functions to regulate cell cycle checkpoints, promoting cell cycle exit and differentiation of progenitors and suppressing cell cycle re-entry of differentiated cells [2]. In the retina, rb1 has been shown to promote the exit of retinal progenitor cells from the cell cycle into the various postmitotic cell types that populate the retinal lamina [9], [10], [11], [12]. To examine rb1 deficient retinas for cell cycle defects, we labeled wild type and rb1 mutant retinas for M-phase positive nuclei with an anti-phosphohistone-H3 antibody (anti-pH3) during the initial phase of RGC birth and axon outgrowth, between 28 and 36 hpf. During this time window, premitotic ath5 positive retinal progenitors divide, with one daughter becoming a postmitotic RGC and the other maintaining its progenitor potency to give rise to other retinal cell types that become postmitotic at later stages of development [21]. Although the total number of M-phase positive increased with time between 28 and 36 hpf in rb1 mutant and wild type retinas, we observed fewer M-phase positive nuclei in rb1 deficient retinas, compared to wild type retinas, at each time point examined (Figure 5).
One possibility is that the reduction of M-phase retinal precursors in rb1 deficient retinas is due to increased cell death. Indeed, compared to wild type retinas, rb1 deficient retinas showed a slight, but significant increase of TUNEL positive nuclei between 28 and 36 hpf (Figure S1). Importantly though, comparing the increased number of TUNEL positive nuclei to the decreased number of pH3 positive nuclei in rb1 mutants at 28, 32, and 36 hpf revealed that apoptosis accounts for only 18–26% of the observed reduction in M-phase positive retinal precursors in rb1 mutant retinas at each time point examined. This suggests that in rb1 mutant retinas cell death contributes only partially to the deficiency of M-phase positive nuclei (Figure S1G). Thus, the reduction in M-phase retinal precursors in rb1 mutant retinas suggests a prolonged terminal cell cycle for the retinal precursors, which need to exit their final cycle to become the earliest population of postmitotic RGCs.
To determine if loss of rb1 function indeed causes an initial delay in the presence of postmitoic, differentiated RGCs, we examined expression of isl2b-gfp, one of the earliest transgenic markers indicative for postmitotic RGCs [22]. We found that in wild type retinas postmitotic, differentiating RGCs marked by isl2b-gfp expression emerged first at 32 hpf, increased significantly in their abundance by 36 hpf, and by 48 hpf isl2b-gfp positive RGCs were densely packed throughout the ganglion cell layer (Figure 6A, 6C, 6E, n = 25, 16, and 29, respectively). In contrast, isl2b-gfp positive RGCs were present in only 13% of rb1 deficient retinas at 32 hpf (n = 23). Because of the cytoplasmic localization of the GFP signal and the density at which RGCs normally populate the ganglion cell layer, it is difficult to determine the total number of isl2b-gfp positive RGCs. Nonetheless, semi-quantitative analysis revealed that by 36 hpf, isl2b-gfp positive RGCs were present in 90% of rb1 mutant retinas (n = 29); however, their distribution with the retina was more similar to that of younger wild type retinas at 32 hpf (Figure 6B, 6D). By 48 hpf, all rb1 mutant retinas harbored isl2b-gfp neurons (n = 32), although differentiation still appeared to lag in 81% (n = 32) of the rb1 mutant retinas compared to the more densely packed ganglion cell layer in wild type retinas (n = 29, Figure 6E–6F). Despite the reduced number of RGCs present at 48 hpf, rb1 mutant isl2b-gfp positive RGCs express DM-GRASP, a late marker of RGC differentiation, demonstrating that mutant RGCs were fully differentiated once becoming postmitotic (Figure S2). Importantly, the number of ath5-gfp positive RGC precursors was unaffected in rb1 mutants (Figure S1 and data not shown). Thus, the rb1 deficiency causes a delay in the transition of RGC precursors to postmitotic RGCs, but not in the specification of RGC precursors.
Aside from the reduced population of early born RGCs, rb1 mutant retinas appear grossly normal, and at 120 hpf, show proper lamination by each retinal cell type [18]. Although we did not determine whether birth dating of other retinal cell types is affected in rb1 mutant retinas, netrin-positive exit glial cells and Muller glia cells are present in appropriate numbers and location, indistinguishable from wild-type retinas (Figure S2). Taken together, these results suggest that a delay in cell cycle exit by rb1 deficient RGC precursors leads to a transient reduction in the early born postmitotic RGCs without consequence to the gross morphology and overall cellular landscape of the rb1 mutant retina.
The early born RGCs are located within the central retina and pioneer the retinotectal tract to the contralateral optic tectum [22]. In the absence of the early pioneering RGC axons, the axons of later born, more peripherally located RGCs fail to exit the eye and project aberrantly within the retina [22]. Given the reduced number of these early born, central RGCs in rb1 mutants, we sought to determine whether peripheral RGC axon trajectories were affected. For this, we labeled small groups of RGCs in the anterior peripheral retina with DiO (green) and in the posterior peripheral retina with DiI (red, Figure 7A–7F). In 120 hpf larvae wild type larvae, all labeled axons from anterior and posterior RGCs fasciculated shortly after sprouting from their soma and extended as an axon bundle, forming a path directly toward the retinal exit point (n = 83, Figure 7A–7B, 7D–7E). In contrast, 91% (n = 65) of rb1 deficient retinas harbored a significant subset of axons that had extended aberrantly throughout the retina and failed to exit (Figure 7C, 7F). These results demonstrate that the delayed differentiation of the early born RGCs in rb1 mutants impairs the ability of later born RGC axons to exit the retina.
The delayed cell cycle exit and differentiation of pioneering RGCs lacking rb1 may also affect axon navigation by later born RGC axons at key choice points: the ventral midline of the diencephalon and/or the optic tectum. To examine these possibilities, we filled the RGC layer of the left and right eyes of wild type and rb1 mutant larvae with either DiI or DiO, respectively (Figure 7G). In wild type siblings, 99% of dye filled optic nerves projected to their appropriate contralateral tectum (n = 946, Figure 7H). In contrast, rb1 mutant optic nerves displayed a variety of phenotypes. The majority of rb1 deficient optic nerves were significantly thinner than their wild type counterparts (37%, n = 663, Figure 7I–7J, 7L), consistent with what we observed in with ath5:gfp (Figure 2). In a significant portion of rb1 deficient optic nerves, 17%, RGCs projected to both the contralateral but also to the ipsilateral tectum, indicative of midline pathfinding defects (n = 663, Figure 7K–7L). Focal DiI/DiO labeling of RGC axons arising from the anterior and posterior retina revealed that retinotopic mapping, a function of retinal cell body location [23], remains intact in rb1 mutants despite the aberrant pathfinding en route to the optic tectum (Figure S3). Finally, in 11% of rb1 mutant retinas, there was a complete failure of RGCs to exit from the eye, even at 120 hpf (Figure 7J). Taken together, these results suggest that rb1 deficient RGC axons make intraretinal and midline pathfinding errors, leading to reduced and incorrect tectal innervation.
By 120 hpf, zebrafish larvae perform an array of sensorimotor behaviors, including responses to visual stimulation. For example, changes in visual field illumination, such as the sudden absence of light or a shift from uniform to focal illumination, elicit specific, stereotyped turning behaviors [24], [25]. We first examined the ability of rb1 mutant larvae to perform positive phototaxis, defined as navigating toward a target light source that is presented after extinguishing the pre-adapted uniform light field [25]. Positive phototactic navigation is characterized by larvae first turning towards the target light source and then swimming forward towards the target. As previously reported, when presented with a target light source wild type larvae facing away from the light target show significant initiation of turns, which are preferentially biased towards the light target (Figure 8A–8B). Once facing the target, wild type larvae initiate forward scoot swims (Figure 8C). In contrast, turn initiation in rb1 mutant larvae facing away from the light target was dramatically reduced (Figure 8A). On the few occasions when they initiated a turn, turning direction was unbiased with respect to the light target (Figure 8B). Moreover, rb1 mutants facing the light target did not show an increase in forward scoot swim initiation above baseline (Figure 8C). To further determine whether rb1 mutants respond to more extreme changes in illumination, we examined their ability to perform an O-bend response to a visual dark flash stimulus, a sudden extinction of light [24]. Again, compared to their wild type siblings, rb1 mutants displayed a minimal O-bend response to dark flash stimulation (Figure 8D). Despite their impaired visual responses, rb1 mutants showed no difference in the spontaneous initiation of turning or swimming behaviors compared to wild type siblings (Figure 8D). Importantly, the kinematic parameters of spontaneously occurring turning and swimming movements were indistinguishable between rb1 mutants and their wild type siblings, demonstrating that the neural circuits required for initiation and execution of turning behaviors are largely intact in rb1 mutants. Together, these results demonstrate that rb1 mutants exhibit visual deficits.
Children with biallelic germline or sporadic inactivation of rb1 are likely to form ocular tumors during early childhood. Initially, the retinas of affected individual show an otherwise grossly normal morphology. In contrast, even conditional rb1 knockout mouse models exhibit ectopic proliferation and cell death leading to significant morphological defects throughout affected retinas. We find that inactivation of the zebrafish rb1 gene through a rb1 causing mutation results in mutant retinas that display very limited signs of cell death, with differentiated retinal cell types that are properly laminated, similarly to childhood retinas lacking rb1. Thus, the fairly ‘normal’ retinal landscape of zebrafish rb1te226a mutants provided us with a unique opportunity to investigate if and how rb1 is required to establish the retinotectal projection. Our analysis reveals a RGC autonomous requirement for rb1 in regulating RGC axon pathfinding within the retina and at presumptive choice points en route to the optic tectum. Moreover, we demonstrate that zebrafish rb1te226a mutants exhibit deficits in visually guided behaviors, suggesting that the retinotectal path defects in rb1 mutants may be sufficient to impair vision. Together, this work reveals a novel role for rb1 in the establishment of RGC axon projections during development and establishes a unique model for understanding the developmental and tumor suppressor roles of the rb1 gene.
Zebrafish rb1te226a mutants harbor a human retinoblastoma causing rb1 gene mutation. The mutant protein is truncated in the B-domain and lacks the cyclin-binding domain, reducing Rb1's capacity to form a ‘pocket’, and reducing its capacity for phosphorylation by cyclin dependent kinases [2], [26]. Consistent with the notion that the rb1te226a mutant allele is largely non-functional, mRNA over-expression in rb1 mutants does not ameliorate the rb1 mutant phenotype. Despite the absence of biological activity of the truncated rb1te226a protein, mutant zebrafish show a significantly milder retinal phenotype compared to conditional or even germline rb1 mouse knockouts [6], [7], [8], [9], [10], [11], [12]. One possible explanation is the strong maternal contribution of rb1 in zebrafish (Figure 4A), which may suppress phenotypic expressivity at early stages of development. Consistent with this idea, formation of the initial scaffold of axon tracts during the first day of development appears unaffected in rb1te226a mutants, yet visual and hindbrain pathways that develop after the first day of development show defects [18].
In humans the rb1 nt1960+1 mutation, which is identical to the zebrafish rb1te226a mutation, causes ocular tumors [13], [14], [15], [16], raising the possibility that zebrafish rb1 mutants might also develop tumors as juveniles. However, rb1 mutants fail to inflate a functional swim bladder, and die ∼7 days post fertilization, precluding the analysis of ocular tumors in juveniles. Although wild type rb1 mRNA injection rescues the early RGC and retinotectal defects through 48 hpf, these transiently rescued rb1te226a mutants do not survive beyond 7 days of development, indicating that rb1 plays a critical role after the injected mRNA has been degraded. Establishing stable, inducible rb1 transgenic lines to rescue developmental deficits will therefore be required to monitor juvenile and adult zebrafish for retinal tumors.
In rb1te226a mutants, a significant subset of RGC axons fail to exit the retina, and many of the exiting axons then project incorrectly to the ipsilateral tectum, revealing a previously unrecognized requirement for rb1 in regulating axon pathfinding. One possible explanation for the RGC guidance defects is that in rb1te226a mutant RGCs the expression of guidance factors might be disrupted. Interestingly, cortical cell migration was shown to be dependent on rb1 regulated neogenin expression [27], suggesting that rb1 deficient RGC axons might lack guidance factors required to navigate towards the retinal exit point and properly cross the ventral midline. To investigate this possibility, we performed microarray gene expression analysis of the retina and brains of 32 hpf rb1te226a mutants (MAW and MG, unpublished). However, this approach did not reveal a significant change in the expression levels of neogenin or other known axon guidance genes. Although it remains possible that rb1 regulates expression of un-identified guidance factors, it is more likely that rb1 regulates axon pathfinding indirectly by ensuring the timely exit of RGC precursors from the cell cycle and hence the appropriate temporal appearance of differentiated RGCs. In fact, genetic ablation of the earliest born RGCs prevents the formation of the retinotectal tract [22]. This suggests that RGC birth order imprints a critical hierarchical pathfinding role on RGC axons, such that axons from the earliest born RGCs pioneer the retinotectal tract that later born RGC axons will follow [22]. The delayed onset of RGC birth in rb1te226a mutants may therefore reduce the population of pioneering RGCs present during a restricted window of environmentally expressed guidance factors.
Zebrafish rb1te226a mutants display deficits in the acoustic startle response [17], [18] and in visually guided behaviors, reflecting the importance of rb1 function for the development of neural circuitry underlying behavior. The deficits in startle behavior are due to defects in a small subset of hindbrain neurons, the spiral fiber neurons [18], and giving the results presented here, it is tempting to speculate that rb1 plays a similar role for the transition of these neurons from precursors to postmitotic neurons. Unfortunately, markers that follow the development of spiral fiber neurons are not available, precluding such analysis. Therefore, we focused on the well-characterized development of RGCs. Although we demonstrate a defect in the early development of these cells and their axonal connectivity, we cannot exclude the possibility that zebrafish rb1 mutants exhibit defects in the development and/or function of other retinal cell types, and that these defects contribute to the deficits in visual behaviors we observe. Future analysis of transgenic lines expressing the wild type Rb1 gene in individual retinal cell types will reveal which cell type(s) and connections are causative of the visual deficit.
In summary, we report a zebrafish mutant carrying a human disease causing rb1 mutation, which reveals novel roles of rb1 in regulating RGC axon pathfinding and visually guided motor behavior. Furthermore, these mutants provide a non-murine vertebrate model of rb1 and offer new potential for identifying the elusive retinoblastoma cell of origin and further insight into the developmental role of rb1.
All experiments were conducted according to an Animal Protocol fully approved by the University of Pennsylvania Institutional Animal Care and Use Committee (IACUC) on January 27, 2011, protocol number 803446. Veterinary care is under the supervision of the University Laboratory Animal Resources (ULAR) of the University of Pennsylvania.
The zebrafish (Danio rerio) strain used in this study was the spcte226a allele (now referred to as rb1te226a) of space cadet [17], [18], maintained on a mixed TLF and Tubingen background. The rb1te226a allele was also crossed into the ath5:gfp and isl2b:gfp transgenic backgrounds for RGC analysis [20], [22]. rb1te226a+/−;ath5:gfp+/− or rb1te226a+/−;isl2b:gfp+/− adults were always crossed with rb1te226a+/−;TLF adults to ensure that rb1te226a embryos analyzed for GFP-expressing RGCs were hemizygous for GFP. Throughout the manuscript, rb1−/−, “rb1 deficient”, and rb1 mutant refers to rb1te226a homozygotes. The other space cadet allele spcty85d [17] was only used where mentioned. Embryos were collected in the morning, maintained on a 14/10 hour light/dark cycle at 28°C, and staged as described previously [28]. Larvae were raised in 6 cm plastic Petri dishes at a density of 20–30 per 7 mL in E3 medium (5 mM NaCl, 0.17 m mM KCl, 0.33 mM CaCl2, 0.33 mM MgSO4) with medium changes at 48 hpf (hours post fertilization) and 96 hpf. Behavioral experiments were conducted on 120 hpf larvae.
A three generation mapping cross between rb1te226a heterozygous and WIK fish was generated, and pools of 25 F2 mutant and F2 sibling 5 dpf larvae were collected in the F2 generation and used for bulk segregant mapping (see Table 2 for simple sequence length and single nucleotide polymorphic markers). Mutant larvae were identified by performing successive, unilateral C-bends to acoustic or tactile stimulation [17], [18]. To identify the mutation, cDNA was prepared following total mRNA extraction from 5 dpf larvae as previously described [29]. rb1 cDNA was amplified with primers (rb1:1–6, Table 2) designed against overlapping regions of the rb1 reference sequence (Ensembl) with the following RT-PCR conditions: 94°C for 3 min and then 40 cycles of 94°C for 45 sec, 57°C for 1 min, and 70°C for 1 min. Products were gel purified and cloned into the pCR2.1-TOPO-TA vector for sequencing. After detecting a frameshift and 4 nucleotide addition to the end of exon 19 in rb1te226a cDNA clones, gDNA was isolated from 5 dpf larvae, and intron 19 was amplified with the rb1:8 primers, using identical PCR conditions to those described above.
For rb1 RNA injection, cDNA was prepared from genotyped homozygous wild type or rb1te226a mutant 5 dpf larvae (dCAPS protocol, see below) and amplified with the rb1:FL primers (similar conditions as above, except extension time increased to 3 min), which includes the coding region of rb1, and cloned into the pCS2+ vector. Wild type rb1 and rb1te226a mRNA was prepared using the mMessage mMachine kit (Ambion, NY) and injected at the 1-cell stage at doses ranging from 1–100 picograms. Embryos injected with 20 or greater picograms of rb1 mRNA showed gross morphological abnormalities and necrosis, whereas embryos injected with 10 picograms or less appeared morphologically normal.
To genotype rb1te226a embryos, we developed a dCAPS assay [30] using the dCAPS program (http://helix.wustl.edu/dcaps/dcaps.html) to design appropriate primers (Table 2). After gDNA isolation, PCR was performed as described above. The PCR product is then digested with SspI (New England Biolabs, Ipswich, MA), cleaving the rb1te226a allele and producing a 120 bp fragment that can be distinguished from the 150 bp wild type allele on a 3% agarose gel containing 1.5% Metaphor agarose (Lonza, Rockland, ME). All genotyping, except for BrdU labeled embryos, was performed following immunolabeling experiments.
For immunostaining, embryos were fixed in 4% paraformaldehyde (PFA) overnight at 4°C, permeabilized with 1 mg/mL collagenase, and blocked for 1 hour with 5% normal goat serum in 0.1 M phosphate buffer. Embryos were then incubated in the primary antibodies anti-GFP (1∶200 mouse JL8, Clontech, Mountain View, CA or 1∶500 rabbit, Invitrogen, Carlsbad, CA), anti-phosphohistone-H3 (Millipore, Charlottesville, VA), 1∶100 anti-BrdU (Roche, Branchburg, NJ), and/or 1∶50 A2-J-22 polyclonal antisera (recognizes carbonic anhydrase II, kindly provided by Dr. P. Linser) overnight at 4°C in blocking solution, washed out, and then detected by the addition of AlexaFluor488 or AlexaFluor594 conjugated secondary antibodies (1∶500, Invitrogen, Carlsbad, CA). TUNEL assay was performed as previously described [31] using Apoptag Peroxidase In Situ Apoptosis Detection Kit (Chemicon, Temecula, CA). After staining, samples were mounted in DAPI containing Vectashield (Vector Labs, Burlingame, CA). Images were acquired with a Zeiss 710 confocal laser scanning microscope (LSM 710) using ZEN2010 software.
For in situ hybridization, digoxygenin-UTP labeled antisense riboprobes for rb1 were synthesized and hydrolyzed from the full length rb1 cDNA construct [32]. Whole-mount in situ hybridization was performed as described previously [33]. Images were acquired with a Zeiss Axioskop compound microscope. For RT-PCR based expression analysis, the rb1:FL and B-actin primers (Table 2) were run against cDNA prepared from total mRNA extracted from 25 embryos/larva at each stage.
120 hpf larvae were anesthetized (0.01% Tricaine) and fixed in 4% paraformaldehyde at 4°C overnight. Larvae were removed from fix, washed briefly in phosphate buffered saline (PBS), and mounted dorsal side up for whole retinal injection or laterally for discreet RGC labeling on glass microscope slides in a bed of 1.5% agarose. To label all RGCs, the vitreal space of each eye was filled with either of the fluorescent lipophilic dyes DiI (red) or DiO (green) (Molecular Probes, Eugene, OR) dissolved in 1% chloroform, using a WPI PV820 picopump injector fitted with a glass micropipette. For discreet labeling, a small region of the exposed eye was labeled with pulses of DiI/DiO dissolved in 0.5% dimethylformamide. Injected larvae were kept moist with PBS and incubated overnight at room temperature in a humidity chamber in darkness. Larvae were then examined for phenotype analysis using a Zeiss Axioplan compound fluorescent microscope. Eyes were carefully removed from selected representative larvae, which were then remounted on coverslips in agarose for imaging. Images were recorded using a Zeiss 510 confocal laser scanning microscope (LSM510) and Zeiss LSM510 analytic software.
For transplant direction wild type donor into space cadet host, wild type transgenic Tg(ath5:gfp) and rb1te226a heterozygous fish were used to generate wild type GFP expressing donor embryos and non-GFP expressing rb1te226a mutant embryos, respectively. For transplant direction space cadet donor into wild type host, rb1te226a; Tg(ath5:gfp) double heterozygotes and either TU or TLF strain wild type mating pairs were used to generate rb1te226 GFP expressing donor embryos and non-GFP expressing wild type embryos, respectively. Once the appropriate donor-host embryos were collected, embryos were immediately placed in E3 medium and kept at room temperature. Donor embryos were pressure injected into the yolk sac at the 1–2 cell stage with the lineage tracer tetramethylrhodamine dextran, 3 Kd, 5% w/v (Molecular Probes, Eugene, OR) dissolved in 0.2 M KCL and filter sterilized. Donor and host embryos were then incubated at 28.5°C in E3 medium in darkness to grow synchronously to the 1000 cell stage. Embryos were then transferred into room temperature complete E2 medium (E2) to retard growth, and dechorionated using Pronase (1∶50 in E2 of 30 mg/ml stock, Roche) in glass 60 mm petri dishes. Dechorionated embryos were washed extensively with E2, transferred using a fire polished glass Pasteur pipette into individual wells in a transplantation dish containing E2, and properly oriented. Transplantation needles were made using #1BBL No Fil borosilicate glass pipettes (WPI), pulled to produce fine tips in a P87 pipette puller (Sutter Instruments, Novato, CA), broken at various diameter openings, and polished using a microforge. Needles were then inserted into a standard pipette holder connected to a modified manual injection apparatus, and mounted in a micromanipulator arm for precision control. Thirty to fifty blastomeres were carefully removed from the donor embryo using the transplantation pipette/manual injector apparatus, and transferred into the adjacent host embryo at the apex of the animal pole (eye/nose region). Operated embryos were maintained in the transplantation dish wells in E2 at 28.5°C in darkness following transplantation, and were allowed to develop undisturbed until epiboly completed. Embryos were then transferred from the transplantation wells/dish into either separate 1.5% agarose coated 60 mm plastic Petri dishes for donors and hosts, or 1.5% agarose coated wells in 12-well tissue culture plates as host-donor pairs, depending on the direction of the transplant, and incubated at 28.5°C for five days. The later was necessary in order to correctly identify rb1te226a donors from each donor-host pair, as the motility phenotype does not manifest itself until 120 hpf. 120 hpf larvae were screened for the presence of GFP expressing RGC clones using a Leica MZFIII fluorescence stereomicroscope, and further analyzed for misprojecting RGC axons using a Zeiss axioplan compound fluorescence microscope. Host larvae suspected of containing misprojecting RGC axons (ie, not exiting the eye, or midline defects) were then fixed and stained with anti-GFP antibody as described above, and imaged using a Zeiss LSM510 microscope and Zeiss LSM510 analytic software. Confocal z-stacks were of sufficient depth (150–220 µm) to insure optic nerves were not inadvertently missed.
Confocal stacks were processed into maximum and/or summation intensity projections using ImageJ for quantification. We used the full width half maximum algorithm to calculate optic nerve diameter from maximum intensity projections of GFP-labeled retinal ganglion cell axons. Tectal innervation was determined by making 20 µm summation projections of GFP labeled tecta, tracing the area of the labeled tectum to determine the Raw Integrated Density (RID) per µm2, and subtracting the RID/µm2 of an unlabeled, background region. TUNEL and anti-pH3 labeled nuclei were counted from 30 µm stacks using Volocity (PerkinElmer, Waltham, MA), with individual cells distinguished by fluorescent intensity and size. Statistical analysis was performed on all data using the Graphpad prism software (www.graphpad.com).
Behavioral experiments were performed on 120 hpf larvae and analyzed with the FLOTE software package as previously described [25], [34], [35]. rb1te226a and wild type siblings were identified based on acoustic startle behavior [17], [18] and then grouped by phenotype for visual behavior testing in 6 cM petri dishes at a density of 12 fish per dish. For all behavioral experiments, N = 48 rb1te226a and 48 wild type sibling larvae. For phototaxis experiments, video recordings were triggered every 500 msec, with each recording covering a 400 msec time window, for a total duration of 4 sec of recorded behavior. Each group of 12 larva were subjected to 3 rounds of phototaxis testing, with 3 min between trials. Orientation of larvae to target light was determined at the beginning of each 400 ms recording as previously described [25], such that the behavior of each larva was tested multiple times and in different orientations with respect to the target light. Therefore, the N for Figure 8A and 8C ranged from 75 to 547 for wild type siblings and 136 to 311 for rb1te226a larvae. In Figure 8B, the N ranged from 39 to 106 for wild type siblings and 13–33 for rb1te226a larvae. For dark flash response experiments, N = 4 groups of 12 larvae. Spontaneous behavior was analyzed on individually housed larvae on a 4×4 grid array.
|
10.1371/journal.pntd.0005165 | Comparison of the Estimated Incidence of Acute Leptospirosis in the Kilimanjaro Region of Tanzania between 2007–08 and 2012–14 | The sole report of annual leptospirosis incidence in continental Africa of 75–102 cases per 100,000 population is from a study performed in August 2007 through September 2008 in the Kilimanjaro Region of Tanzania. To evaluate the stability of this estimate over time, we estimated the incidence of acute leptospirosis in Kilimanjaro Region, northern Tanzania for the time period 2012–2014.
Leptospirosis cases were identified among febrile patients at two sentinel hospitals in the Kilimanjaro Region. Leptospirosis was diagnosed by serum microscopic agglutination testing using a panel of 20 Leptospira serovars belonging to 17 separate serogroups. Serum was taken at enrolment and patients were asked to return 4–6 weeks later to provide convalescent serum. Confirmed cases required a 4-fold rise in titre and probable cases required a single titre of ≥800. Findings from a healthcare utilisation survey were used to estimate multipliers to adjust for cases not seen at sentinel hospitals. We identified 19 (1.7%) confirmed or probable cases among 1,115 patients who presented with a febrile illness. Of cases, the predominant reactive serogroups were Australis 8 (42.1%), Sejroe 3 (15.8%), Grippotyphosa 2 (10.5%), Icterohaemorrhagiae 2 (10.5%), Pyrogenes 2 (10.5%), Djasiman 1 (5.3%), Tarassovi 1 (5.3%). We estimated that the annual incidence of leptospirosis was 11–18 cases per 100,000 population. This was a significantly lower incidence than 2007–08 (p<0.001).
We estimated a much lower incidence of acute leptospirosis than previously, with a notable absence of cases due to the previously predominant serogroup Mini. Our findings indicate a dynamic epidemiology of leptospirosis in this area and highlight the value of multi-year surveillance to understand leptospirosis epidemiology.
| Leptospirosis is an infectious disease that causes a fever. It can be severe or fatal. Understanding how many people get leptospirosis helps to determine priorities in allocating resources for disease diagnosis, treatment, and prevention. There are few data about leptospirosis incidence in sub-Saharan African countries. The only mainland estimate is from northern Tanzania for the years 2007–08. To see if leptospirosis incidence had changed since 2007–08, we measured leptospirosis incidence in the same location in 2012–2014. To do this, we systematically approached people at two hospitals in the Kilimanjaro Region and tested them for leptospirosis. We adjusted the number of identified cases of leptospirosis found at the hospitals to account for people with fever who did not come to hospital for testing and care. We also adjusted for imperfect testing methods. We found that the number of people who developed leptospirosis annually had dropped from 75–102 cases per 100,000 people during 2007–08 to 11–18 cases per 100,000 people during 2012–14. Also, the subtype of leptospirosis responsible for the most cases during 2007–08 was not present during 2012–14. The number of people developing leptospirosis was not stable, highlighting the value of measuring how commonly leptospirosis occurs over several years.
| Leptospirosis is a major cause of illness worldwide with an estimated 1.03 million cases, 59,000 deaths, and 2.90 million disability adjusted life years lost annually [1, 2]. The burden of disease is thought to be greatest in tropical countries, although reported estimates of incidence in continental Africa are scarce [3, 4]. Accurate estimates of incidence are important for estimation of disease burden and consequently, appropriate allocation of resources for diagnosis, treatment, and prevention. Challenges in estimating incidence that may account for the scarcity of reports of incidence in Africa include lack of availability of diagnostic tests [5], low clinician awareness [6], and non-specific presentation.
Although active, population-based surveillance is an ideal method for accurately determining incidence, resource and logistic challenges often preclude its use. Multiplier methods have been used successfully to estimate the incidence of acute infectious diseases in resource-limited settings by extrapolating from hospital based data [7, 8]. Specifically, multiplier methods were used to determine the incidence of acute leptospirosis in the Kilimanjaro Region during 2007–08 [9]. Using hospital based prevalence data and multipliers from a linked health-care seeking behaviour survey [10], the annual incidence of acute leptospirosis was estimated as 75–102 cases per 100,000 [9]. This estimate of incidence based on empirical data was substantially higher than an estimate (7–38 cases per 100,000 population) for Tanzania based on a modelling approach using incorporated data from a systematic review of risk factors [2]. Leptospirosis may cause endemic disease, but is also capable of causing epidemics during flooding or other extreme weather events [11]. As such, data gathered from the same location from multiple time periods can provide insights into the dynamics of disease incidence over time, distinguish periods of endemic and epidemic transmission, and help determine more representative burden of disease estimates.
We sought to estimate the incidence of acute leptospirosis in northern Tanzania from 2012 until 2014 using a similar methodology to the previous estimate in the same region in order to describe trends over multiple year periods.
We calculated the incidence of acute leptospirosis using a multiplier study design. Briefly, we combined a healthcare utilisation survey performed in two districts within the Kilimanjaro Region with a hospital-based surveillance involving systematic evaluation for leptospirosis in febrile patients at the two major referral hospitals in the Kilimanjaro Region. We multiplied the number of identified cases of acute leptospirosis by a number of factors designed to account for incomplete data using the surveillance pyramid model (Fig 1.) [9].
Study Site: We studied patients at two referral hospitals in Moshi, Tanzania. Moshi is the administrative centre for the Kilimanjaro Region that has a population of 1.6 million. Moshi is situated at an elevation of approximately 890m and has a tropical climate with rainy seasons from October through December and March through May. Aside from urban Moshi, the region is rural with inhabitants practicing cultivation and small-holder farming. Kilimanjaro Christian Medical Centre (KCMC) is a 450 bed hospital and the zonal referral centre for several regions in Northern Tanzania. Mawenzi Regional Referral Hospital (MRRH) is a 300 bed hospital and the referral centre for the Kilimanjaro region.
Enrolment procedures: From 20 February 2012 through 28 May 2014 the study team approached all adult patients who were admitted to KCMC with a febrile illness as well as all adult or paediatric patients who were admitted at MRRH. In addition we approached every second patient who presented with fever to the outpatient department at MRRH. Hospitalized participants were eligible for enrolment if they had a history of fever within the previous 72 hours or an axillary temperature of >37.5°C or a tympanic, oral or rectal temperature of ≥38.0°C at admission. Non-hospitalized patients were eligible if they had an axillary temperature of >37.5°C or a tympanic, oral or rectal temperature of ≥38.0°C. All adult study participants provided written informed consent. For those under 18 years, a parent or guardian provided written informed consent. In addition, written assent was provided for those aged 12 to 18 years. This study differed from the previous Kilimanjaro Region incidence study in its enrolment of outpatients and the enrolment of children at MRRH rather than at KCMC.
Enrolment occurred only on weekdays. Enrolled patients underwent phlebotomy, with blood allocated for acute leptospirosis serology only if there was sample available after blood parasite microscopy and blood culture. Participants were requested to return for collection of convalescent serum 4–6 weeks after enrolment. For those who did not attend the scheduled follow up, we attempted to contact them and encourage attendance. Additionally, we recorded inpatient death. Unlike the previous study estimating leptospirosis incidence in the Kilimanjaro Region, we did not record inter-hospital transfer.
Laboratory methods: Serology for leptospirosis was performed on acute and convalescent serum samples using the standard microscopic agglutination test (MAT) with a panel of 20 Leptospira serovars belonging to 17 serogroups at the United States Centers for Disease Control and Prevention. These included serogroups: Australis (represented by L. interrogans serovar Australis, L. interrogans serovar Bratislava), Autumnalis (L. interrogans serovar Autumnalis), Ballum (L. borgpetersenii serovar Ballum), Bataviae (L. interrogans serovar Bataviae), Canicola (L. interrogans serovar Canicola), Celledoni (L. weilii serovar Celledoni), Cynopteri (L. kirschneri serovar Cynopteri), Djasiman (L. interrogans serovar Djasiman), Grippotyphosa (L. interrogans serovar Grippotyphosa), Hebdomadis (L. santarosai serovar Borincana), Icterohaemorrhagiae (L. interrogans serovar Mankarso, L. interrogans Icterohaemorrhagiae), Javanica (L. borgpetersenii serovar Javanica), Mini (L. santarosai serovar Georgia), Pomona (L. interrogans serovar Pomona), Pyrogenes (L. interrogans serovar Pyrogenes, L. santarosai serovar Alexi), Sejroe (L. interrogans serovar Wolffi), and Tarassovi (L. borgpetersenii serovar Tarassovi).
Case definitions: We defined confirmed acute leptospirosis as participants who demonstrated a four-fold rise in agglutinating antibody titres between acute and convalescent serum samples. Cases were defined as probable if a participant’s serum had a single agglutinating titre of at least 1:800. These definitions were identical to those used to obtain the previous incidence estimate [9, 12]. The predominant reactive serogroup, for confirmed cases was defined as the serogroup containing the serovar with the largest rise in titres between acute and convalescent sera. For probable cases, we used the serovar with the highest titre to define the serogroup.
A time multiplier of 1.40 was used to account for enrolment occurring only on weekdays (5 of every 7 days). Additionally a study duration multiplier of 0.44 was included to calculate annual incidence from a study that enrolled for 27 months (20 February 2012 through 28 May 2014). We applied enrolment and blood draw multipliers to account for eligible patients who either did not enrol or for whom blood was not available for leptospirosis serology. Calculations of these multipliers are presented in the results. We were unable to include a transfer multiplier in the current study as details of inter-hospital transfer of participants were not recorded. For the estimation of incidence based solely on confirmed cases, a paired sera multiplier was applied to account for those patients who did not have paired sera drawn. Diagnostic test multipliers were used to account for the sensitivity and specificity of MAT serology. The sensitivity was estimated at 100% for paired sera, 48.7% for participants with solely acute sera and 93.8% for those with solely convalescent sera. The specificity was estimated at 93.8%. The estimates are based on a published evaluation of diagnostic tests [9, 13] and matched those used in the 2007–08 study.
A healthcare utilisation survey was carried out in the Moshi Urban (population 184,292) and Moshi Rural (population 466,737) Districts of Kilimanjaro Region between 13 June and 22 July 2011 as previously reported [9, 14]. Briefly, 30 of the 45 wards were selected randomly using a population-weighted approach. A study member collected data from the heads of the first 27 households encountered within the ward. A total of 810 households were sampled, comprising 3,919 household members. All households had at least one member >15 years of age, 361 had at least one member aged between 5 and 15 years of age, and 198 households had at least one member aged below 5 years. The demographic characteristics from the healthcare utilisation survey have been previously compared to the 2002 Tanzanian Census [9]. Age specific population data has not yet been released from the 2012 Census [14]. Questions relating to health-care seeking behaviour in the event of febrile illness were used to identify participants likely to present to KCMC or MRRH. These questions included, ‘what is the name of the health care facility with an inpatient ward where you/your family would go if you/your family had fever?’ and ‘what will you do if a [household member subdivided by age bracket] has a fever for ≧ 3 days?’. The hospital multipliers are presented in Table 1. Each multiplier is the reciprocal of the proportion of survey participants who responded that they would attend KCMC or MRH as their first or second choice healthcare provider.
We used population totals from the 2012 census [14]. As age specific population data were not available from the 2012 census, we multiplied age specific proportions from the 2002 census by the 2012 population total to estimate age-specific populations. The 2007–2008 Kilimanjaro Region incidence estimate used population totals from the 2002 census.
We compared incidence by using the estimate of incidence derived from confirmed and probable cases from each of the study periods and the estimated population sampled as the denominator. As shown in Table 2, the estimated population sampled was calculated by multiplying the total population by the proportion of participants in the healthcare utilization survey that identified KCMC or MRRH as hospitals they would attend in the event of febrile illness. We compared the highest estimates of incidence in each of the study periods.
We repeated all calculations using both probable and confirmed cases and then using confirmed cases only. Additionally we performed a one-way sensitivity analysis by varying hospital multipliers according to answers to alternative relevant questions in the healthcare utilisation survey that might also reflect the behaviour of participants and diagnostic test multipliers by using a range of alternative plausible sensitivity values for MAT [15–17].
Data was entered using the Cardiff Teleform system (Cardiff, Inc., Vista, CA, USA) into an Access database (Microsoft Corporation, Redmond, WA, USA). Incidence calculations were carried out using Microsoft Excel 2010 (Microsoft Corporation. Redmond, WA, USA) spreadsheets. Other analyses were performed using STATA, version 13.1 (STATA-Corp, College Station, TX, USA). We used a test of proportions to compare the incidence between 2007–08 and 2012–14. All p values are 2 sided and statistical significance was set at p<0.05.
This study was approved by the KCMC Research Ethics Committee (#295), the Tanzania National Institutes for Medical Research National Ethics Co-ordinating Committee (NIMR1HQ/R.8cNo1. 11/283), the Institutional Review Board of Duke University Medical Center (IRB#Pro00016134) and the University of Otago Human Ethics Committee (Health) (H15/055).
Of 1,115 participants enrolled from within the study districts, 1,017 (91.2%) had blood drawn for leptospirosis testing. Of the 1,115 participants, 409 (37.7%) <5 years, 111 (10.0%) 5–14 years, and 595 (53.4%) were aged ≥15 years. A total of 593 (46.9%) participants were male. A total of 758 (74.6%) participants reported having a fever for at least 3 days.
Of 1,017 participants tested for leptospirosis, 12 (1.2%) met the case definitions for confirmed leptospirosis and an additional 7 (0.7%) met the case definitions for probable acute leptospirosis. The predominant reactive serogroups among confirmed and probable cases of leptospirosis are summarised in Table 3.
Of both confirmed and probable cases, there were seven (1.7%, 95% confidence interval [CI] 0.4–2.9%) cases among 416 outpatients and 12 (1.9%, 95% CI 0.8–3.1%) cases among 601 inpatients. There was no statistically significant difference in the prevalence of leptospirosis between inpatients and outpatients (p = 0.72).
The annual incidence of acute leptospirosis in the Moshi Urban and Rural Districts (2012–2014) was 11–18 cases per 100,000 population using hospital multipliers derived from the question ‘To which facility would you go if you were unwell with a fever lasting ≧3 days?’. When using respnses to the question, ‘What is the name of the health care facility with an inpatient ward where you/your family would go if you/your family had fever?’ the incidence was 9–18 cases per 100,000 population. The annual incidence was highest in adults, ranging from 13 to 29 cases per 100,000 population. Details of the calculation and age-specific incidences are included in Table 5. The estimated incidence for confirmed and probable cases was (18 cases per 100,000 population) was statistically significantly lower than the estimate of 102 cases per 100,000 population from 2007–08 (p<0.0001).
Estimates of annual incidence in six monthly time blocks is summarised in Table 6. These data show a higher incidence during the first few months of the study.
The results of the one-way sensitivity analysis are presented in Table 7. When we derived hospital multipliers from alternative questions from the healthcare utilisation survey that might also reflect the behaviour of participants, the estimated annual incidence ranged from 8–37 cases per 100,000 population. When we varied the estimated sensitivity of MAT from the lowest to highest plausible values [5, 13, 16, 18, 19], the estimated annual incidence varied from 10–25 cases per 100,000 population.
This study highlights the dynamic nature of leptospirosis epidemiology in the Kilimanjaro Region, Tanzania. Our study shows that both incidence and serogroup dominance among human cases have changed. The overall annual estimate of 11–18 cases of acute leptospirosis per 100,000 population is substantially lower than the estimated incidence of 75–102 cases per 100,000 population per year from 2007–08 [9]. In addition the age-specific incidence appears to have changed, with children estimated to have a lower incidence than in the 2007–08 study. Despite the lower incidence during 2012–14, leptospirosis still appears to be an important cause of fever in our region.
The explanation for the wide variation in incidence is uncertain, but may reflect changes in climatic conditions, transient presence of an infected reservoir host, changes in human-animal interactions, or changes in rodent-livestock interactions. Regarding climate, leptospirosis is recognised as an important public health problem following extreme weather events in other parts of the world such as Malaysia, Philippines, and Thailand [11]. The strong El Niño conditions of 2006–07 were associated with flooding and epidemics of other diseases, such as Rift Valley Fever. El Niño conditions may have influenced leptospirosis incidence during 2007–08 [20–22]. Urbanisation or reduced livestock ownership are unlikely to explain the change in incidence, because, firstly population increased in rural areas between the 2002 and 2012 censuses and secondly livestock numbers are thought to have increased over the study period [14, 23]. Livestock vaccination coverage to any disease remains low in Tanzania, and we think that increased vaccination against leptospirosis is unlikely to account for the reduced incidence [23].
There has also been a change in the most common predominant reactive serogroup. During 2007–08, serogroup Mini was the most commonly implicated serogroup, whereas during 2012–2014 there were no cases of leptospirosis in which Mini was the dominant serogroup [24]. Our findings are consistent with an interpretation of unstable serogroup transmission dynamics, with increased transmission of a serovar from within the Mini serogroup during 2007–08. Research to understand the reservoir hosts, sources, and risk factors for serogroup Mini in northern Tanzania may help to explain the apparent unstable transmission. Aside from the fever surveillance work at KCMC and MRRH in 2007–08, there are no reports of Leptospira serovars belonging to the Mini serogroup causing human disease in East Africa, although Mini serovars have been isolated from people and small mammals in the western Indian Ocean islands of Mayotte and Madagascar [25–28]. Mini is an uncommonly tested serogroup, but serological reactivity to Mini has been shown in cattle [29–32], wild game animals [33, 34], and rodents [35]. Pilot serological study of cattle slaughtered for meat in Moshi Municipal District in 2014 found seroreactivity to L. borgpetersenii serovar Mini in 14 (24.1%) out of 58 animals tested [36]. This finding indicates that livestock may be an important source for human leptospirosis in the area. In addition, our data from both humans and livestock suggests that a representative of the Mini serogroup should be included in future MAT panels in East Africa.
In 2012–14, the most common predominant serogroup was Australis, which was the second most commonly identified predominant reactive serogroup during 2007–08. Agglutination was observed against both test serovars (L. interrogans serovar Australis and L. interrogans serovar Bratislava) in 2007–08 and 2012–14. Continued identification of this serogroup suggests that at least one serovar from the Australis serogroup is endemic in the region. Serovars from the Australis serogroup have been isolated from rodents in Tanzania [37], a human in Kenya [38], an African grass rat in Nigeria [39] and cattle in Zimbabwe [40]. Serologic studies of animals in East Africa, have reported infrequent sero-reactivity to Australis serogroup in sheep, goats and pigs [41] and cattle [42]. Seroreactivity against serogroup Australis serovars was also observed in cattle slaughtered for meat in the Moshi area in 2014 [36]. Rural residence has previously been identified as a risk factor for acute leptospirosis in northern Tanzania [24]. These findings raise the possibility of livestock as an important source of infection. However, care is needed in interpreting infecting Leptospira strains from serological data as cross-reactions between Mini, Sejroe, and Hebdomadis serogroups are common and using serological data to infer infecting serovars is unreliable [43, 44]. Therefore, Leptospira spp. isolates from humans and animals, and studies investigating risk factors are needed.
One of our study’s strengths is that the estimate of incidence used the same hospital surveillance system and the same healthcare utilisation survey within the same districts as the earlier study, allowing a direct comparison of distinct time periods. Differences in enrolment practices and multipliers between the 2007–08 incidence estimate and the current study may have influenced the difference in estimated incidence. The 2012–14 study enrolled both those hospitalised and those treated as outpatients, where as the 2007–08 study enrolled only inpatients. However, in our study, the incidence did not vary by admission status. Other minor alterations in multipliers include the addition of a blood-draw multiplier and the omission of a transfer multiplier as these data were not collected. These multipliers ranged from 0.7–1.2 and if applied, would have widened the apparent difference between incidence estimates from 2007–08 and 2012–2014. The human population of the study area has been estimated from the 2012 census in the 2012–14 study and the 2002 census for the study of Biggs et al. The total population rose from 545,168 to 651,028 between 2002 and 2012 [14]. It is likely therefore that the 2007–08 study underestimated the true population at the time of the study, and consequently will have overestimated the incidence by a small margin.
There are other limitations in our study that influence interpretation of the results. We chose to estimate incidence using multiplier methods as resource limitations precluded active surveillance in the entire population. Although it is an accepted method of incidence estimation, it means our estimate is based on a small number of cases, and some of the variation may be due to random error. In addition multiplier methods rely on many assumptions. In particular, we assumed that those who presented to the two tertiary referral hospital study sites were representative of the community sampled in the healthcare utilisation survey and that the care seeking behaviour of those surveyed is representative of the population. We are unaware of any substantial changes to the health system or health seeking behaviour from 2007 through 2014, and hence any healthcare utilisation survey error is likely to be consistent between the studies. In the 2007–08 study, Biggs et al identified differences between the age and sex distribution of the 2002 census population and those who participated in the healthcare utilisation survey [9]. We were unable to compare the demographics of the population in the 2012 census to those who took the healthcare utilisation survey but it is possible that the survey respondents are not representative. In order to understand the effect that changes in care-seeking behaviour would have on our estimate, we have performed a sensitivity analysis to provide a range of estimates. In this sensitivity analysis we varied the multipliers according to different questions from the healthcare utilization survey asking about fever shorter than 3 days. The larger multipliers derived by these questions reflect the fact that patients with shorter durations of fever are less likely to present to tertiary hospitals. We think that these estimates are less accurate than our final estimate as three quarters of our participants reported a fever of at least 3 days.
We are likely to have underestimated the incidence of disease through our use of serological diagnosis and enrolment from referral hospitals and may have biased inclusion towards those with severe disease. Firstly, we assumed that those who provided serum for testing were similar to those who did not. Secondly, we used test multipliers to attempt to account for the insensitivity of MAT in the acute phase of illness. We performed calculations using a sensitivity of 100% for MAT on paired serum samples and 48% for MAT on those providing only acute serum samples to ensure comparability with the 2007–08 incidence estimate. However, there is increasing evidence that MAT is less sensitive than the values of 100% for paired sera and 48.7% for solely acute serum that we used. It is therefore likely that the true incidence is towards the upper end of the estimate in our sensitivity analysis [16, 18, 19].
In conclusion, our study demonstrates substantial variation in leptospirosis incidence between time periods at the same site in continental Africa. Leptospirosis incidence appears to have declined from 2007–08 to 2012–14, although it remains an important cause of fever. This appears partly due to unstable transmission of a serovar from within the Mini serogroup. Our findings indicate the value of leptospirosis surveillance over multiple year time periods to understand the epidemiology of the disease. Our findings also highlight that much more work is needed to identify the animal reservoirs of Leptospira spp. in Africa in order to understand the human epidemiology.
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. Use of trade names and commercial sources is for identification only and does not imply endorsement by the US Department of Health and Human Services or the Centers for Disease Control and Prevention.
|
10.1371/journal.pcbi.1004801 | Early Transcriptome Signatures from Immunized Mouse Dendritic Cells Predict Late Vaccine-Induced T-Cell Responses | Systems biology offers promising approaches for identifying response-specific signatures to vaccination and assessing their predictive value. Here, we designed a modelling strategy aiming to predict the quality of late T-cell responses after vaccination from early transcriptome analysis of dendritic cells. Using standardized staining with tetramer, we first quantified antigen-specific T-cell expansion 5 to 10 days after vaccination with one of a set of 41 different vaccine vectors all expressing the same antigen. Hierarchical clustering of the responses defined sets of high and low T cell response inducers. We then compared these responses with the transcriptome of splenic dendritic cells obtained 6 hours after vaccination with the same vectors and produced a random forest model capable of predicting the quality of the later antigen-specific T-cell expansion. The model also successfully predicted vector classification as low or strong T-cell response inducers of a novel set of vaccine vectors, based on the early transcriptome results obtained from spleen dendritic cells, whole spleen and even peripheral blood mononuclear cells. Finally, our model developed with mouse datasets also accurately predicted vaccine efficacy from literature-mined human datasets.
| Vaccines are designed to elicit effective immune responses against antigens. The various vector platforms used in vaccine development are diverse and complex, rendering the selection of promising vaccines vector challenging. We have designed a modeling strategy that predicts the propensity of vaccine vectors to elicit strong late T-cell responses using transcriptome material obtained 6 hours after vaccination. Our model, designed with mouse datasets, also predicted vector efficacy from mined human data. Thus, molecular signatures obtained 6 hours after vaccination can predict vaccine efficacy at 2 weeks post vaccination, which should help in vaccine development.
| The development of vaccines against complex chronic diseases such as HIV or cancer has been largely unsuccessful so far. Novel vaccine technologies are rationally designed to generate appropriate protective immune responses [1], notably efficient T-cell responses. Such vaccine vectors include plasmid DNA, viral and bacterial vectors, and virus-like particles (VLPs). The intrinsic characteristics of these vectors, including their capacity to stimulate innate immunity and to activate and target the antigen to antigen-presenting cells, determine in large part their immunogenicity and thus their potency as vaccine or gene therapy vectors [2–4]. However the rational design of vectors is limited by various aspects, such as the partial understanding of the factors governing the induction of optimal immunity (i.e. the activation of the innate immune system by various vector components, the effect upon adaptive immunity…) or the possible dependence of vector efficacy on the specificity of the target diseases.
Systems biology has been introduced in vaccine development to assist in circumventing these limitations and shorten the vaccine development process. Systems biology may not only help to better understand, analyze and reconstruct the complex immune interactions between the pathogen/vaccine and host immune system, but may also improve the in silico testing models for vaccine candidates. Systems biology approaches have proven capable to predict immune responses induced after vaccination [5,6]. For example, expression patterns of genes associated with the efficient processing of peptides for major histocompatibility complex presentation have been identified as useful surrogate markers of vaccine efficacy, obviating the need to perform challenge studies [7]. Signatures derived from antibody repertoire profiling on peptide microarrays during the natural course of influenza infection were shown to be predictive of the efficacy of influenza vaccines [8]. Multivariate analysis performed on human peripheral blood mononuclear cell (PBMC) microarray data, obtained 3 days after vaccination, identified innate immune response–related signatures that predicted the late adaptive immune response to the YF-17D yellow fever vaccine [9].
In this manuscript, we describe a methodology that enabled us to successfully predict the adaptive immune responses induced by large sets of vaccine vectors of different classes, ranging from infectious particles to VLPs and DNA. All these vectors expressed the same antigen, the immune response to which was measured using a validated standardized method. We developed our model based on the analysis of transcriptomic data, obtained 6 hours after vaccination, that could predict the antigen-specific immune responses induced at the peak of the response, 5–10 days later. It is noteworthy that this model, developed in mice, successfully predicted vaccine-induced responses from literature-mined human datasets.
Forty-one vectors classified in 13 categories of vaccines and all expressing the same antigen were evaluated and compared for their ability to induce an adaptive T-cell immune response after vaccination (S1 Table). The forty-one vectors included (i) recombinant viral vectors derived from adenovirus (rAd), vaccinia (VACC), modified vaccinia Ankara (MVA) and lentivirus (LV), (ii) recombinant bacteria vectors derived from Bacille de Calmette et Guérin (BCG), (iii) recombinant VLPs made of the AP205 [10] or Qbeta (Qb) [11] proteins from bacteriophage, the VP2 proteins from murine polyoma virus (MPY) [12] or murine pneumotropic virus (MPT), the Gag capsid proteins from murine leukaemia virus (MLV) [13], the core from hepatitis B virus (HBc), and (iv) plasmid encoding a recombinant protein (DNA) or recombinant MLV-VLPs (plasmoVLPs) [13,14]. Each vaccine platform was engineered to display or express the immunodominant LCMV gp33-41 epitope model antigen [15] in order to compare the different vaccine-induced CD8+ T-cell specific responses. In the framework of CompuVac (www.compuvac.eu), we standardized the method for measuring the gp33-41-specific T-cell response using tetramer staining (Fig 1A). Mice were immunized with each vector and we evaluated the gp33-41-specific T-cell response in PBMCs at days 5, 7 and 10, following the frequency of circulating gp33-41/H-2Db tetramer+ CD8+ T cells. In each experiment we included control mice that were injected with PBS or rAd (rAd_1 batch) to provide negative and positive controls. Data for each experimental group were normalized as the experimental to rAd vector response ratio allowing cross-laboratory data comparisons.
We observed a wide range of immune responses that were triggered by the different vectors. The maximal CD8+ T-cell expansion was induced with bacteriophage-derived VLPs, while very low but significant responses were observed with MPT and HBc VLPs (Fig 1B). Interestingly, different vector designs within the same vector platform led to different responses. As an example, Qb-derived VLPs induced variable CD8+ T-cell expansion depending on their production processes that were designed to modify their TLR-ligand composition (i.e. Qb_5 devoid of viral RNA and CpG in contrast to Qb_1; Fig 1A). We took into consideration all the vectors and performed hierarchical clustering on normalized values that defined 3 clusters (C). The first cluster comprised vectors with low ratio values, characterizing weak inducers of antigen-specific T cells, hereafter referred as “Weak” vectors. The other 2 clusters included vectors inducing high or intermediate responses, defining the “Strong” vector class. This class comprised the different recombinant viral vectors (rAd, MVA, VACC, LV) expressing rather than displaying the antigen, and which have been extensively developed as CD8+ T-cell vaccines [16–18]. It also contained bacteriophage-adjuvanted VLPs, in agreement with previous reports [10,19].
As dendritic cell activation is key to the initiation of immune responses, we investigated whether transcriptome data from sorted spleen dendritic cells (DCs) sampled 6 hours after immunization could be predictive of the antigen-specific T-cell response measured several days later, at the peak of the response. To address this question, we devised a stepwise modelling scheme. DC-sorted transcriptome datasets were initially produced for 19 vectors on the Codelink platform, corresponding to 7 different vaccine platforms, for which the antigen-specific T-cell response was also measured (S1 Table).
The rationale for looking at signatures instead of individual genes was motivated by (i) the need to detect slight gene expression modifications (captured as the overall expression changes of correlated genes), (ii) the technical constraints of working on different microarray platforms (CodeLink, Illumina and Affymetrix), and (iii) the objective of producing a predictive model working across microarray platforms. Thus, our modelling scheme was based on our recently described strategy for signature discovery, using independent component analysis (ICA) followed by gene set enrichment analysis (GSEA) [20]. This allows circumventing the limitations due to the use of different platforms when analyzing individual gene expressions, by comparing statistical signature’s enrichment across datasets.
ICA is an unsupervised algorithm extracting independent components Y from original datasets X by searching for the demixing matrix W:
Y=X×W
W matrix is calculated by maximizing the non gaussianity of the components measured as the negentropy J:
J(y)=H(yGauss)−H(y),
where H(y) and H(yGauss) are the Shannon entropy for a vector y and a random Gaussian vector with same variance as y [21].
The use of ICA to analyze microarray data is justified by the hypothesis that X is a mix of signals from underlying cellular pathways. Therefore, columns of Y contain a summary of gene contributions in the extracted components. The RNA expression value of a gene is thus the superposition of several signals of this gene in each component which add up. From each component y, two reduced gene sets can be extracted by selecting genes with critical contribution on both sides of the distribution [22].
We first performed ICA on the 19 available datasets, yielding 210 molecular signatures characterizing the variability within each dataset, and likely linked to vector properties. We then analyzed the differential gene expression between the controls and the tested vectors using bootstrapping [23,24], in order to increase the model’s sensitivity. Bootstrapping consists in sampling series of additional datasets by randomly drawing samples with replacement of equal size from an original dataset, as described in Fig 2. We sampled 100 consecutive bootstrapped datasets from each of the 19 original datasets and generated 100 corresponding ranking lists of genes based on modified t-test statistics. The previously identified signatures were then tested for their behavior vis-à-vis the gene lists using GSEA, generating normalized enrichment scores (NES). Molecular signatures from GSEA software (>5000) were added at this step in order to increase the efficiency of the normalization procedure. NES of molecular signatures from ICA were then extracted for the next steps. This yielded a matrix, containing 1900 columns (100 bootstrapped datasets for each of the 19 original datasets) and 210 lines (the number of calculated NES). This matrix was then used to create random forest (RF) classification models (Fig 2). NES values and T-cell response classification were used as predictors and dependent variables, respectively, in the randomForest package, which as output provides classification results and associated probabilities for each T-cell response class.
An initial predictive model was built with 9 vector datasets (in red in Fig 1B) for which the antigen-specific T-cell responses were available (900 bootstrapped datasets and 100 signatures). Predictions of 10 additional datasets, including independent experiments done with the same or different batches of these vectors, were very consistent (see Tables 1 and S2). The model sensitivity for the “Weak” and “Strong” vector classes (respectively equal to the specificity for the “Strong” and “Weak” classes) are 0.89 and 0.98, respectively. The positive predictive value (PPV) is stable for the two classes (“Weak”: 0.96, “Strong”: 0.93). This 9-vector model is already efficient to classify the vector platform with 0.94 accuracy. These results led us to construct the final predictive model (called RFM model) including all the 19 datasets, based on the analysis of the 210 signatures across the 1900 bootstrapped datasets. This complete training set contained enough information to discriminate clearly between the 2 vector classes, as demonstrated by the misclassification rate parameter reaching zero after 100 simulated trees.
The RandomForest algorithm provides a ranked list of the signatures based on their importance to the efficacy of the classification in the model. This score is based on the decrease of the Gini impurity criterion for each child node of a split. The result of this calculus is the mean of this decrease for each signature present in the trees of the forest. 27 most important signatures, having a mean decrease score higher than ten, were selected. Clustering methods were then applied (i) on NES values of these 27 signatures calculated on original datasets (Fig 3A) and (ii) on the mean NES values calculated on the bootstrapped datasets (Fig 3B). The interest of bootstrap is clearly revealed with clusters more explicitly defined after bootstrap.
We then asked whether RFM was biased toward particular vector datasets. We first used the leave-one-out methodology, where 19 models were iteratively built using only 18 out of 19 datasets, and then assessing how accurately such models predict the 100 bootstraps from the left-out dataset. All vectors were classified as expected for at least 96 of the 100 bootstrapped datasets, except MPY_3 for which 16 bootstrapped datasets were misclassified (S3 Table). This result shows overall very high prediction stability and no significant bias of the RFM model.
We verified that RFM was not biased for a given vector platform. One hundred new models were constructed, each based on one randomly selected representative of the 7 vector platforms (rAd, AP205, MVA, MPY, MPT, MLV and BCG). For each vector, the probabilities to be classified as expected were calculated and the prediction distribution across the 100 models is shown in Fig 4. Vaccines from the “Strong” vector class (in red) showed good consistency in their prediction distribution, with no value under 0.6 (100% confidence). Vaccines from the “Weak” vector class showed more variability: in particular, 2 MPY vaccines (MPY_3 & MPY_3bis; same vector batch (#3) used in 2 independent experiments) were not classified as expected in 16 models out of 100 (84% confidence); these 16 misclassifying models all used MPY_2 as the MPY representative. Note that this specific preparation (#2) of MPY vaccine was produced using baculovirus machinery in insect-derived cells, while the other MPYs were produced in yeast.
RFM was then used to predict the vector class of 4 new vectors belonging to 3 vector platforms: 2 batches of lentivirus (LV) vectors -a category of vaccine not represented during the model establishment, one new batch of AP205 (AP205_3) and one of MLV (MLV_2). We had independently determined that LV vectors induced strong antigen-specific T-cell responses after immunization and were classified in the “Strong” vector class (Fig 1). As shown in Tables 1 and 2A, these 4 bootstrapped datasets were classified as expected with high precision (>95%) while sensitivity and PPV of the model increased compared to the 9-vector model, especially the sensitivity for the “Weak” vector class now reaching 0.97 (from 0.89) with RFM. These results highlight that RFM (i) is not vaccine platform-dependent, (ii) correctly predicts a vector platform unknown to the model, and (iii) efficiently predicts both “Weak” and “Strong” vectors.
RFM was built on transcriptome data obtained from sorted spleen DCs. In our next experiment, we assessed whether RFM would be sensitive enough to classify transcriptome datasets derived from whole spleen samples obtained 6 hours after immunization, where DCs represent 1–2% of total splenocytes. As summarized in Table 2B, all bootstrapped datasets from whole spleens were well classified, with at least 91% of the expected classification, thus demonstrating our model’s sensitivity in classifying vectors in whole spleen transcriptome datasets.
We then tested microarray datasets for whole spleen samples obtained 6, 48 and 72 hours after vaccination with one vector, the rAd vector that we used as a standard. Strikingly, only datasets sampled 6 hours after injection were classified as expected (as “Strong”) (Table 2D).
Similarly, we tested the performance of our model in classifying vectors using PBMC-derived microarray datasets. The rationale for this experiment is that PBMCs, less than 1% of which are DCs, offer a more accessible sample source than spleen, especially in humans. As shown in Table 2C, all but one vectors were classified as expected with high precision (≥ 90%). AP205_1 was classified as expected, though with less confidence (73%).
Finally, we tested whether our model could classify datasets obtained from the literature. We found datasets from the Merck Ad5/HIV trial reported by Zak et al. [25] PBMC transcriptome data were generated from samples obtained at 6, 24 and 72 hours after vaccination. We bootstrapped the samples of Zak et al., taking patient-paired samples before and after vaccination. 100% and 91% of the bootstrapped paired samples were predicted as “Strong” at 24 and 72 hours, respectively (Table 3), in line with the authors’ original observations. The same analysis performed with the 6-hour time point gave a “Strong” prediction for 31% of the bootstrapped paired samples. The latter finding is consistent with the conclusion of Zak et al. that transcriptomic modifications at 6 hours were not significant. These results demonstrate the capacity of RFM generated from mouse DC transcriptome datasets to classify human PBMC datasets.
Biological annotation of the 27 most important signatures of RFM reveals one signature (Sig1) with statistical functional enrichments related to immune processes (FDR p-values 10−4–10−8). This signature is highly focused on STAT-1 with 51 genes having strong biological connections (Fig 5A). Interestingly, Sig1 is upregulated in all the vectors, but with higher intensity in the “Strong” as compared to the “Weak” vectors.
No specific molecular pathway was clearly identified by QIAGEN’s Ingenuity Pathway Analysis (IPA) functional analysis for the other 26 important signatures in our model. However, visual inspection of these signatures identified the CH25H gene as highly modulated by strong vectors. Since this gene has been recently described as playing a role in DC maturation [26], we analyzed its network of connected genes with IPA (Fig 5B). This network was also globally more modulated by “Strong” rather than “Weak” vectors, and comprised genes implicated in DC function such as MYD88, DUSP5 and ABCG1.
Understanding and predicting innate immune response to vector platforms is primordial for fast and effective production of new vaccination or gene therapy protocols. Systems biology tools efficiently extract information from large datasets in computing predictive models and have already played a major role in recent discoveries in this field [5,27]. In this paper, we initially focused on early transcriptomic changes of DCs since these are first-line players in the innate immune response and directly contribute to the triggering of the adaptive response. Our aim was to identify transcriptomic signatures predictive of the late CD8+ CTL responses to the LCMV gp33-41 model antigen conveyed by a variety of vaccine vectors.
Based on molecular signatures extracted using the non-supervised ICA method [20,22], we produced and validated a prediction model taking into account 19 available datasets generated with different vector platforms. We chose the random forest learning algorithm for its reported efficiency among classification methodologies [28–30]. The originality of our strategy was the use of signatures rather than genes to classify samples. Our results showed that this model consistently predicts both “Weak” and “Strong” vectors, with greater confidence for the latter. This suggests that there are shared gene expression modifications induced by “Strong” vectors, while changes induced by “Weak” vectors are more diverse. Consistent with this, Li et al. recently reported that different types of vaccine lead to different transcriptomic modifications in humans 3 days after vaccination [31], with vaccines inducing high transcriptomic modifications being those that induce robust antibody responses.
Among the 27 signatures selected for their importance in the RFM model, one (Sig1; see S4 Table) is related to immune components, including “viral infection”, “role of RIG1-like receptors in antiviral innate immunity” and “interferon signalling” pathways. Previous studies have characterized gene expression modifications in the early stages of vaccination consistent with Sig1 annotation. Querec et al. investigated the transcriptome of patient PBMCs at days 0, 1, 3, 7 and 10 after vaccination with yellow fever vaccine [9]. Of 65 regulated genes, 26 were related in part to interferon and the antiviral response, including MX1, IFIT1, IFIT2, IFIT3, OAS1, OAS2, OAS3 and OASL, and 7 were related to signal transduction, including STAT1 and IRF7. Similarly, Zak et al. [25] applied the modular transcriptome analysis framework described in Chaussabel et al. [32] to study the innate immune response to MRKAd5/HIV in PBMCs 6, 24, 72 and 168 hours after patient vaccination. They identified genes highly regulated at 6 and 24 hours, including STAT1, STAT2, IFITs, MXs and OASs (also identified in Querec et al.). Strikingly, all these genes are also part of Sig1, emphasizing further their key role in the early response to the vaccine. Furthermore, DDX60, a newly described antiviral factor that induces Rig-1-like receptor-mediated signaling [33], present in Sig1, was reported by Querec et al. as well [9]. Interestingly, Sig1 is upregulated in vaccinated samples compared to control group, but to a lesser extent in “Weak” vs. “Strong” vectors (see Figs 3 and 5).
Our cross-analysis of Zak et al.’s microarray data on Merck Ad5/HIV-vaccinated human PBMC samples, which yield good predictions for the 24- and 72-hour time points, demonstrates that our prediction model, solely based on mouse DC-sorted transcriptome data, efficiently predicts human transcriptome data. This can be explained by the high similarity of gene expression in immunological cell lineages between mice and humans [34], although the kinetics of the immune response to vaccine is different.
No specific molecular pathway was clearly identified by IPA annotations for the other 26 important signatures in our model. This is somewhat surprising since these signatures have been selected by the model to best distinguish “Strong” and “Weak” vectors and are therefore expected to represent differentially regulated biological pathways. In this line, none of the 27 signatures corresponds to a peculiar behavior of a vector but they rather reveal similar behavior within “Strong” or “Weak” groups (Fig 3). Moreover, the identified signatures were extracted from 13 out of 19 different vector datasets (9 “Strong” and 4 “Weak” vectors). We believe that these signatures are unlikely artifactual but related to yet undefined biological processes. Indeed, the constant improvement of annotation databases can reveal secondary or additional functions of genes. For example, CH25H, a gene found in one of the 26 signatures and clearly upregulated in “Strong” vectors, is primarily involved in cholesterol metabolism, but has recently been shown to play a role in the early stage of DC maturation [26]. Fig 5 shows how the expression of this gene is related to dendritic cell through direct or indirect interactions with STAT-1 or IFNγ, both members of Sig1, and with several genes known to be important in early dendritic cell activation: for example, MYD88 is a gene involved in toll-like receptor signaling [35], DUSP5 is known to be upregulated during dendritic cell maturation [36], and ABCG1 is a gene playing a role in adaptive immune responses [37]. The comparative analysis of gene expression modulation of this interaction network shown in Fig 5 reveals a similar pattern of differential expression for “Strong” vectors (rAd_1, AP205_1) different than that observed for “Weak” vector datasets (BCG_2, MPY_3bis). This again points at a significant difference in early dendritic cell activation-related gene behavior in “Strong” vs. “Weak” vectors.
Altogether, our results underline the relevance of the CompuVac initiative that consisted in producing, in a standardized manner, immunological and transcriptome data related to vaccine candidates in order to predict their capacity to elicit strong antigen-specific responses. Our model was based on transcriptome data from sorted spleen DCs of mice vaccinated with various “Strong” and “Weak” T-cell inducer vectors. This prediction model accurately predicted the behavior of these and other candidate vaccines only 6 hours after injection. The model was powerful enough to produce a relevant vector classification even when using whole mouse spleen and PBMCs, or even human PBMCs (Fig 6), and across 3 microarray platforms (CodeLink, Illumina and Affymetrix). The accuracy and sensitivity of the model are likely high because it is built with very different vaccine platforms therefore representative of possible vector behaviors in triggering the early immune response. This study further supports the potential of systems immunology approaches in facilitating the development and characterization of vaccines, offering robust in silico solutions to study the early events of the immune response to vaccines.
Experimental protocols complied with French law (Décret: 2001–464 29/05/01) and EEC regulations (86/609/CEE) for the care and use of laboratory animals and were carried out under Authorization for Experimentation on Laboratory Animals Number 75-673-R. Our animal protocol (Ce5/2009/042) was approved by the “Charles Darwin” Ethics Committee for Animal Experimentation (CNREEA 05) and performed in the licensed animal facility A75-13-08.
Recombinant adenovirus- and MVA-derived viral vectors, BCG-derived bacterial vector, AP205 [10] or Qb [11] bacteriophage-, MPT- and MPY- [12] or MLV-derived [13] VLPs used as an antigenic platform and DNA vaccines were included in this study. According to the CompuVac evaluation scheme, each vaccine platform was engineered to display / express the LCMV gp33-41 model antigen [15] in order to measure the vaccine-induced T-cell specific responses and dendritic cell transcriptome changes (see following sections). The sequence IITSIKAVYNFATCGILAL corresponding to the GP33-41 epitope flanked upstream and downstream by 5 of its natively neighboring amino acids was used. The 53 vectors considered in this paper (S1 Table) are displayed in 13 vector platforms 7 of which were used for a training set (rAd, MVA, AP205, MPT, MPY, MLV and BCG) and 2 for prediction of new platforms (LV and Qb).
Groups of three to five 7-week-old female C57BL/6 mice (Charles River, France and Germany) were immunized with a controlled quantity of vector particles as defined in CompuVac assay protocols (www.compuvac.eu). For monitoring T-cell responses, each vector was injected with its “best” route of administration: subcutaneously for VLP vectors; intramuscularly for recombinant antigen-expressing vectors and by intra-dermally by gene gun for DNA vaccines. Control mice were injected with 100 μL of phosphate buffered saline solution (PBS). For each vector (n = 41), the T-cell immune response measurement was performed independently one to three times. T-cell immune responses induced against the LCMV gp33-41 model antigen were measured by MHC-I gp33-41/H-2Db tetramer (ProImmune, UK) staining of PBMCs at 5, 7 and 10 days after injection. The highest measure was kept for each mouse and the mean value was then calculated for the group. Values were normalized against measures monitored in parallel in mice immunized with the rAd_1 control vector.
Experimental groups comprised of 3 to 6 mice immunized with vaccine candidates by the intravenous route. Mice were sacrificed 6 hours after immunization. Spleen DCs were purified with CD11c+-conjugated MACS magnetic beads (Miltenyi Biotec) according to the manufacturer's instructions. After incubation for 20 minutes at 4°C, cells were washed and passed over a MACS column. Purity was checked routinely by FACS and found to be greater than 96±2%. 2x106 CD11c+ cells were used for total RNA extraction using Nucleospin RNAII (Macherey Nagel). For test dataset generation, whole PBMCs and/or whole splenocytes and/or sorted spleen DCs were collected at 6 hours, and at 48- and 72-hour time points for the kinetic follow-up. RNA was checked for quality using gel electrophoresis and for quantity using a Nanodrop spectrophotometer (Thermo Scientific). Microarrays were performed using either Applied Microarrays (CodeLink Mouse Whole Genome Bioarray) or Illumina (WG6 Mouse BeadArray) technologies (S1 Table). The MessageAmp II aRNA Amplification Kit (Ambion) was used for cDNA and cRNA production from 1 μg of total RNA. 10 μg of amplified cRNA was subsequently fragmented and hybridized for 20 hours using the Applied Microarrays hybridization and washing buffer kit. Slides were scanned using the GenePix Personal 4100A scanner for CodeLink array or the Illumina BeadArray 500GX Reader for Illumina array. Hybridization and raw data extraction were performed using either GenePix Pro 6.0 (for CodeLink array) or BeadStudio (for Illumina array) software, respectively (GEO accession GSE66991).
Each tested vector dataset comprised “vector-immunized” and corresponding PBS control samples. Quantile normalization was performed with the limma package [38] on R software [39], and then a log2 transformation was applied. Probes with a detection p-value above 0.05 in all samples in a dataset were discarded.
Following our two-step ICA→GSEA signature discovery strategy [20], signatures were extracted using the fastICA algorithm R package [40] following modifications in [22]. Parameters were set as default, except for the unmixing matrix A-1 convergence threshold set to 10-6. Ranked gene lists were calculated using the limma modified t-test. ES were calculated using GSEA [41] with the pre-ranked gene list protocol. Normalized ES are then calculated based on the permutation performed on gene sets collection, allowing comparison between experiments. The ICA-extracted signature database was complemented with the MsigDB C2 (curated gene sets of biological pathways) and C5 (Gene Ontology gene sets) databases (www.broad.mit.edu/gsea) in order to increase universe of genes available for permutation of gene sets. Signatures with fewer than 7 detected genes were ignored.
For each model produced in the Results section, classification was performed on a matrix of fastICA extracted signature NES values (see above section) calculated for bootstrapped vector datasets (100 bootstraps per vector dataset), using the random forest algorithm implemented in the randomForest R package to produce a forest of 2000 trees [42]. The number of randomly selected signatures used at each of the 2000 runs was set according to the mtry function implemented in the randomForest package. The class prediction of the new dataset was deduced by the probability to be “Weak” or “Strong” > 0.5. The overall vector class was then obtained as the majority of “Weak” or “Strong” class assignments over the 100 bootstraps.
For classification model validation, we implemented the leave-one-out methodology consisting in creating models with n-1 datasets, where n is the total number of datasets, and classifying the dataset left out. In addition, we implemented a “multi-model” methodology based on the classification of bootstrapped datasets over 100 models created as above. Each model was computed on an NES matrix of a random selection of one representative vector dataset of each of the 7 represented vector platforms (see Vector platforms section and S1 Table). Vector mean probabilities were calculated as the average probability of being “Weak” or “Strong” over the 100 bootstrapped vector datasets, and their distribution over the 100 models was analyzed.
For biological insight evaluation of the signatures, microarray data were analyzed through the use of QIAGEN’s Ingenuity Pathway Analysis (IPA, QIAGEN Redwood City, www.qiagen.com/ingenuity).
|
10.1371/journal.ppat.1000977 | Oxidation of Helix-3 Methionines Precedes the Formation of PK Resistant PrPSc | While elucidating the peculiar epitope of the α-PrP mAb IPC2, we found that PrPSc exhibits the sulfoxidation of residue M213 as a covalent signature. Subsequent computational analysis predicted that the presence of sulfoxide groups at both Met residues 206 and 213 destabilize the α-fold, suggesting oxidation may facilitate the conversion of PrPC into PrPSc. To further study the effect of oxidation on prion formation, we generated pAbs to linear PrP peptides encompassing the Helix-3 region, as opposed to the non-linear complexed epitope of IPC2. We now show that pAbs, whose epitopes comprise Met residues, readily detected PrPC, but could not recognize most PrPSc bands unless they were vigorously reduced. Next, we showed that the α-Met pAbs did not recognize newly formed PrPSc, as is the case for the PK resistant PrP present in lines of prion infected cells. In addition, these reagents did not detect intermediate forms such as PK sensitive and partially aggregated PrPs present in infected brains. Finally, we show that PrP molecules harboring the pathogenic mutation E200K, which is linked to the most common form of familial CJD, may be spontaneously oxidized. We conclude that the oxidation of methionine residues in Helix-3 represents an early and important event in the conversion of PrPC to PrPSc. We believe that further investigation into the mechanism and role of PrP oxidation will be central in finally elucidating the mechanism by which a normal cell protein converts into a pathogenic entity that causes fatal brain degeneration.
| The protein only theory, a widely accepted model describing the prion agent, assumes that the mechanism underlying prion disease pathogenesis includes a conformational change of the α-helix rich, soluble and protease sensitive PrPC into an aggregated and protease resistant β-sheet rich PrPSc form. Until recently, no covalent modification was known to be associated with such a conversion, making it difficult to follow the individual fate of each PrP form or to associate cellular events as stress-response or inflammation with the formation of prions. We now show that before PrPC initiates its conversion from proteinase K sensitive to resistant and from soluble to aggregated in the pathway to becoming PrPSc, it first undergoes oxidation of the most hidden Met residues located in a protein region exhibiting sequence identity for all species. While the cellular events promoting such oxidation in this transmissible disease remain unclear, we present evidence that PrP molecules carrying a mutation ascribed to the most common familial prion disease spontaneously oxidizes at these same Met residues. Our data provide new insights into the mechanism underlying familial Creutzfeld Jacob disease (CJD) and contribute to our general understanding of the fundamental processes related to prion pathogenesis.
| Prions are infectious agents that cause neurodegenerative diseases, such as scrapie, bovine spongiform encephalopathy (BSE) and CJD. They are believed to be composed mainly of PrPSc, a misfolded form of the GPI-anchored glycoprotein termed PrPC[1]. While the function of PrPC has not been fully elucidated, it has been suggested that this protein plays a role in the protection of cells from copper-induced oxidative stress [2]–[5]. Until recently, and mainly in the absence of convincing data to the contrary, the two PrP isoforms were believed to differ from each other only by their high-order structures; mostly an α-helical fold for PrPC, and largely a β-sheet assembly for PrPSc[6]. Nevertheless, while investigating the epitope of an α-PrP monoclonal antibody (mAb) with an uncommon recognition pattern (IPC2), we came to the conclusion that at least one of the Helix-3 methionine residues of PrPSc, M213, is differentially oxidized [7]. The oxidation of PrPSc was also confirmed by chemical reduction experiments, state of the art mass spectrometry and detection by an antibody generated against a MetO rich maize protein [8]. The finding that M213 as well as the other conserved Helix-3 Met residue, M206, were oxidized in PrPSc was first reported in the seminal work of Stahl et al. following sequencing of the PrP27-30 endoLysC peptides [9]. The fact that these specific Met residues are oxidized in PrPSc is particularly intriguing since they are the most buried residues among methionines in the 3D PrP α-fold and thus are less accessible to reactive oxygen species (ROS) [10]. So is the case for Met 205, present in PrP proteins from some species, which when mutated to both Ser or Arg destabilizes the protein structure [11]. However, if and when they are oxidized, Helix-3 Met residues may not be targeted by the methionine reductase (Msr) system, which reverses oxidation of accessible Met residues [12], [13]. Indeed, it was shown that while mice overexpressing superoxide dismutase (SOD), which inhibits oxidation, presented prolonged incubation periods upon RML infection, ablation of the MsrA system did not reduce the time from infection to disease outbreak [14].
The time course of Helix-3 Met oxidation as related to PrP conformational conversion is of great mechanistic importance. If this specific oxidation takes place after PrPSc is formed and accumulated in brain cells, then Met oxidation, while being an interesting covalent marker of PrPSc, may not participate in the sequence of events leading to prion formation and disease manifestation. Conversely, if Met oxidation occurs on the PrPC form and mediates the subsequent conformational change, then methionine oxidation may constitute an early and important step in prion formation. Along these lines, theoretical investigations have predicted that the polarity increase of Met 206 and 213 residues upon sulfoxidation may induce destabilization of the PrP helical conformation [15]. This prediction agrees with the destabilization of the native α-fold and the appearance of proaggregating properties observed in PrP chains with either methoxinine or serine substitutions of Helix-3 Met residues [16], [17]. To further establish the role sulfoxidation in PrPSc formation, we aimed to generate pAb antibodies against linear PrP human/mouse sequences, which include reduced and oxidized Helix-3 Met residues. As opposed to the complex IPC2 epitope, which precludes simple recognition of most PrP forms upon disulfide bond reduction [7], an antibody raised against a linear sequence should detect all denatured PrP forms unless a covalent modification in the amino acid chain interferes with such recognition. Furthermore, the use of such antibodies would allow quantitative investigation of the different PrP forms under similar conditions, avoiding the need for distinct purification protocols, which alone may change the properties of the tested proteins.
Consistent with this prediction, we now show that all human PrPSc and most mouse PrPSc chains were not detected by antibodies generated against reduced PrP Helix-3 Met residues unless these brain proteins were previously reduced by strong chemical reagents. In addition, our antibodies did not detect PrPSc expressed in prion-infected cells or partially aggregated PrPs present in gradient fractions of prion-infected brains, indicating that both newly formed PrPSc as well as intermediate PrP forms could be oxidized. Intriguingly, this was also the case for a PK-sensitive mutant PrP form linked to the most prevalent familial prion disease rHuPrP(23–231) E200K[18]. Our results establish the presence of sulfoxides in the Helix3 methionines in all pathogenic forms of the prion protein and indicate that such oxidation most probably precedes the conversion of PrPC into proteinase K (PK) resistant PrPSc.
Figure 1A shows the sequences of the PrP Helix-3 region for various species. These sequences are very similar for all species listed and even identical for the 206–214 regions, which includes both Met206 and Met213. Some species (such as human, mouse and cow) also present a Met residue at position 205. To generate specific Ab to reduced and sulfoxidized Helix-3 PrP forms, we immunized rabbits with several KLH-coupled peptides (Figure 1B). These peptides include KLH coupled to the Hu/Mo 203–214 sequence, which covers the three Met residues in these species. As oxidized antigens, we inoculated rabbits with two peptides prepared by different methods, including Hu/MoPrP 201–214 coupled to KLH, synthesized with MetO residues, and KLH-C-204-213, which was oxidized with H2O2 after synthesis and coupling. Following several rounds of immunization (see Methods), the rabbits (two for each peptide) were bled, and isolated serum was tested against normal brain homogenates from different species. Next, positive homogenates were immunoblotted with the designated antiserum preincubated with an array of small PrP peptides (Figure 1C) to determine the recognition site of each antibody on the protein sequence by competition. Finally, the characterized serum samples were tested against prion-infected samples.
No reactivity against any form of PrP was detected using the serum from the rabbits immunized with the oxidized KLH-KM peptide (not shown). Properties of the other serum samples are described in Figure 2. As shown in panel A, both the antiserum raised against the KLH-conjugated and oxidized TC peptide (pAb RTC) and the antiserum raised against the reduced VC peptide (pAb RVC) clearly recognized Mo and Hu PrPC, although they did not recognize Ha and Bo PrPs. Indeed, while the Mo and Hu PrP sequences are identical in this region, other species present slight individual differences in the 203–205 sequence (Figure 1a).
As opposed to the similarity in species reactivity of the pAbs RTC and RVC, the RTC and RVC epitopes on the PrP sequence were found to be different, as determined by the inhibition of the PrP immunoblotting signal with an array of peptides. While the activity of the pAb RVC was inhibited by the KM peptide, which includes the three Met residues, this same peptide did not affect recognition of PrP by the pAb RTC in either the reduced or H2O2-oxidized form. In contrast, the activity of RTC was inhibited by the TM peptide (201–205), suggesting that this pAb recognizes the TVDK or TVDKM sequence (present in Hu and Mo PrP), N-terminally to the relevant Met residues. This finding implies that both oxidized PrP peptides failed to generate an immune response to the oxidized Met rich region, consistent with investigations in other fields indicating that charged and oxidized epitopes are mostly unrecognized by T-cell receptors [19], [20].
Next, we examined the capacity of the RTC and RVC pAbs to recognize PrPSc forms. To this effect, we immunoblotted brain samples from normal and prion infected brains (digested in the presence or absence of proteinase K at 37°C) with a panel of antibodies generated against PrP epitopes upstream and downstream of the helix 3 Met area to ensure that PrPSc was present in its full length under these conditions. Figure 2C shows that the α-PrP mAb 6H4 (Prionics), which recognizes PrP in all species (at residues 145–152), the pAb RTC and also the recombinant R1 antibody (at residues 225–231 in rodent PrP [21]) detected PrP isoforms in normal and prion-infected brain samples. However, the PrP recognition pattern of the pAb RVC was significantly different. While this reagent recognized PrP from normal brain samples as well as from undigested brain samples of RML-infected mice and genetic and sporadic CJD patients (believed to comprise both PrPSc and PrPC) the pAb RVC could not detect all of HuPrPSc and detected only low levels of MoPrPSc after PK digestion.
To confirm that the recognition pattern of the pAb RVC vis a vis proteinase K resistant PrPSc forms indeed resulted from Met oxidation, we subjected PK digested extracts from prion-infected human and mouse brains to N-methylmercaptoacetamide (MMA), a specific MetO reducing agent [22]. Figure 2D shows that following MMA treatment, the pAb RVC easily recognized both human and mouse PrPSc at detection levels similar to those of the α-PrP mAb 6H4 both before and after MMA treatment. Similar but less striking results were obtained for detection of the reduced samples by IPC2 because, as described in our previous publication [7], full detection of reduced PrP forms by this mAb requires deglycosylation of PrP forms by PNGase.
To establish whether oxidation of Met residues is essential for the conversion of PrPC to PrPSc, we asked whether Met oxidation occurs first on PrPC or whether oxidation is a delayed effect related to the long-term accumulation and reduced clearance of proteinase K resistant and misfolded prion protein in the brains of the affected subjects. To separate these mechanistic possibilities, we studied by pAb RCV the oxidation status of newly formed PK resistant PrPSc generated in cells permanently infected with prions, such as ScN2a [23] and ScGT1 cells [24]. Since these cells constantly proliferate, PrPSc produced by them can be considered relatively new, as opposed to the PrPSc molecules that accumulate in infected brains.
For this experiment, extracts from ScN2a cells (infected with the RML mouse prion strain) and from the ScGT1 cell line (infected with both the RML and the 22L prion strains) [25] were treated in the presence or absence of proteinase K and immunoblotted with the anti-PrP IPC1mAb, which recognizes all forms of Mo and Ha PrP ([7], Sigma), and the pAb RVC, which properties were described above. Extracts from the uninfected cell lines (N2a and GT1) and brain samples from normal and scrapie-infected mice were included in this study. Figure 3a shows that while the IPC1 mAb recognized all forms of PrP in cells and brains, the pAb RVC failed to detect proteinase K resistant PrP forms in any of the infected cell systems and barely detected bands characteristic for PrPSc before proteinase K digestion. These results indicate that newly made PrPSc may be quantitatively oxidized, as was shown here for PrP from two different cell lines and for two prion stains (ScN2a-RML, ScGT1-RML and ScGT1- 22L). Similar results were obtained for an RML infected GT1 line expressing chimeric Mo-Ha PrP [26], [27] (not shown here for the pAb RVC, see bellow for a similar antibody). Contrary to PK resistant PrPSc in the cells, and as depicted also in Figure 2, pAb RVC could detect low levels of PrPSc in infected mouse brains as well as some prion-related bands in the undigested parallel samples. Therefore, we conclude that in infected murine brains, as opposed to infected human brains or infected mouse cells, low levels of proteinase K resistant PrP are present in a fully or partially reduced form. Whether such PrPSc molecules are formed independently or join a seed of oxidized PrPSc molecules after formation is currently unknown.
Similar to brain PrPSc, detection of cell PrPSc by the pAb RVC could be restored in both cell lines when samples were reduced by MMA before being subjected to immunoblotting (Fig. 3B). This finding is consistent with the notion that the lack of PrPSc recognition by RVC indeed relates to the oxidative state.
Previous studies on prion-infected cells demonstrated that the formation of PrPSc from PrPC is a slow multistep process, which may include a variety of intermediate PrP states [28], [29]. To investigate whether PrP Helix-3 Met oxidation occurs before the acquisition of PK resistance, we examined the oxidation state of putative PK sensitive intermediate forms. Several experimental approaches indicated the presence of such intermediates, denominated either PrPSc-sen (from PK sensitive) or PrP* [30]–[32]. While these forms were never proven directly to be infectious, they were shown to present characteristic PrPSc properties. Intermediate PK sensitive PrP forms may be present as aggregates and require extensive or partial denaturation to be recognized by anti-PrP antibodies, as is the case for PrPSc [33]. To determine the oxidation state of intermediate forms of PrP, we subjected Sarkosyl extracted control and prion-infected brain samples from murine and human subjects to sucrose gradients, as previously described for PrP* [30]. Fractions of the centrifuged gradients (light to heavy) were collected and digested in the presence or absence of PK before immunoblotting with either the mAb 6H4 or pAb RVC. As shown in Figure 4, while PrP was similarly detected by both antibodies in the lighter gradient fractions of control samples (normal mouse and normal human brains), immunoblots of the prion-infected samples (RML infected mouse and CJD E200K heterozygous familial cases) with each of the antibodies showed very different results. Before proteinase K digestion, PrP was recognized by mAb 6H4 in most fractions of both human and mouse gradients, although the banding pattern of the protein resembled PrPC in the lighter fractions (1–3) and PrPSc in the heavier fractions. After proteinase K digestion, only the heaviest gradient fractions (mostly fractions 9–10) presented any form of PrP signal, indicating that while proteinase K resistant PrPSc is the most aggregated, partially aggregated PrP-sen forms (fractions 4–7) may also present the PrPSc banding pattern [22]. In contrast, when the undigested gradient fractions from the prion-infected brains were immunoblotted with the pAb RVC, the pattern of PrP recognition mostly resembled that of the normal brain homogenates. No PrP forms were detected in any of the intermediate or heavy fractions, except low levels of mouse PrPSc in the heaviest fraction. Following proteinase K digestion, the PrP signal mostly disappeared form all infected fractions, except traces in the last fraction of the mouse gradient, consistent with the experiments described in Figures 2 and 3. Similar results were obtained for brain samples from sporadic CJD patients (data not shown). Since the lack of recognition of PrP by the pAb RVC in the intermediate gradient fractions indicates that PrPSc-sen forms are as oxidized as the PrPSc-res forms, we conclude that oxidation of PrP accompanies the conformational change required for PrP aggregation and precedes the acquisition of proteinase K resistance during PrPSc formation. The fact that the low levels of reduced mouse PrPSc were detected by PVC only in the most aggregated fraction, both before and after PK digestion, further suggests that non-oxidized mouse PrP may join the prion seed following its formation from oxidized PrP molecules.
Figure 2C shows that as opposed to PrPC in normal human brains and undigested PrP in the brains of sporadic CJD patients, PrP was poorly detected in brains of heterozygous E200K PrP fCJD patients [18]. This finding indicates that the mutant E200K PrP molecules may be oxidized in these brains even in their initial conformational state. Indeed, the Met rich area of PrP Helix-3 is located C-terminally to the residue 200, which mutated form, E200K, is the most abundant among familial CJD patients. In fact, peptides embracing this region and comprising either E (peptide 195–213) or K (peptide 185–205) at position 200 were used more than a decade ago for the generation of specific (to wt or mutant) anti- PrP pAbs [34]. The pAb raised against the HuPrP peptide containing E at position 200 (designated in Figure 1b as the pAb RGM) did not recognize proteinase K sensitive PrP forms expressed in fibroblasts from homozygous E200K patients, suggesting that the pAb RGM specifically detected wt PrP as opposed to the mutant E200K form [34]. Next, brain extracts from heterozygous CJD E200K patients were immunoblotted with this antiserum. The results showed convincingly that the pAb RGM did not detect proteinase K-resistant PrP forms. Due to the general belief at the time that no covalent modification separates PrPC from PrPSc and that the only difference between the mutant and wt PrP proteins could be the mutation itself, it was concluded that in heterozygous E200K patients only the mutant protein (K at codon 200) acquires the proteinase K resistance property during disease [34]. This conclusion was then generalized using other methods and additional PrP mutations [35].
Based on the results described above for the pAb RVC, demonstrating that antibodies directed against Helix-3 methionines may not recognize PrPSc and since the peptide used for generation of the RGM antibody comprised both the 200 residue and the Helix-3 methionines, we now investigated whether this reagent does not recognize E200K PrPSc specifically or otherwise cannot detect all forms of human PrPSc, as described above for pAb RVC. To this effect, we immunoblotted brain homogenates from RML infected-mice as well as from sporadic or familial E200K CJD human cases and analyzed them in parallel with both anti-PrP mAb 6H4 and pAb RGM. As depicted in Figure 5a, pAb RGM, similarly to pAb RVC, did not recognize proteinase K resistant HuPrP in both sporadic and E200K familial CJD samples, and in addition detected poorly undigested forms of HuPrP E200K. Consistent with the results obtained with pAb RVC, pAb RGM detected low levels of MoPrPSc from infected brains, but did not detect PrPSc from infected cells lines, as depicted here for GT1 cells expressing chimeric Mo-Ha PrP. Similar results were obtained for PrPSc from ScN2a cells (not shown).
To assess whether pAb RGM has separate recognition sites for E at position 200 and for the Helix 3 Met residues, which may explain why this antibody did not detect 200K PrP in fibroblasts from E200K homozygous subjects [34], we repeated the inhibition experiments described for RTC and RVC in Figure 2 using pAb RGM. We found that the activity of pAb RGM, which detected only mouse and human PrP (not shown), was totally inhibited by several peptides covering the Helix-3 Met residues, including the one comprising the 203–211 PrP sequence. This prevents the residue at codon 200, regardless E or K, from forming part of the pAb RGM epitope (Figure 5b), indicating that the lack of recognition of the mutant PrP by pAb RGM is not related to the presence of K instead of E at position 200. In addition, and since the epitope of this antibody does not comprise M213, these results constitute the first direct evidence that oxidation of M206 (and/or M205) can also be considered as a covalent signature of PrPSc, as predicted by the theoretical studies.
To investigate why pAb RGM was unable to recognize the mutated PrP even though its epitope does not include the 200 residue, we examined the recognition of wt and rHuPrP E200K by RGM as well as by a panel of antibodies designed to detect oxidized and non-oxidized PrP forms. As shown in Figure 5C, while α-PrP mAb 3F4 recognized the wt and mutant rHuPrP chains equally, pAb RGM did not detect the mutant recombinant protein, as described before for mutant PrP expressed in fibroblasts from E200K homozygous patients [34]. Similar results were obtained when wt and mutant recombinant PrPs were immunoblotted with α-PrP mAb IPC2, the epitope of which includes Met 213 and the adjacent disulfide bond, both distant from the site of the E200K mutation. In contrast, only the rHuPrP E200K was recognized by pAb DZS18, a pAb raised against a MetO rich maize repetitive sequence, which was shown to recognize enriched PrPSc as well as other oxidized proteins [8]. These results suggest that Helix-3 methionines in PrP E200K may undergo facilitated or spontaneous oxidation both in cells [34] and in its α-folded recombinant form. Indeed, Figure 5D shows that the monomers of wt and E200K HuPrP (23–231) are indistinguishable by far-ultraviolet CD spectroscopy at 25°C and pH 4.5, but they differ in their thermal denaturation profile. Curve fitting yielded Tm values of 60±0.5°C and 54.5±1°C for the wt and E200K chains, which agrees with previously reported destabilization of this mutant PrP under a different setup [36]. These results, as well as previous experiments showing charged-induced alterations of E200K PrP [37] suggest that changed dynamics of Helix-3 in the mutated protein might favor transient exposures of the contained methionines to ROS. The spontaneous oxidation of E200K PrP also explains the poor recognition of undigested PrP from E200K patients brain by both RGM and RVC pAbs (Fig. 2, 5).
We have shown here that antibodies generated against reduced Helix-3 PrP Met residues could not recognize the majority of PrPSc forms. This finding applied to most PrPSc accumulated in scrapie-infected mouse brains and for all PrPSc accumulated in human CJD brains, as well as for all newly formed PrPSc in several prion-infected cell lines. Since reduction by MMA restored the recognition of brain and cell's PrPSc by these antibodies, we conclude that most Helix 3 Met residues in PrPSc, both as newly made in cells, or as long term accumulated in infected brains are oxidized. Our results also indicate that Met oxidation is also present in intermediate PrP forms, such as proteinase K sensitive and partially aggregated PrPs found in human and mouse infected brains, indicating that oxidation accompanies aggregation and precedes acquisition of proteinase K resistance by the nascent PrPSc molecules. In addition, we show here that pathogenic mutant PrP forms, as is the case for E200K PrP [18], are mostly oxidized even in the monomeric state. Taken together, our results are consistent with the conclusion that Helix-3 Met oxidation is an early event in the conversion of PrPC into proteinase K resistant PrPSc and thus in prion formation and subsequent disease pathogenesis.
From a structural point of view, Met oxidation involves the transformation from a moderated hydrophobic to a hydrophilic side chain. While in protein exposed residues this chemical change may not have major structural effects, sulfoxidation of buried Met may impact the stabilization interactions maintaining the proteins 3D fold. Indeed, this intuitive prediction is in agreement with our theoretical studies, which showed that changing the sulfur atom of Met206 and M213, both single or in combination, by a sulfoxide destabilizes the native α-folded [15], thereby allowing for a conformational conversion. Indeed, increasing the polarity side chain at any of the conserved Helix-3 Met residues (Met205, Met206 and M213) impedes the native state folding and the appearance of proaggregating states [11], [16], [17], [38], [39]. Then, from these studies it can be proposed that the tolerance for the PrP α-fold is determined by the redox state of the Helix-3 Met residues and that the intolerance for the native state increases the probability of the productive conversion pathway.
Surprisingly, our results suggest that raising antibodies specific for PrPC is not a difficult task. The Met rich area in Helix-3 appears to be quite immunogenic, as deduced by the fact that even immunization of rabbits with the relatively large peptide spanning amino acids 195–213 yielded antibodies against the Met rich area (see Figure 6). So was the case for the mAb IPC2, which was produced following the immunization of mice with full length recombinant mouse PrP [7]. Reagents similar to pAb RGM and RVC may have been produced in many laboratories, but their true meaning not understood. In contrast, raising antibodies against oxidized PrP peptides that may be specific for proteinase K sensitive and resistant forms of PrPSc has been unsuccessful thus far. This difficulty may relate to the well-established immunological barrier that precludes recognition of oxidized peptides by T-cell receptors [19], [20]. Such immunological phenomena may partially explain the apparent lack of immune response against PrPSc in all species.
While the failure of our antibodies to recognize aberrant forms of mouse and human PrP was mostly quantitative, a recent MS study failed to detect high levels of Met213 oxidation in hamster PrPSc [40]. Indeed, while MS may be the method of choice to establish the presence of covalent modifications in proteins, its use for quantification of oxidation in this specific case may be limited. First, the labile character of sulfoxidation of Met residues does not allow for accurate separation of in vivo and in vitro modifications [41]. Furthermore, as opposed to detection of full length mature proteins by specific antibodies, MS detection operates on soluble peptides produced by proteolysis, each of which has different recovery pattern efficiencies, even in the same protein. Indeed, it was previously shown that recovery of the PrP tryptic peptide including M213 is quite poor, and that the recovery is even less efficient for the peptide including M206 [42]. In this study, we were unsuccessful in recovering the Helix3 area of HuPrP (23–230) E200K for MS analysis, likely because the mutation, which adds a digestion site for trypsin, generated labile peptides that could not be identified with significant yield. Eventhough, and given the right conditions, we assume that a combination of immunological reagent sand MS are the right methods to look for modifications in this and other proteins.
We also describe in this study how an antibody believed to detect E at codon 200 of wtPrP actually recognized a reduced Met Helix-3 sequence [34]. The reason for such misconception was that, if, as generally believed, no covalent modifications separated between the different forms of PrP, than the epitope of an antibody that recognizes wt PrP (E at codon 200) but not mutant PrP (K at codon 200) should include E at codon 200. We have shown here that despite the accuracy of the old results, the previous interpretation, suggesting that only mutant PrP converts into proteinase K resistant PrP in the brains of heterozygous patients may be mistaken in view of our new knowledge. While pAb RGM indeed did not detect PrPSc in brains from heterozygous E200K CJD patients (Figure 5), similar results were obtained for PK resistant PrPSc in brains from sporadic CJD patients, which comprise E at position 200. In addition, our results suggest that pAb RGM did not recognize E200K PrP in cells from homozygous subjects not because it carries K at codon 200, but because this mutant PrP form may be present in an oxidized form, as shown here for rHuPrP E200K. Most importantly, the finding that E200K PrP can undergo spontaneous oxidation at Helix-3 Met residues constitutes the first mechanistic clue explaining the late onset spontaneous appearance of CJD in carriers of pathogenic PrP mutations. Once oxidized, the conformation of the mutant PrP may be irreversibly impaired. We speculate that oxidative events may facilitate spontaneous CJD outbreaks in subjects carrying designated PrP mutations, as is the case for E200K mutation carriers. Indeed, the prevalence of familial CJD increases with age [43], [44], as in the case for oxidative insults [45]. Whether oxidized mutant PrP can serve as a template for wt PrP conversion in heterozygous cases remains to be established.
While our results suggest that the oxidation of PrP forms may play a role in the formation of PrPSc, we have still to elucidate the conditions, kinetics and mechanism that lead to the initial irreversible oxidation of wt PrP Helix 3 Met residues. Interestingly, it has recently been shown that when fibrillar assemblies of recombinant PrP chains are annealed (by heat), they can transmit prion infectivity to wt animals [46], a result that could not be obtained with other recombinant PrP preparations [47]. It would be interesting to test whether synthetic prions as well as prions arising from diverse PMCA protocols [48]–[50] include oxidized PrP forms.
Based mainly on the fact that PrP ablated mice did not suffer from severe neurological damage [51], it was generally assumed that the function of PrPC is not associated with prion disease pathogenesis. However, we show here that oxidation of Met residues on PrPC, which may relate to its proposed antioxidative function [5], may be an essential step in acquisition of the aberrant PrPSc conformation. In fact, the association between oxidative stress and PrP conversion may link the activity of the prion proteins with other neurodegenerative conditions affected by stress and oxidation, such as ALS, AD and Parkinson's diseases [2], [52], [53], as well as to normal aging [45].
Animal experiments were conducted under the guidelines and supervision of the Hebrew University Ethical Committee, which approved the methods employed in this project. Brain human samples were received following postmortem examinations from the Pathology Department of the Hadassah University Hospital. Immunobloting analysis, as that described in this manuscript (in search of PrPSc), is part of the routine pathological protocol applied on brains from suspected CJD cases. Our laboratory in the Hadassah Department of Neurology is the national referral center for CJD diagnosis (genetic and biochemical testing). The testing of these samples was approved by both the safety and ethical authorities of the Hadassah University Hospital. Since all cases of CJD and alike negative controls are unable to sign for such tests long before their death due to their medical condition, the relatives of these patients provided informed written consent for PM studies. Enabling close relatives to provide such consent is the standard policy of the Israeli Ministry of Health.
PrP peptides were synthesized on a Liberty peptide synthesizer with a Discover single mode microwave module, using standard Fmoc chemistry. Amino acids were purchased from Luxembourg Bio Technologies, except for Fmoc-Met(O)-OH, which was purchased from Novabiochem. Peptides were cleaved from the resin by treatment with a mixture 95% trifluoroacetic acid, 2.5% water, 2.5% triisopropylsilane, and precipitation with cold diethylether. The peptides were purified on a Vydac C8 semipreparative column using gradients of 5% to 60% acetonitrile in water, with 0.1% trifluoroacetic acid (TFA) in both solvents. The mass of the peptides was measured using an Applied Biosystems Voyager-DE Pro MALDI TOF mass spectrometer and verified to be within ±1 Da of the theoretical mass. The purified peptides were lyophilized with 30% acetic acid to remove residual TFA.
Recombinant HuPrP(23–230) wt (with M129) and E200K chains were produced, purified and refolded into the α-form from their pET11a constructs using oxidized glutathione for disulfide bond formation and including 2 mM Met in refolding buffers [54], [55]. The mutant chain was generated by site-directed mutagenesis using QuickChange protocols with the following primers: 5′-GAAGTTCACCAAGACCGACGTTAAG-3′ (forward) and 5′–CTTAACGT CGGTCTTGGTGAACTTC-3′ (reverse). Before their use, proteins were equilibrated by dialysis in 10 mM NaAc pH 4.5 containing 50 mM NaCl and 0.5 mM citrate and characterized both as monomers by dynamic light scattering using DynaPro Titan spectroscatter (Wyatt Technology). CD spectra were recorded using a Jasco-810 spectropolarimeter operating at 25°C, and using 0.1 cm pathlength cuvettes and about 13 µM protein concentration solutions. Thermal denaturation experiments were performed by following the changes in the ellipticity at 220 nm as the samples were heated from 15°C to 85°C at the rate of 1 degree/min.
Designated PrP peptides were coupled to activated KLH (Sigma) and inoculated into rabbits while emulsified into Complete Freund's Adjuvant for the first immunization and Incomplete Freund's Adjuvant for subsequent injections. Most peptide immuniziations were performed at the animal facility of the Hebrew University-Medical School, except the ones for the RVC antibody which was produced by GenScript Inc (NJ, USA). KLH coupled with the Cys-KM peptide was first oxidized with 20 mM H2O2. After 15 min incubation at 37°C, the reaction was quenched by addition of 20 mM of free methionine before addition to the adjuvant. Following 3 lines of immunization, serum samples from all immunized rabbits were tested for their anti-PrP activity. Rabbits with positive sera were immunized once again before final collection of blood. Antisera were purified by affinity chromatography, using for retention either peptides (RVC antibody) or Protein A (RTC and RGM antibodies).
Brain samples of normal humans and patients with confirmed sporadic and genetic E200K CJD were obtained from Hadassah University Hospital Pathology department. Brains from mice infected with the RML scrapie prions and from golden hamsters inoculated with Sc237 prions were provided by the Animal Facility of the Hebrew University-Medical school. Brain homogenates (10% w/v) were prepared by repeated extrusion through an 18-gauge followed by a 21-gauge needle in phosphate-buffered saline (PBS), aliquoted and maintained at −70°C until use.
Naïve and scrapie infected ScN2a[23] and ScGT1 cells [24] were washed, collected and lysed in 1 ml lysis buffer (100 mM Tris pH 7.4, 100 mM NaCl, 1% NP40, 1 mM EDTA) for 10 min. Samples were then centrifuged at 2000 rpm for 15 min at 4°C, and the supernatant was concentrated by methanol precipitation. Pellets were resuspended in 2% sarkosyl/STE buffer (10 mM Tris–HCl, pH 7.5, 10 mm NaCl, 1 mM EDTA). Protein content was determined by a BCA kit (Pierce). Equal amounts of protein were treated in the presence and absence of 40 µg/ml proteinase K for 30 minutes in 37°C. Digestion was stopped by the addition of a protease inhibitor complex (Complete Protease Inhibitor Cocktail Tablets, Roche) before subjecting the samples to denaturation by boiling in the presence of sample buffer. Samples were then immunoblotted with the designated anti-PrP antibodies.
Normal and prion-containing brain samples were homogenated at 10% (W/V) in 10 mM Tris, pH 7.4 and 0.3 M sucrose. Proteinase K digestions were performed by incubating 30 µl of 10% prion-infected brain homogenates with 2% sarkosyl for 30 min at 37°C with 40 µg/ml protease. Control samples were incubated at 37°C in the absence of proteinase K). After boiling in sample buffer, samples were subjected to SDS PAGE and immunoblotting with the diverse anti-PrP antibodies. For the inhibition experiments, nitrocellulose sheets comprising the transferred proteins were subjected either to a 1∶2000 dilution of the designated antibody alone or preincubated for at least 2 hours with the appropriate synthetic peptide (2 µg/ml). Immunoblots were developed with α mouse or α rabbit antibodies AP or HRP-conjugated secondary antibodies (Promega, Madison WI).
Proteinase K digested prion-infected cells or brain homogenates were treated with 6 M N-methylmercaptoacetamide (MMA) [22]. After 15 h of incubation at 37°C, samples were precipitated with 9 volumes of methanol (1 h, −80°C) and then centrifuged (10000 rpm, 30 min, 4°C). Pellets were washed twice with methanol and processed for SDS-PAGE analysis. When immunoblotting with IPC2, sample buffer was devoid of β-mercaptoethanol [7].
Sarkosyl extracted brain extracts from human and mouse (normal and prion infected were subjected to a sucrose gradient as previously described [30]. Shortly, 140 µl of 10% brain homogenates (mouse:normal and scrapie infected; human: normal and CJD), extracted in the presence of 2% Sarkosyl were overlaid on a sucrose gradient composed of layers of increasing concentrations of sucrose (10–60%). Gradients were then centrifuged for 1 h at 55000 rpm in a Sorval mini-ultracentrifuge and subsequently 11 samples of 120 µl were collected from the top to the bottom. In the prion infected gradient fractions were digested in the presence and absence of 40 µg/ml proteinase K before immunobloting with either α PrP mAb 6H4 or pAb RVC.
Human Prion Protein: P04156 (PRIO_HUMAN), Mouse Prion Protein: P04925 (PRIO_MOUSE), Hamster Prion Protein: P04273 (PRIO_MESAU).
|
10.1371/journal.ppat.1005027 | Age-Dependent Cell Trafficking Defects in Draining Lymph Nodes Impair Adaptive Immunity and Control of West Nile Virus Infection | Impaired immune responses in the elderly lead to reduced vaccine efficacy and increased susceptibility to viral infections. Although several groups have documented age-dependent defects in adaptive immune priming, the deficits that occur prior to antigen encounter remain largely unexplored. Herein, we identify novel mechanisms for compromised adaptive immunity that occurs with aging in the context of infection with West Nile virus (WNV), an encephalitic flavivirus that preferentially causes disease in the elderly. An impaired IgM and IgG response and enhanced vulnerability to WNV infection during aging was linked to delayed germinal center formation in the draining lymph node (DLN). Adoptive transfer studies and two-photon intravital microscopy revealed a decreased trafficking capacity of donor naïve CD4+ T cells from old mice, which manifested as impaired T cell diapedesis at high endothelial venules and reduced cell motility within DLN prior to antigen encounter. Furthermore, leukocyte accumulation in the DLN within the first few days of WNV infection or antigen-adjuvant administration was diminished more generally in old mice and associated with a second aging-related defect in local cytokine and chemokine production. Thus, age-dependent cell-intrinsic and environmental defects in the DLN result in delayed immune cell recruitment and antigen recognition. These deficits compromise priming of early adaptive immune responses and likely contribute to the susceptibility of old animals to acute WNV infection.
| While West Nile virus (WNV) infection preferentially causes severe neuroinvasive disease in elderly humans, the basis for this epidemiological linkage has remained uncertain. Here, we studied the impact of aging on WNV pathogenesis and immune responses using a mouse model of infection. Old mice showed increased lethality after WNV infection compared to adult mice, and this phenotype was associated with delayed antibody responses, and higher levels of virus infection in the blood, spleen, and brain. Detailed immunological and microscopic analyses revealed defects in germinal center development in the draining lymph node and impaired migratory capacity of naïve CD4+ T cells within days of WNV infection. This deficit was worsened by a separate age-related deficiency in the production of several chemokines that recruit leukocytes to inflamed lymph nodes. Thus, age-dependent defects in naïve CD4+ T cell trafficking and in the draining lymph node environment result in delayed initiation of antiviral responses and a failure to control WNV, which ultimately contributes to higher rates of mortality after infection.
| Aging is linked to a decline in immunity that causes increased susceptibility to infectious diseases and reduced vaccine efficacy in the elderly population. This process, termed immunosenescence, is the consequence of age-dependent changes to multiple components of the innate and adaptive immune responses. For example, elderly individuals display elevated basal levels of pro-inflammatory cytokines and type I interferon (IFN) [1], and demonstrate reduced signaling following pathogen-associated molecular pattern (PAMP) recognition [2]; collectively this results in dysregulated cytokine responses [3]. Additionally, elderly individuals exhibit a shift from a naïve to a memory repertoire in CD4+ T cells, CD8+ T cells [4–6], and B cells [7], which decreases the number of cells capable of recognizing new antigens encountered during primary infections. Cell-intrinsic defects also have been identified in different arms of the adaptive response. Naïve CD4+ T cells from old individuals display reduced activation, immunological synapse formation, and proliferation upon T cell receptor (TCR) engagement [8,9]. This leads to a diminished ability of CD4+ T cells to prime B cells during germinal center reactions [10,11] and, along with intrinsic B cell defects [12], results in reduced affinity maturation, class switching, and memory responses [11].
West Nile Virus (WNV) is a globally important mosquito-transmitted RNA virus that causes severe disease preferentially in the elderly as a consequence of enhanced dissemination to the central nervous system (CNS) and infection and injury of neurons [13]. After inoculation in the skin, WNV infects dendritic cell subsets and traffics to the draining lymph node (DLN) where it replicates locally and spreads to other secondary lymphoid and visceral organs [14]. In most human cases, WNV is cleared after induction of a robust innate and adaptive immune response. However, in the elderly, the virus crosses the blood-brain barrier at increased frequency and infects neurons of the brain and spinal cord; this is associated with a 5 to 10% case-fatality rate [14]. Based on animal studies, rapid clearance of WNV infection requires the concerted actions of multiple aspects of the immune system, including antiviral cytokine responses by myeloid cells [15], complement activation [16], the rapid induction of virus-specific neutralizing antibodies [17,18], helper, effector, and regulatory functions of CD4+ T cells [19–22], and priming of cytotoxic CD8+ T cells [23,24].
Analogous to humans, old mice are more vulnerable to WNV infection. One study reported a decreased number and function of WNV-specific CD8+ T cells [21], although the mechanistic basis of this defect was not defined. We hypothesized that age-dependent defects in B cell responses also contributed to disease susceptibility in old mice. Following WNV infection of young adult mice, an early T-independent and T-dependent IgM response develops, which limits viremia and prevents spread to the brain [17,18]. Subsequently, T-dependent class-switching and affinity maturation leads to the development of plasma cells that secrete higher affinity IgG antibodies, which neutralize virus in peripheral tissues [22]. Some of these antibody-secreting cells migrate to the CNS where they likely help to control and clear WNV infection [25].
In this study, we detected age-dependent defects in the early antigen-specific IgM and IgG response that correlated with increased viral titers in the serum, spleen and brain, and higher rates of WNV mortality. Using a series of adoptive transfer experiments coupled with two-photon intravital microscopy, we identified cell-intrinsic defects in naïve CD4+ T cells as well as reduced cytokine and chemokine levels in the DLN environment, both of which contributed to reduced cellular accumulation, delayed germinal center development, and immune responses. The cell-intrinsic defect was observed specifically in naïve CD4+ T cells and not in CD19+ B cells and was associated with an inability to migrate efficiently into and within the inflamed DLN. Thus, age-dependent immune defects that occurred within the first few days of WNV infection resulted in delayed immune cell recruitment, antigen recognition, and priming. These very early defects contributed to the failure to control WNV infection and prevent death.
We evaluated disease severity after WNV infection in adult and old mice. Compared to 4 month-old adult C57BL/6 mice, 18 month-old old syngeneic mice exhibited increased mortality (56% versus 15%, P < 0.01, Fig 1A) following subcutaneous infection with WNV (New York 2000 strain), corroborating results from a published report [21]. To understand the basis for the increased lethality, we measured viral burden in tissues. Older mice sustained increased levels of WNV in the serum at day 6 after infection (3.5-fold, P < 0.001 Fig 1B), in the spleen at days 4 (7-fold, P < 0.05) and 6 (17-fold, P < 0.01 Fig 1C) after infection, and in the brain at day 9 after infection (20-fold, P < 0.01 Fig 1D), which likely caused the higher mortality rate.
The early humoral immune response limits WNV infection and dissemination in adult mice [17,18]. We hypothesized that age-dependent defects in the WNV-specific humoral immunity might contribute to the more severe clinical phenotype in old mice. Accordingly, we assessed the WNV-specific antibody responses at days 5, 8, and 15 after infection in adult and old mice. Serum from old mice had less neutralizing activity than that from adult mice at all time points tested (2.7 to 4-fold, P < 0.05, Fig 1E). Also WNV-specific IgM and IgG antibody titers against the viral envelope (E) protein were lower at day 8 after infection in the aged mice (2.9 and 5.1-fold, P < 0.01, Fig 1F and 1G). By day 15, overall serum WNV E protein-specific IgM and IgG responses were equivalent in the surviving adult and old mice although a qualitative defect remained in the recognition of a dominant neutralizing epitope on domain III of the E protein [26,27] in old mice in both the IgM and IgG fractions (data not shown). Surviving old mice also had defects in durable humoral immunity, as their serum had reduced neutralization capacity at 90 days post infection (3.5-fold, P < 0.05, Fig 1H), which correlated with fewer WNV-specific long-lived plasma cells in the bone marrow (4.5-fold, P < 0.01 Fig 1I and 1J). Thus, old mice had delayed and decreased humoral immune responses during both the acute and memory phases after WNV infection.
To assess the basis for the defects in the early B cell response in WNV-infected old mice, we focused on the DLN. Following subcutaneous infection in the footpad, WNV traffics to the popliteal LN where an early B cell response is initiated [28]. Within the germinal centers of the DLN, T follicular helper (TFH) cells provide signals to germinal center B (GC B) cells to promote class switching and affinity maturation [29]. The accumulation of TFH (CD4+ PD1+ CXCR5+) and GC B (CD19+ Fas+ GL7+) cells was decreased in old mice relative to adult mice at days 4 and 6, although nearly equivalent levels of TFH and GC B cells were present at 8 days post infection (Fig 2A and 2C). These findings were corroborated by immunofluorescence microscopy analysis, which showed a reduced number of germinal centers at day 6 after infection (Fig 2E). The delayed germinal center development in the old mice matched the temporal pattern of antibody titers. At 6 days after infection, similar levels of CD28 were observed on CD4+ T cells and TFH cells from old and adult mice (S1A and S1B Fig), suggesting that decreased expression of co-stimulatory molecules was not responsible for the delayed germinal center development in old mice. Consistent with this pattern, old mice had a proportionately higher number of CD4+ T cells and TFH cells expressing the co-stimulatory molecule OX40 (CD134) (S1C and S1D Fig). TFH cells from old mice did exhibit a small yet statistically significant reduction (11%, P < 0.01) in CXCR5 expression at 6 days post infection (S1E Fig). By 8 days after infection, however, a higher percentage of the CD4+ T cells from old mice were phenotypically TFH, and adult and old mice had equivalent numbers of GC B cells (Fig 2A, 2B and 2D). Thus, in old mice, the blunted humoral response to WNV correlated with a delay in the kinetics of germinal center development.
During isolation of the popliteal lymph nodes for GC analysis, we noticed DLN from adult mice were substantially larger than those from old mice after WNV infection (Fig 2E and 2F). Because of this observation, we assessed how aging affected the total number of immune cells in the DLN at different time points after infection. In naive mice, we found increased cell numbers in the adult DLN relative to those from the old mice (6.6 x 105 versus 3.0 x 105 cells, P < 0.01 Fig 2G). After subcutaneous WNV inoculation, cell numbers increased rapidly in the adult mice, peaking at day 6. Cell accumulation was delayed in the DLN of old mice with fewer cells observed at days 2, 4, and 6 after infection (Fig 2G).
Cell subset analysis revealed that old mice had fewer CD4+ T and B cells in the DLN after WNV infection (Fig 2H and 2K), which could reflect reduced in situ proliferation and/or defects in cell trafficking. We monitored cell activation, as judged by surface acquisition of the marker CD69, and proliferation through expression of the nuclear protein Ki-67. As reported previously [28], CD19+ B cells in the DLN of adult mice were activated shortly after WNV infection with greater than 85% of cells expressing CD69 by day 2 (Fig 2L). The kinetics of CD69 up-regulation were equivalent on B cells in adult and old mice, although in old mice CD69 expression remained higher at day 8 (Fig 2L and data not shown). An analogous pattern of CD69 expression was seen in adult and old CD4+ T cells, although a smaller subset of cells was activated (e.g., 31% at 2 days after infection) (Fig 2I). We observed a similar small fraction of B cells and CD4+ T cells staining for Ki-67 in the adult and old mice between days 0 and 4 after infection (Fig 2J and 2M). An increase in Ki-67 expression was noted at day 6 for CD4+ T cells in both age groups, with higher levels in old mice at both days 6 and 8 after infection (Fig 2J). We also detected a small decrease in the frequency of Ki-67-positive proliferating B cells in old mice at days 6 and 8 after infection (3.5% to 2.3% at day 6 and 5.5% to 4.2% at day 8, P < 0.05, Fig 2M), which could reflect delays in the germinal center reaction observed in the old DLN (Fig 2A and 2C). These findings suggest that impaired activation or proliferation of lymphocytes did not account for the reduced cellular accumulation in the DLN of the old mice.
Diminished cellularity of the DLN of old mice was not unique to viral infection, as similar data was obtained with an inflammatory stimulus. Old mice immunized with ovalbumin emulsified in complete Freund’s adjuvant (CFA) had fewer CD4+ T cells and B cells in the DLN at early time points compared to adult mice (Fig 3A–3C). Old mice also had fewer IL-2-secreting antigen-specific CD4+ T cells at day 7 after immunization, and fewer germinal center B cells and TFH cells at day 6 after immunization (Fig 3D–3F).
We hypothesized that the reduced lymphocyte accumulation in the DLN of old mice soon after WNV infection or ovalbumin immunization could be due to cell-intrinsic defects in the lymphocytes and/or environmental defects in the DLN. To evaluate these possibilities, we sorted naïve CD4+ CD44- CD62L+ T cells from lymphoid tissues of adult and old mice and labeled purified cells ex vivo with different fluorescent dyes (Fig 4A). As previously reported, old mice had fewer naïve CD4+ T cells compared to adult mice (S2 Fig and [5]). Equal numbers of differentially labeled adult and old naïve sorted CD4+ T cells were mixed at a 1:1 ratio and injected into recipient adult or old mice that had been infected two days earlier with the Kunjin strain of WNV (WNV-KUN), a less pathogenic variant that can be studied under A-BSL-2 conditions. Importantly, subcutaneous infection of old mice with WNV-KUN showed similar defects in cellular accumulation in the DLN compared to adult mice, with kinetics that mirrored that seen with the virulent WNV New York strain (S3 Fig and Fig 2). One hour after transfer of the labeled donor T cells into the WNV-KUN infected mice, the DLN of recipient mice were harvested. In both the adult and old recipients, greater numbers of adult donor T cells were detected within the DLN suggesting an intrinsic trafficking defect of the naïve CD4+ T cells from old mice (Fig 4B). We also observed reduced accumulation of adult and old donor cells in the old compared to adult recipients (Fig 4B), suggesting that the DLN environment separately contributed to the lower numbers of naïve CD4+ T cells recruited after infection. Similar defects were observed in the homing of cells to the spleen indicating that reduced trafficking could be a general property of naïve CD4+ T cells and the lymphoid environment of old mice (Fig 4C).
We used intravital two-photon microscopy to evaluate the trafficking patterns of adult and old naïve CD4+ T cells in vivo within lymphoid tissues. Dye-labeled naïve CD4+ T cells from adult or old mice were mixed at a 1:1 ratio and adoptively transferred into adult recipient mice that had been infected with WNV-KUN two days prior. Donor cells from both adult and old mice efficiently rolled and arrested on the luminal side of the high endothelial venules (HEV) in the DLN of recipient infected mice (Fig 4D and 4E, S1 Movie and S2 Movie), although the appearance of donor T cells from old mice in the HEVs was delayed by ~15 minutes relative to donor cells from adult mice. Remarkably, naïve CD4+ T cells from old mice extravasated less efficiently compared to T cells from adult mice, with many cells unable to undergo the rapid cell shape deformation required for transendothelial migration across HEV (Fig 4F). This result suggested that although naïve CD4+ T cells from old mice could bind HEV, they were less efficient at entering the parenchyma of the DLN due to defective diapedesis. When the motility of extravasated T cells was analyzed in explanted lymph nodes from WNV-KUN infected adult recipient mice, adult donor naive CD4+ T cells moved with greater speed and displacement than the old donor naïve CD4+ T cells (Fig 4G–4I, S3 Movie and S4 Fig). However, neither adult nor old donor naive CD4+ T cells showed signs of chemotactically biased motion (Fig 4K). Computational analysis [30] predicted that the motility defects of naïve CD4+ T cells from old mice would delay encounter with antigen 500 μm away by ~30 hours (Fig 4L), which would contribute to the slowed kinetics of the humoral immune response in these animals. A trend towards reduced migratory capacity of old, compared to adult naïve CD4+ T cells also was observed in the LN of naive mice, although the magnitude of the difference was smaller (Fig 5A–5E).
Decreased expression of surface adhesion and chemoattractant receptors on naïve CD4+ T cells from old mice might explain the reduced transmigration and motility phenotype. However, when we assessed the expression of the surface adhesion molecules CD62L, PSGL1, LFA1, and VLA-4 as well as the chemokine receptors CCR7 and CXCR4 on naïve CD4+ T cells from adult and old mice, no significant differences in the geometric mean fluorescence intensity of expression of any of these molecules were observed (S5 Fig). We also did not find differences in the levels of filamentous actin in adult and old naïve CD4+ T cells following CCL19 (MIP-3β) and CCL21 (Exodus-2) stimulation (S6A Fig), suggesting that cytokine recognition, signaling through the Rac/Rho pathway, and actin polymerization/depolymerization were intact. Reduced T cell binding to ICAM-1 also could lead to diminished transmigration [31]. However, we observed no difference in the ability of naïve CD4+ T cells from adult or old mice to bind ICAM-1 coated substrates or migrate through ICAM-1 coated transwells towards CCL19 and CCL21 chemoattractants (S6B and S6C Fig).
To link the reduced GC responses in old mice (Fig 2) with cell-intrinsic migratory defects of the naïve CD4+ T cells (Fig 4), we adoptively transferred sorted naïve CD4+ T cells from adult or old mice to recipient TCR β/δ -/-mice (lacking T cells) that had been infected 2 days earlier with WNV (Fig 6A). We quantified CD4+ T cell and GC responses 6 days later at 8 days after infection. In all recipient TCR β/δ -/-mice, very low GC responses were observed in the DLN, possibly because few antigen-specific CD4+ T cells reached this site (data not shown); however, GC responses were detected in the spleen. Animals receiving donor cells from old mice had a reduced percentage of CD4+ T cells in the spleen relative to those receiving adult donor cells (Fig 6B), and this corresponded with a lower percentage of TFH and GC B cells (Fig 6C and 6D). These data suggest that intrinsic CD4+ T cell defects alone, independent of other age-dependent immunological defects, contribute to the reduced GC response following WNV infection in old mice.
We next tested whether B cells also had a trafficking defect by adoptively transferring a 1:1 ratio of differentially dye-labeled adult and old B cells into adult recipient WNV-KUN infected mice and quantifying cell recruitment. No major differences were observed in the numbers of adult and old donor B cells in lymphoid tissues at 6 hours after transfer (Fig 7A and 7B), suggesting that impaired cell accumulation was not a general phenomenon that affected all lymphocyte subsets from old mice. Similar results were obtained in the spleen at 1 hour after transfer (data not shown). CD19+ B cells also did not show defects in migration in explanted lymph nodes (Fig 7C–7F and 7K and 7L) or spleen (Fig 7G–7J and S4 Movie).
To further assess possible cell-intrinsic defects in B cells from old mice we adoptively transferred an equal number of bone marrow cells from adult or old CD45.2 mice into separately irradiated CD45.1 adult recipient mice. We also transferred bone marrow cells from adult B cell-deficient (μMT) to provide equivalent T cell help (Fig 8A). Twelve weeks later, the transfers were confirmed by analyzing circulating CD19+ B cells in blood (Fig 8B). Recipient mice were infected with WNV and serum was harvested for evaluation of WNV-specific antibody responses at days 5, 8, and 15 after infection. Notably, recipient mice reconstituted with B cells from adult or old mice had similar WNV-specific IgM and IgG responses (Fig 8C and 8D).
The reduced accumulation of adult or old donor cells in old recipient lymphoid tissue (Fig 4B and 4C) suggested a separate environmental defect independently contributed to the reduced migration of naïve CD4+ T cells. To address the basis for these findings, we measured the levels of pro-inflammatory cytokines and chemokines in the inflamed DLN (Fig 9). Within one to two days of WNV infection, lower levels of several chemokines were apparent in the homogenates of DLN from old mice. This included chemoattractants for monocytes (MCP-1 (CCL2)), granulocytes (MIP-1α (CCL3)), NK cells (MIP-1β (CCL4)), B cells (BLC (CXCL13)), and was associated with the subsequently decreased production of proinflammatory cytokines including IL-1α, IL-2, IL-6, and IFN-γ, and the anti-inflammatory cytokine IL-10. Relevant to our naïve CD4+ T cells findings, the DLNs from old mice had nearly a 10-fold decrease in the naïve T cell chemoattractant CCL21 at day 2 after infection (Fig 9N). Consistent with the blunted production of chemokines, we also observed diminished accumulation of NK cells, γδT cells, macrophages, and dendritic cells in the DLN of old mice after WNV infection (Fig 10). Thus, in addition to cell-intrinsic defects of naïve CD4+ T cells in old animals, the DLNs of old mice separately have attenuated inflammatory cytokine and chemokine responses, which reduced the recruitment of other cell types that orchestrate adaptive immunity against WNV infection.
The decline of the immune system with age correlates with increased susceptibility to many infectious diseases and a decreased response to vaccines. In this study, using a mouse model of aging in the context of WNV infection, we identified cell-intrinsic defects in naïve CD4+ T cells as well as environmental defects that resulted in delayed and reduced lymphocyte accumulation in the inflamed DLN of old mice at very early time points. These defects delayed the development of the early germinal center reaction, which resulted in blunted humoral immune responses that likely contributed to increased viral titers in the blood and peripheral organs and enhanced vulnerability to lethal infection. Even small differences in WNV-specific antibody levels at early times after infection affect dissemination to the brain and clinical outcome [17,18]. As mice lacking B cells uniformly succumb to WNV infection [17], the delayed adaptive immune response in old mice still provided a significant level (~50% survival) of protection. Old mice surviving the acute stage of infection eventually developed similar overall levels of WNV-specific antibody titers and germinal center B and TFH cells compared to adult mice, although qualitative defects in the humoral immunity remained. In addition, old mice had deficits in their long-lived plasma cell memory response relative to the adult mice. This suggests that despite GC B cells and TFH cells reaching equivalent levels in the surviving adult and old mice after WNV infection, functional defects persist in the germinal center reactions of old mice, which contribute to blunted memory responses. While further study is warranted, part of the age-dependent differences in memory phenotypes could be associated with differential expression of pro-survival (e.g., Bcl-2) or pro-apoptotic (e.g., TRAIL) proteins in B cells [32,33] or less likely the lower levels of CXCR5 observed on TFH cells.
Although the association between aging and susceptibility to viral infection has been described [34,35], our results support a model in which immune system deficits occurring shortly after infection determine clinical outcome. Within 48 hours of WNV infection, we detected a diminished immune response in the DLN of old mice that included reduced accumulation of NK cells, dendritic cells, macrophages, and B and T cells. Associated with this phenotype were decreased levels of several cytokines and chemokines, including the naïve T cell chemoattractant CCL21 and the B cell chemoattractant CXCL13. These results are consistent with reports showing disrupted cytokine production after pattern recognition receptor stimulation in innate immune cells from old mice and humans [3]. Defects in the cellular production of cytokines and chemokines in the LN of old mice likely contribute to the reduced accumulation of innate and adaptive immune cells after inflammatory or infectious stimuli. Consistent with this idea, reduced levels of CCL19 and CCL21 were detected in the spleen of old mice immunized with OVA absorbed on alum [36]. These chemokine and recruitment defects also may lead to defective priming of CD8+ T cells, as the DLN has an important role in priming these cells during a systemic viral infection [37]. Thus, our data provides an explanation why qualitative and quantitative defects in antigen-specific CD8+ T cells responses also exist after WNV infection of old mice [21].
Prior studies have reported reduced frequencies of naïve CD4+ T cells in old animals, similar to our observations, due to long-term antigen exposure, conversion to memory phenotypes, and repertoire contraction [5]. We also observed a separate, independent cell-intrinsic defect in the naïve CD4+ T cells from old mice. By adoptively transferring adult or old naïve CD4+ T cells into recipient mice lacking T cells, we found that intrinsic defects in the naïve CD4+ T cells of the old mice contributed to reduced GC responses independent of other factors, analogous to previous immunization models with transgenic donor cells [10]. Although others have identified naïve CD4+ T cell-intrinsic defects following antigen encounter [38], we identified an earlier, age-dependent deficit in the migratory capacity of the naïve CD4+ T cells. Soon after adoptive transfer of naïve CD4+ T cells from adult and old donors into adult recipient mice, cells from the old mice accumulated in DLNs to lower levels as compared to cells from adult mice. Intravital microscopy revealed that the trafficking of naïve CD4+ T cells from old mice was impaired at the diapedesis step specifically, as cell rolling and adhesion in HEVs was delayed only slightly relative to cells from adult mice. The reduced trafficking of naïve CD4+ T cells to DLN in old mice suggests that immune surveillance declines during aging. Once naïve CD4+ T cells from old mice entered the parenchyma of the inflamed LN, we also observed reduced motility. Although the T cell trafficking defects may seem modest, mathematical modeling [30] predicted that the reduced motility in the DLN could lead to many hours of delay in finding cognate antigen. Combined with the additional homing defects, this would hinder the induction of antigen-specific adaptive immune responses substantially. Due to the rapid nature of viral replication and dissemination, such a lag could and likely does alter disease outcome.
What is the cell-intrinsic defect of naïve CD4+ T cells from old animals that results in a reduced migratory capacity? A diminished capacity of naïve CD4+ T cells from old animals to activate and proliferate following TCR stimulation has been reported [38]; this has been ascribed to reduced cytoskeletal rearrangement [39,40], aberrant surface glycosylation [41], and altered phosphorylation of key signaling molecules [42]. These defects in naïve CD4+ T cells from old animals can be overcome to some extent by cytokine (e.g., IL-2) stimulation [43] or treatment with O-sialoglycoprotein endopeptidases [44]. We found no differences in the expression of cell-surface molecules important for chemokine binding or migration in adult or old mice. We also did not detect differences in the ability of naïve CD4+ T cells from adult and old mice to bind ICAM-1 or undergo cytoskeletal rearrangement after chemokine activation. Instead our studies suggest a key cell-intrinsic migratory defect in naïve CD4+ T cells from old mice that occurs in vivo prior to antigen stimulation at very early times relative to infection. Impaired function of non-muscle myosin IIA in CD4+ T cells from elderly humans and old mice has been linked to defective migration ex vivo [40] and an inability to deform nuclei [45] respectively, which are processes critical to successful diapedesis; however, we found no difference in the ability of naïve CD4+ T cells from adult and old mice to migrate across an ICAM-1-substrate ex vivo. Alternatively, mTOR inhibition rescues some of the age-related functional defects in naïve CD4+ T cells from old mice, suggesting that a reduced metabolic capacity could affect energy stores [44,46]. As migratory activities require extensive energy-dependent cytoskeletal rearrangements, generation of leading edges and uropodia, lower energy reserve levels could compromise this process.
In comparison to naïve CD4+ T cells, and even though humoral responses were blunted, we did not observe significant defects in the homing, motility, or function of CD19+ B cells from old mice. Intravital microscopy in explanted lymphoid tissues showed similar movement and displacement by B cells from old and adult mice. Moreover, in experiments in which bone marrow from old and adult mice were transferred into irradiated adult recipient mice and then infected with WNV, differences in the antiviral antibody response were not apparent; this confirms previous reports suggesting that newly derived B cells are functionally more intact than long-lived B cells in old animals [47]. These findings suggest that the defects in the naïve CD4+ T cells from old mice are not associated with a general functional decline of all lymphocyte subsets.
In summary, using a mouse model of WNV infection, we identified a specific age-dependent decline in the trafficking of naïve CD4+ T cells. Efficient T cell trafficking is required for optimal antigen recognition in the DLN and for timely initiation of T-helper dependent antibody induction. Because the rapid production of neutralizing and inhibitory antibodies is essential for limiting WNV replication and dissemination [17,18], even a modest delay in the kinetics of the humoral response in elderly persons could increase susceptibility to virus infection. As functional deficits in the cells that coordinate early adaptive immune response and the LN microenvironment exist, strategies to correct these defects in at-risk individuals could improve responses to vaccines and limit replication of microbes that promote rapid disease pathogenesis.
This study was carried out in strict accordance with the recommendations in the Guide for the Care and Use of Laboratory Animals of the National Institutes of Health. The protocols were approved by the Institutional Animal Care and Use Committee at the Washington University School of Medicine (Assurance Number: A3381-01). Dissections and footpad injections were performed under anesthesia that was induced and maintained with ketamine hydrochloride and xylazine, and all efforts were made to minimize suffering.
Cell Trackers: Cell Tracker Orange CMTMR (5-(and-6)-(((4-Chloromethyl) Benzoyl) Amino) Tetramethylrhodamine), CellTrace CFSE (carboxyfluorescein diacetate, succinimidyl ester) and CellTrace Violet were purchased from Life Technologies. The following fluorochrome or biotin-conjugated antibodies were purchased from BD Biosciences, Biolegend, and eBioscience (clone name in parenthesis): CD3ε (145-2C11), CD4 (RM4-5), CD8α (53–6.7) CD11b (M1/70), CD11c (HL3), CD19 (1D3), CD28 (E18), CD44 (IM7), CD45R (RA3-6B2), CD62L (MEL-14), CD69 (H1.2F3), OX40/CD134 (OX-86), Ki67 (SolA15), PD1 (29F.1A12), FAS (Jo2), GL7, F4/80 (BM8), TCRγ (eBioGL3), NK1.1 (PK136), VLA4 (R1-2), LFA-1 (2D7), PSGL-1 (2PH1), CCR7 (4B12), CXCR4 (L276F12), and CXCR5 (2G8).
The WNV strain (3000.0259) was isolated in New York in 2000 and passaged once in C6/36 Aedes albopictus cells. The WNV-KUN strain (CH16532) was a gift of R. Tesh (World Reference Center of Emerging Viruses and Arboviruses, Galveston, TX) and amplified as described previously [48]. Viral titer was determined by plaque assay using Vero cells, as described previously [49].
C57BL/6 mice were purchased from The Jackson Laboratory (4 month-old) (Bar Harbor, ME) or acquired from the National Institute of Aging dedicated breeding colony at Charles River Laboratories (18 month-old) (Wilmington, MA). TCR β/δ-/- mice were purchased from The Jackson Laboratory and bred in the animal facilities at the Washington University School of Medicine. All mice were housed in a pathogen-free mouse facility at the Washington University. Mice were inoculated subcutaneously via footpad injection with 102 plaque-forming units (PFU) of WNV-NY or 104 PFU of WNV-KUN diluted in 50 μl of Hanks balanced salt solution (HBSS) supplemented with 0.1% heat-inactivated fetal bovine serum (FBS). For intravital imaging of the draining cervical lymph node, mice were infected subcutaneously in the nape of the neck.
At specified time points after WNV infection, serum was obtained by intracardiac heart puncture, followed by intracardiac perfusion (20 ml of PBS), and organ recovery. Organs were weighed, homogenized using a bead-beater apparatus, and titrated by plaque assay on BHK21-15 cells as described previously [49].
Bone marrow cells were collected from adult or old C57BL/6 (CD45.2) mice and adult B-cell deficient μMT mice. Adult or old bone marrow cells were mixed with an equal number of μMT bone marrow cells and 1 x 107 cells were transferred adoptively by retroorbital injection into 8 to 12-wk-old 800 cGy–irradiated B6.SJL (CD45.1) mice (National Cancer Institute). Twelve weeks later, reconstitution was validated by flow cytometry and mice were infected with WNV.
Mice were injected subcutaneously in the footpad with an emulsion containing CFA and 10 nmoles of OVA323–337 peptide, and DLN were analyzed at indicated time points. The ELISPOT assay has been described previously [50,51]. To evaluate MHC class II specific T cell responses in adult and old mice, 5 x 105 cells from the DLN were harvested at day 7 post-immunization and plated on 96-well plates with Immobilon-P membrane coated with anti-mouse IL-2 antibody. After overnight incubation, plates were washed, incubated with a secondary antibody and developed as previously described [50,51]. Spots were analyzed on dedicated equipment and software (Cellular Technology Ltd along with Immunospot software Version 5.0.9) to calculate the total number of antigen-specific CD4+ T cells in DLN.
The levels of WNV-specific IgM and IgG were determined using an ELISA against purified WNV E protein [52]. Focus reduction neutralization assays were performed on Vero cells after mixing serial dilutions of serum with a fixed amount (100 FFU) of WNV [49].
WNV antigen-specific antibody secreting cells were quantified via an enzyme-linked immunospot assay [28]. Briefly, mixed cellulose esters filter plates (Millipore) were coated with 20 μg/ml of purified WNV E protein. Bone marrow cells were harvested from infected mice 90 days post infection, and 200,000 cells were seeded per well with 8 wells per mouse. Antibody spots were developed with chromogen substrate (3-amino-9-ethyl-carbazole) and wells were imaged and enumerated with an ImmunoSpot plate reader from Cellular Technology Ltd.
To isolate naïve CD4+ T cells, splenocytes were isolated from mice and erythrocytes depleted with ACK lysis buffer (Invitrogen). Cells were then stained in buffer (PBS, 1% FBS, and 2 mM EDTA) with biotinylated antibodies against CD8, CD45R (B220), NK1.1, and MHC class II as well as fluorescently conjugated antibodies against CD4, CD44, and CD62L. Cells were washed with 10 mL of staining buffer and stained with anti-biotin magnetic beads (Miltenyi Biotec). Magnetically labeled cells were then depleted by passing through a LS magnetic column (Miltenyi Biotec) leaving roughly 90% pure CD4+ T cells. CD4+ CD44- CD62L+ were then isolated on an Aria-II (BD Biosciences) fluorescence-activated cell sorter (see S2 Fig). B cells were isolated using a B Cell Isolation Kit (Miltenyi Biotech). Briefly, cells were stained with a biotinylated-antibody cocktail followed by anti-biotin labeled magnetic beads. Cells were loaded onto a magnetic LS column and the flow-through was collected.
Naïve CD4+ T cells or B cells from uninfected adult or old mice were sorted as detailed above. For cell tracking, isolated cells from adult and old donors were labeled differentially with vital dyes (CellTracker Orange CMTMR, CellTrace CFSE, or CellTrace Violet) according to the manufacturer recommendations. Importantly, dyes were switched between old and adult cells in some experiments and dye-dependent effects on cell trafficking were not observed under our staining conditions (data not shown). Isolated cells were incubated at 37°C and 5% CO2 in pre-warmed CO2-Independent Medium (Invitrogen) with 2 μM of CellTrace CFSE, 5 μM CellTracker Orange, or 10 μM of CellTrace Violet (Invitrogen) for 15 minutes. Subsequently, cells were centrifuged (300 x g, 5 minutes) and the pellet was resuspended in pre-warmed medium for another 30 minutes. The incubation was stopped after two washes with cold PBS. Finally, cells were counted and adjusted to equal number for homing or microscopy imaging experiments. Cells were transferred adoptively via an intravenous route into recipient mice that had been infected subcutaneously with WNV-KUNV 48 hours earlier. For transfer into TCR β/δ-/- mice, 3 x 106 sorted naïve CD4+ T cells were injected intravenously into each recipient mouse that had been infected subcutaneously with WNV-NY 48 hours earlier.
For analysis of T cell homing to the LN and spleen, naïve CD4+ T cells or B cells were sorted from adult and old mice and labeled with different dyes. Cells were suspended in PBS at a ratio 1:1 (adult to old) and injected intravenously (2 x 106 from each donor). One hour post transfer, LN and spleens were harvested and processed into single cell suspensions. Cells were analyzed by flow cytometry on an LSRII (BD Biosciences).
We analyzed the trafficking behavior of lymphocytes in lymphoid organs using published methods for ex vivo (LN explants) and with minor modifications for in vivo imaging of cervical LNs [53]. Recipient mice (adult or old) infected with WNV-KUN 48 hours earlier were given 5 x 106 cells CellTrace CFSE or Violet labeled naïve T cells (from adult and old donors). For ex vivo analysis, LN were harvested 6 to 8 hours later, glued to plastic cover slips with VetBond (3M), placed under the flow of warm (35°C to 37°C) oxygenated medium (DMEM, without phenol red) [54]. Intravital imaging of T cell extravasation was performed as described [55] with the following modifications. Mice were anesthetized with avertin (250 mg/kg) and connected to a MiniVet Mouse Ventilator (Harvard Apparatus) to reduce movement artifacts and improve mouse viability. The skin of the neck was opened and the cervical DLN exposed on a skin flap and the inner surface was attached to a plastic cover slip using VetBond with the LN protruding through a hole in the center. Imaging was performed on A custom-build two-photon microscope equipped with an Olympus 1.0 NA 20x water-immersion objective, a Vision II Coherent Ti:Sapphire laser and running ImageWarp (A&B Software) for hardware and acquisition control [56]. Dyes were excited at 820 nm and detected using 480 nm and 570 nm dichroic filters. T cell rolling in vessels was imaged at 25 f/sec and 3D images collected ~ 2 per minute with image dimensions of ~250 x 225 x 75 μm (X,Y,Z; 2.5 μm Z steps). Blood vessels in the LN were labeled with QTracker 655 (10 μl in 50 μl of PBS given intravenously, Life Technologies).
Multidimensional data rendering and cell tracking were performed using Imaris software (version 7.3, Bitplane). For quantitative and statistical analysis of T cell motility we used either, T Cell Analyzer Software (version 1.7.0, Dr. J. Dempster, University of Strathclyde, UK) and an open-source software suite developed by Dr. J Textor (University of Ultrecht, The Netherlands). For quantitative and statistical analysis of T and B cell motility, we used the X and Y dimensions of the cell tracks to avoid bias arising from the intrinsically poorer Z resolution [57]. Displacement factors were defined as the distance between the endpoints of a track divided by the square root of its duration. The unpaired Mann-Whitney U test was used to compare track motility parameters. To control for variation between experiments, data were normalized prior to comparison by equalizing the medians across independent experiments. Hotelling's test for directionality was performed as described previously [58]. Briefly, cell tracks were broken down into steps (each step being a pair of consecutive cell positions), and the mean speed vector was computed per population of interest. Under the null hypothesis, the cells perform a random walk and the mean speed vector is zero in each dimension. Ellipses in Figs 4K, 5E, 7F and 7J show 95% confidence regions for the mean speed vector. If an ellipse does not contain the origin (0,0), then the null hypothesis is rejected for that population, and there is some directionality in the cell motion. For all cell populations studied in this paper, directionality was of a small magnitude (<1 μm/min at mean speeds of 5–10 μm/min).
Adult or old draining popliteal LN were harvested from mice at the indicated times post infection. To generate a single cell suspension, LNs were placed in staining buffer and gently crushed with the back of a syringe plunger and cells passed through a 70 μm cell strainer. The total number of live cells was quantified by trypan blue exclusion in a hemocytometer. Cells were stained in buffer (PBS, 1% FBS, and 2 mM EDTA) with fluorescently conjugated antibodies for 30 minutes on ice. Cells were washed in buffer and then processed on an LSRII flow cytometer (BD Biosciences).
Draining lymph nodes were harvested at 6 days post infection with WNV-KUN. DLNs were frozen in O.C.T. compound, and 9-μm-thick sections were cut with a cryostat. Sections were mounted on microscope slides and stained with anti–GL7-PE (1:50), anti–IgD-FITC (1:50), and DAPI (1 μg/ml) in PBS for 1 hour at room temperature. Slides were washed in PBS and a coverslip mounted on top of the stained tissue sections.
Splenocytes were isolated and erythrocytes were lysed with ACK buffer. Cells were stained with anti-CCR7 in staining buffer for 30 minutes at 37°C. Cells were washed and then incubated with anti-CD4, CD19, CD44, CD62L, LFA1, VLA4, PSGL1, and CXCR4 in staining buffer for 30 minutes on ice. Cells were processed on an LSRII flow cytometer. For filamentous actin staining, isolated CD4+ T cells were incubated with 100 ng/mL of CCL19 and CCL21 at 37°C in PBS + 0.1% BSA. Reactions were stopped by adding paraformaldehyde (3.6% final concentration), and cells were stained with anti-CD4, CD44, CD62L, and Alexa Fluor 647 phalloidin (Life Technologies). Cells were analyzed using an LSRII flow cytometer.
Mice were infected subcutaneously with 102 PFU of WNV in the footpad. Draining popliteal LNs were collected at 24 or 48 hours post infection and placed in 200 μl of PBS containing 0.5% (w/v) BSA. DLN were homogenized and cytokines were quantified using a Bio-Plex Pro 23-plex group I cytokine kit (Bio-Rad) and Bio-Plex 200 (Bio-Rad). CXCL13 and CCL21 were measured using ELISA kits from R&D Systems.
ICAM-1 Fc (R&D Systems) was adsorbed to a polystyrene Petri dish at 4 or 40 μg/ml in PBS. After a one-hour incubation at 37°C, plates were rinsed with PBS and non-specific binding sites were blocked with 1% human serum albumin (Sigma-Aldrich) for one hour at room temperature. Freshly isolated naïve CD4+ T cells (5.5 x 105) were added to Petri dishes in 1.5 ml of L15 media containing 0.5% FBS and 10mM HEPES and supplemented with 200 ng/ml of CCL19 and CCL21. Dishes were incubated for one hour at 37°C and then fixed with 1% paraformaldehyde. Unbound cells were removed after extensive rinsing with PBS supplemented with 0.2% human serum albumin, 1 mM CaCl2, and 1 mM MgCl2. Dishes were imaged at 2.5X magnification and the number of adherent cells was counted. For the transwell migration assay, a 96-well modified Boyden chamber with 5 μm pore size was purchased (Neuro Probe). The top membrane was coated with 0, 4, or 40 μg/ml of ICAM-1 Fc in PBS for one hour at room temperature. CCL19 or CCL21 (0.1, 1, or 10 ng/ml) in migration buffer (RPMI 1640, 0.5% FBS, and 100U/ml penicillin-streptomycin) was plated in the bottom chamber. Freshly isolated naïve CD4+ T cells (5 x 104) were added to the top chamber and incubated for 4 hours at 37°C. Subsequently, the membrane was removed and cells in bottom chamber were counted with a hemocytometer.
All data was analyzed using Prism software (Graph Pad 6, San Diego, CA). Flow cytometry data was analyzed using FlowJo software (Tree Star Inc.). Kaplan-Meier survival curves were analyzed by the log rank test. All other statistical analyses were performed with Mann-Whitney or unpaired t-tests as indicated in the Figure legends.
|
10.1371/journal.pntd.0002153 | Circulation of Japanese Encephalitis Virus in Pigs and Mosquito Vectors within Can Tho City, Vietnam | Japanese encephalitis virus (JEV) is a mosquito-borne, zoonotic flavivirus causing encephalitis in humans and reproductive disorder in pigs. JEV is present in large parts of Asia, where urbanization is high. Households within and outside Can Tho city, South Vietnam, were selected to monitor circulation of JEV. A nested RT-PCR was established to detect the presence of JEV in mosquitoes whereas sera from pigs belonging to households within the province were analyzed for the presence of antibodies to JEV. A total of 7885 mosquitoes were collected and divided into 352 pools whereof seven were JEV-positive, six of which were collected within the city. Fragments from four pools clustered with JEV genotype III and three with genotype I. Of the 43 pigs sampled inside the city 100% had JEV antibodies. Our study demonstrates exposure to JEV in pigs, and co-circulation of JEV genotype I and III in mosquitoes within an urban environment in South Vietnam. Thus, although JEV has mainly been considered a rural disease, the potential for transmission in urban areas cannot be ignored.
| Japanese encephalitis (JE) is a serious disease, especially in children, in large parts of Asia. It is transmitted by mosquitoes and mainly known as a disease in the rural, rice-producing areas in South and Southeast Asia. Here, the authors show that of 43 pigs sampled in Can Tho city, South Vietnam, all had antibodies to JE virus (JEV), and since some of these were born in the city, they must have been infected within the urban area. Mosquitoes were collected during night at households both in the city and in a rural area, using traps that attract mosquitoes with light. The mosquitoes were analyzed to see if they had JEV in them. Two different types of JEV were found in mosquito samples collected at urban households. One of these, genotype I, has been emerging in Northern Vietnam and was now shown to be circulating together with the older genotype III. These findings suggest a risk of JEV transmission from pigs to humans in urban areas, or in other words, that JE should not be regarded as a rural disease only.
| Japanese encephalitis (JE) is a zoonotic disease spread over large parts of Asia. It is one of the most important arboviral encephalitis in humans, with an estimated 10 million cases over the last 60 years, with 30% case fatality [1]. Pigs and wading birds are amplifying hosts of the causative Japanese encephalitis virus (JEV), and do not display clinical signs, except for pregnant sows that may abort or have stillborn piglets [2], [3]. Japanese encephalitis virus is a mosquito-borne flavivirus which is divided into five genotypes [4], and the virus has been isolated from more than 25 mosquito species, although not all are equally important in the epidemiology of JEV [5]. One of the most important vectors is Culex tritaeniorhynchus, a zoophilic mosquito that commonly breeds in irrigated rice fields, and therefore the disease is mainly considered rural [1], [3], [6].
Keiser et al. [6] calculated that 1.9 billion people live in rural areas with endemic or epidemic JE. Today, however, more than half of the world's population live in cities [7], [8] and the urbanization, especially in low-income countries, creates needs and possibilities for urban animal keeping to supply city inhabitants with food. Therefore, transmission of emerging zoonotic diseases in urban areas is increasingly important. Since some of the most populated cities in the world are in JE infested countries, the number of people at risk would increase dramatically if JEV is also transmitted in urban areas. It has been shown that two prerequisites for spreading JEV, the presence of competent vectors and the main amplifying host (the pig), are met in urban settings [9], [10]. In an urban area, the vector Cx. tritaeniorhynchus has been shown to increase in number by the presence of pigs, whereas the number of another vector, Culex quinquefasciatus, increases by the presence of humans [10]. However, the presence of JEV in urban areas has not been studied extensively previously, although previous studies in other cities in Asia have shown seropositivity in humans [11], [12].
In Vietnam, the land area used for rice production is increasing along with pig production, two factors likely to contribute to increased transmission of JEV [13]. Cases of encephalitis in humans are usually reported as acute encephalitis syndrome in Vietnam, and the incidence in south Vietnam has been 1.9 cases annually per 100 000 inhabitants between 1998 and 2007, with a mean case fatality of 6.4% [14]. One of the regions with the most JE cases, the Mekong Delta region [9], [15], has both extensive pig farming and rice production and disease is present all year [16], [17]. Although JEV is known to circulate in the rice-producing rural areas here, little is known about the circulation of JEV within urban areas.
The aim of the present study was to investigate the presence of JEV in pigs and vectors in a city in an endemic area, in order to contribute to the risk assessment of JE in humans. Pigs within and outside Can Tho city in the Mekong delta region, Vietnam, were examined serologically for JEV, and mosquitoes from the same locations were investigated for presence of JEV viral RNA. A high seroprevalence in pigs kept in the urban area is reported here, as well as the presence of JEV genotype I and III in mosquitoes collected in the city.
All animals in this study were treated according to the ethical standards of Can Tho University, Vietnam, and all animal handling was approved by the head of the Department for Veterinary Medicine, Can Tho University. Blood collections were performed by jugular venopuncture, which is the international recommended method [18] and also adhere to the Swedish guidelines for sampling blood from pigs in research [19]. A vacutainer system collecting maximally 10 ml was used. Household and pig owners were informed about the purpose and the methods of the study, and provided oral informed consent and answered questionnaires as a written consent.
Can Tho city is a central province in the Mekong delta region, comprised of eight districts, which are subdivided into wards. The most urbanized district is Ninh Kieu; it is comprised of the actual Can Tho city, which itself is subdivided in 13 wards, and has a total human population density of 7500 persons/km2 and a pig density of 94 pigs/km2 [20], and the more rural districts of the province have human population densities between 380 and 1400 persons/km2 and pig densities between 65 and 123 pigs/km2.
Ten urban wards in Can Tho city with different ratio of pigs/people were selected to represent different parts of the city. In these wards, households were included if sampling was allowed and if it was possible to affix traps for mosquito collections. In total 14 urban households that kept pigs and five households without pigs were included. Three households at pig farms in the rural Co Do district were also included as a comparison.
Mosquitoes were collected during two three-week periods in February–March (spring) and October–November 2009 (fall), using un-baited CDC mini light traps (Bioquip Products, California, USA) as described by Lindahl et al. [10]. Briefly, two traps were operated from dusk to dawn in the same household, close to human dwellings, and if the household had pigs, one of these traps was placed close to the pigs, immediately adjacent to the pig pen. Mosquitoes were identified according to Reuben et al. [21]. In catches containing more than 300 mosquitoes, 300 were identified to species, since this number always represented more than 10% of the catch and thus considered sufficient to estimate the existing species composition. The remaining non-identified specimens were counted and pooled unsorted.
Mosquitoes were pooled according to collection site with 1–60 mosquitoes per pool. To each pool, TRIzol Reagent (Invitrogen, Carlsbad, CA), corresponding to at least 10 times the total mosquito volume, was added. The mosquito pools were then stored at −20°C before analysis, except during transport to the National Veterinary Institute, Uppsala, Sweden, when samples were kept cold in a box with ice packs.
The mosquito pools were subsequently homogenized in a TissueLyzer (Qiagen GmbH, Hilden, Germany) for 2×1 min at 30 rpm after addition of one 5 mm steel ball to each tube. Extraction of the homogenates was performed according to the manufacturer's protocol for TRIzol Reagent with dilution of the resulting pellet in 20 µl of nuclease-free water and subsequent storage at −70°C.
To create a sensitive method for JEV detection in mosquito samples a nested RT-PCR protocol was established. When mosquito pools are analyzed for the presence of viral RNA using PCR, there is a risk for false negative results due to inhibition caused by the mosquitoes [22], [23]. To assess the inhibitory effect of mosquitoes, Swedish mosquitoes, assumed to be negative to flaviviruses as no known mosquito-borne flaviviruses are transmitted in Sweden [24], were used for spiking. Mosquitoes were caught by hand-net in Uppsala. 750 µl of TRIzol Reagent (Invitrogen, California, USA) were added to pools with 5 and 50 mosquitoes and homogenized as above. The Nakayama JEV strain, provided as a TRIzol Reagent suspension by the Swedish Institute for Control of Communicable Diseases (Solna, Sweden), was used. The suspension was diluted 1∶100 and 1∶1000 in TRIzol Reagent, and 1∶1000 in the homogenates of 5 and 50 mosquitoes, in order to mimic pools with low viral contents under field conditions.
RNA was extracted from all spiked virus suspensions, using TRIzol Reagent as described for the mosquito samples. All extractions were thereafter diluted in RNA safe buffer (RSB) [25] in dilution series to evaluate sample dilution as a method to avoid inhibition in a simple and cost-effective way. To further evaluate if inhibition occurs already in the extraction step, the extraction from the 1∶100 dilution of virus in TRIzol Reagent was diluted 1∶10 in the extraction from 50 homogenized mosquitoes without JEV, to reach the same concentration of extracted viral RNA as the other extractions.
The nested RT-PCR was performed using Agpath-id One step RT-PCR and Path-id PCR (Applied Biosystems, Foster City, CA, USA) according to the manufacturer's instructions with T3000 Thermocycler (Biometra, Goettingen, Germany) and Rotorgene3000 (Qiagen/Corbett Research, Sydney, Australia) respectively. The outer set of primers, emf1 and vd8 (Table 1), amplify an approximately 650 bp sequence from the non-structural protein 5 (NS5)-3′untranslated region (UTR) of all flaviviruses [26], whereas the inner set of primers and probe was specific for the NS5 region of JEV [27]. The probe (Table 1) was modified with a degeneration in the middle to improve sensitivity, since preliminary studies indicated a variation in nucleotide sequence (results not shown).
The first RT-PCR was set up with final primer concentrations of 160 nM, in a 25 µl reaction with 2 µl RNA template. The subsequent qPCR was performed with 1 µl of the product from the first PCR reaction in a 25 µl reaction using a primer concentration of 400 nM and the probe in a concentration of 150 nM. All Vietnamese mosquito samples were diluted with RSB and run in two dilutions (1∶10 and 1∶100), to avoid inhibition, according to the results using the spiked mosquitoes (Table 2).
The product from the first RT-PCR was reamplified using the outer forward primer (emf1) and the inner reverse primer. The product was visualized on a 1.5% agarose gel, excised and extracted using QIAquick Gel extraction kit (Qiagen Gmbh, Hilden, Germany) according to the manufacturer's protocol. Sequencing was performed using ABI PRISM Big Dye Terminator Cycle Sequencing v3.1 Ready Reaction kit (Perkin Elmer, Waltham, MA, USA) on an ABI PRISM 310 genetic analyzer (Applied Biosystems, Foster City, California, USA) according to the manufacturer's instructions. The resulting 133 bp sequences were aligned using BioEdit Sequence Alignment Editor [28] with reference sequences from different JEV genotypes derived from the NCBI GenBank database. Phylogenetic analysis was performed in MEGA version 5 [29] using the neighbor-joining method. Bootstrap probabilities were calculated using 10 000 replicates and evolutionary distances were computed using the p-distance method. To verify that the reference strains clustered the same way, a phylogenetic tree was also created the same way with the reference strains of their entire length.
The infection rate in the mosquitoes was calculated using two methods. The minimum infection rate (MIR) in the mosquitoes was calculated as the number of positive pools divided with the total number of mosquitoes, assuming that at least one mosquito was positive in the positive pool. The maximum likelihood estimate (MLE) for the mosquito infection rate was calculated using the software by Biggerstaff [30] (www.cdc.gov/ncidod/dvbid/westnile/software.htm). As both blood-fed and unfed mosquitoes were used in the analysis, the MLE here may be an overestimation of the infection rate, since some of the mosquitoes may have had virus only in a blood meal.
Blood sampling of pigs was performed during the sampling period in the rainy season, and restricted to sows and gilts over six months of age, hereafter referred to as female pigs. As many female pigs as the owner would allow, up to eleven pigs per household, were sampled. Blood samples were collected from the jugular vein using vacutainers, and kept cool with ice packs, until centrifuged the same day, after which the serum was stored at −20°C. Samples were transported frozen in a box with ice packs and inactivated for 60 min at 60°C at arrival, and stored at −20°C until analyses.
At the time for blood sampling, pig keepers were interviewed by a native Vietnamese using a written questionnaire. For every pig, data was collected on age, breed, parity, how many matings that were required to achieve the last pregnancy, the number of piglets born in total and alive in the last litter, if the sow had ever aborted or had stillborn piglets, if she ever had weak born or piglets displaying neurological symptoms at birth (e.g. shivering), how long she had been in the household and vaccination routines. The locations of the households where pigs were sampled are shown in Figure 1.
Serological analyses were performed using competitive IgG ELISA and IgM MAC ELISA as previously described [31], [32] and antigen, antibodies, conjugate and controls were provided from Australian Animal Health Laboratories (AAHL, Geelong, Victoria, Australia). All samples were tested twice on plates coated on different days. In the IgG ELISA a sample was considered positive if the inhibition was above 65% on both tests compared to the negative control. If a sample yielded different results it was considered inconclusive. In the IgM MAC ELISA, the ratio of sample optical density (OD) to the OD-value of the lowest concentration of the positive control was calculated. A sample was considered positive if it had a ratio greater than one.
Swedish pig sera were used to verify the specificity of the ELISAs. The sera originated from routine surveillance of infectious diseases in Swedish control programs. To evaluate the effect of heat inactivation on the specificity of the assays, the Swedish samples were aliquoted and one part was inactivated for 56°C for 60 min, and changes in OD compared to the unheated part, run on the same plate, were monitored.
The results of the dilution series with Swedish mosquitoes spiked with JEV are shown in Table 2. It was not possible to detect JEV in spiked Swedish mosquito pools unless they were diluted. In a pool with five mosquitoes it was necessary to dilute the extracted RNA 1∶10, and in pools with 50 mosquitoes 1∶100. The same results were obtained if extracted JEV RNA was added after the RNA extraction of the mosquitoes (results not shown).
A total of 7885 mosquitoes, divided into 352 pools, were screened for JEV. Six mosquito pools from four of the urban households were positive for JEV by the nested PCR (Table 3). One positive mosquito pool was found in one of the rural households.
Minimum infection rate was 1.0 per 1000 mosquitoes (excluding identified males, 95% CI 0.25–1.71). The MLE for infection rate was 1.0 per 1000 mosquitoes. The MLE in Cx. tritaeniorhynchus was 1.6 per 1000 mosquitoes and in Cx. quinquefasciatus 1.3 per 1000 mosquitoes. When MLE was calculated separately for mosquitoes collected within the city, it was lower, except for Cx. quinquefasciatus (Table 4).
All positive mosquito pools contained both blood-filled mosquitoes and mosquitoes without visible blood content, and all pools had been collected close to the pigs. At two households, I and S (Table 3), the positive mosquito pools contained non-identified specimens. The 300 identified mosquitoes from the corresponding collection at household I consisted of 46% Cx. tritaeniorhynchus and 36% Cx. gelidus. The 300 identified mosquitoes at household S consisted of 78% Cx. tritaeniorhynchus.
The phylogenetic analysis showed that three sequences clustered with genotype I strains, and four with genotype III strains (Figure 2). The pairwise sequence comparison between the fragments showed between 90 and 99% similarity. At household S, sequence fragments from the two positive samples clustered with different genotypes in the phylogenetic tree. The location of samples with the different genotypes within Ninh Kieu district is shown in Figure 1.
All 43 female pigs from the urban households and 30 out of the 31 from rural households were positive in the IgG ELISA giving an overall seroprevalence of 99% (95% CI 96%–100%). The test result from one sample was considered inconclusive. All samples tested negative or inconclusive in the IgM ELISA. According to the pig keepers, 24 of the 43 pigs in the urban area and 28 of the 31 pigs in the rural area originated from the household.
In the sows that had at least had one litter (n = 51), repeated estrus was reported for 29%, 8% had aborted, 29% had stillborn or mummified fetuses and 24% weak born piglets, while 4% delivered piglets that shivered at birth. Out of the 51 sows, 28 had never shown any of these symptoms.
Here we demonstrate the concurrent circulation of two JEV genotypes in mosquitoes within the urban area of Can Tho city and extensive seropositivity in pigs born in the city. Worldwide, demographic changes, human behavior, and increased globalization are suggested drivers for emergence of arboviruses such as JEV [33]. However, increased urbanization and the establishment of anthrophilic vectors have been suggested to be the most important catalysts [34]. Even though human JE cases have been shown to occur in cities without extensive pig keeping [11], the increases in the number of vectors close to pigs in urban areas [10], and the fact that all positive mosquito pools had been collected close to pigs in the present study, together indicate increased risks associated with pigs in urban animal farming.
In the present study JEV-RNA fragments from positive mosquito pools clustered with isolates of both genotypes I and III. Genotyping is normally based on sequencing of the E or the prM genes, which both have extensive variation, whereas NS5, the virus polymerase, is highly conserved between viral strains [35], which is confirmed by the limited variation between the sequences found here. Two of the positive samples from urban households were highly similar to the one at the rural household X, 20 km away. Although JEV likely is active throughout the area, explanations for this similarity could be a trade of viremic pigs, or that infected vectors or birds move between rural and urban areas.
Genotype III was previously the most common genotype in Vietnam, but, through sequencing of old and new isolates from northern Vietnam, it has been shown that genotype III gradually has been superseded by genotype I [36], a shift that has been observed in other Asian countries as well [37], [38]. The results of the present study indicate that there is a co-circulation of the two genotypes in southern Vietnam.
In an earlier study by Thu et al. [39] in the rural areas of Can Tho province, one positive pool of JEV was found from 22 048 mosquitoes, yielding a MIR of 0.05/1000 mosquitoes, similar to a MIR of 0.046/1000 Cx. tritaeniorhynchus in suburban Bangkok found by Gingrich et al. [40]. In the present study the MLE was 1.0/1000 mosquitoes and 1.6/1000 Cx. tritaeniorhynchus. Although it is possible that the higher MLE here is explained by higher infection rates in the mosquitoes in Can Tho city, it could also partially be due to higher sensitivity of the nested RT-PCR used. Also, all positive mosquito pools contained blood-filled mosquitoes as well as unfed. Thus, since it is possible that not all mosquitoes with JEV were actually infected, the actual infection rates may be lower than those calculated here. However, since the mosquitoes were all competent vectors for JEV, even presence of JEV in the blood meal indicates a risk of transmission.
Three of the PCR positive mosquito pools contained Cx. tritaeniorhynchus, a known competent vector for JEV [3]. The number of Cx. tritaeniorhynchus has been found to be significantly associated with the number of pigs in a household [10] and in the present study the positive pools were also consistently collected close to the pigs. One JEV positive mosquito pool contained Cx. quinquefasciatus, another competent vector. This species is anthropophilic and feed to a higher extent on humans than on pigs [41]. The positive mosquito pool was collected close to pigs, indicating that Cx. quinquefasciatus may have an important role as a bridge vector within Can Tho city.
The procedures used for mosquito handling and the established nested RT-PCR may provide a robust, economic and sensitive method for screening mosquito pools for JEV, which makes it suitable in tropical and low-income countries where the disease burden from mosquito-borne infections is high. TRIzol Reagent was used since it inactivates virus and helps preserving the RNA of RNA viruses for future analyses, even in samples stored at room temperature [42] and extraction with TRIzol Reagent does not require expensive equipment. We demonstrate that inhibition of PCR may cause problems for detection of arbovirus when screening mosquito samples, and that this should be taken into account. We also conclude that it is possible to avoid inhibition caused by mosquitoes simply by diluting samples.
The IgG sero-prevalence in the present study approached 100%. This was higher than the results by Lindahl et al. [43] when a commercial JEV ELISA kit was used for detection of JEV in sows in the rural area surrounding Can Tho city. However, ten years have passed between the samplings, which may explain the different results, and in addition, two different ELISA methods were used. The competitive IgG ELISA used in the present study has been shown to be cross-reactive with other flaviviruses in the JEV serological group, such as Murray Valley encephalitis virus and Kunjin virus [31], [32], but none of these flaviviruses have been demonstrated in southern Vietnam. Apart from JEV, dengue virus is the only vector-borne flavivirus in the region that infects mammals, and serological cross-reactions with JEV have been demonstrated [44]. However, cross-reactions with dengue virus occur to a lesser extent than with viruses in the JEV serological group [45], [46] and we therefore consider it unlikely that the positive results in our study would be due to cross reactions. Another flavivirus which could cause cross-reactions with JEV is Tembusu virus, which is present in Southeast Asia [47], although there are no reports of the virus in the Mekong delta region in Vietnam. More than half of the female pigs in the urban households were born at the farm where they were sampled and must thus have been infected at their present location. The negative and inconclusive results using the IgM ELISA could indicate that the JEV infections are not recent, or that infection with JEV occurred while the pigs were still partly protected by maternal IgG antibodies, thus inducing less IgM production [48]. JEV seropositivity has mainly been studied in humans in urban areas [35], [12] but since humans tend be mobile, pigs born in the city may be better indicators for JEV transmission within an urban environment. Notably, half of the female pigs had experienced reproductive symptoms that could be related to JEV infection. However, there are many other pathogens circulating in the Mekong Delta [49], [50] that can cause similar reproductive symptoms, although a previous study in Can Tho province could find an association between seropositivity to JEV and reduced reproductive performance in female pigs less than 1.5 years [44].
Whether humans are infected in the urban area or not, is not known. In the entire Can Tho city province the reported incidence of acute encephalitis has been on average 2.4 cases per 100 000 inhabitants during 2009–2012, with the majority of cases being in children under 6 months of age [51]. As in other endemic areas the number of clinical cases in adults is relatively low, due to the acquired natural immunity in the adult population [52], but the risk for clinical disease may be much higher for non-immune visitors from non-endemic areas. Vaccination against JEV is increasing in Vietnam although not all children are covered yet [14]. With increasing indications of risks for urban transmission of JEV there may be cause to revise vaccination policies.
In conclusion, the present study demonstrates the presence of JEV within an urban area by finding both serological evidence of widespread infections in pigs and mosquitoes PCR-positive for the virus.
|
10.1371/journal.pgen.1005942 | The MRX Complex Ensures NHEJ Fidelity through Multiple Pathways Including Xrs2-FHA–Dependent Tel1 Activation | Because DNA double-strand breaks (DSBs) are one of the most cytotoxic DNA lesions and often cause genomic instability, precise repair of DSBs is vital for the maintenance of genomic stability. Xrs2/Nbs1 is a multi-functional regulatory subunit of the Mre11-Rad50-Xrs2/Nbs1 (MRX/N) complex, and its function is critical for the primary step of DSB repair, whether by homologous recombination (HR) or non-homologous end joining. In human NBS1, mutations result truncation of the N-terminus region, which contains a forkhead-associated (FHA) domain, cause Nijmegen breakage syndrome. Here we show that the Xrs2 FHA domain of budding yeast is required both to suppress the imprecise repair of DSBs and to promote the robust activation of Tel1 in the DNA damage response pathway. The role of the Xrs2 FHA domain in Tel1 activation was independent of the Tel1-binding activity of the Xrs2 C terminus, which mediates Tel1 recruitment to DSB ends. Both the Xrs2 FHA domain and Tel1 were required for the timely removal of the Ku complex from DSB ends, which correlates with a reduced frequency of imprecise end-joining. Thus, the Xrs2 FHA domain and Tel1 kinase work in a coordinated manner to maintain DSB repair fidelity.
| Genomic DNA provides the essential blueprint for life, and therefore living organisms have several mechanisms for maintaining the stability of their own genomes. DNA double-strand breaks (DSBs) are one of the most severe forms of DNA damage, which, without precise repair, can provoke a loss of genetic information, leading to tumor formation. DSBs are repaired by two distinct pathways, homologous recombination (HR) and non-homologous end joining (NHEJ), which can be precise or imprecise. In addition, the DNA damage response (DDR) is essential in the cell to integrate multiple events that need to occur after damage: activation of DNA repair enzymes, selection of repair pathway and control of cell cycle progression, transcription, and so on. Here we show that different domains of Xrs2, a central DSB repair protein in budding yeast whose human ortholog, Nbs1, is linked to a human hereditary disorder with a high risk of cancer, is required not only for repair pathway choice but also for full activation of DDR. This result indicates that DSB repair and the DDR are coordinated at multiple levels to ensure precise repair and thus to maintain genomic integrity.
| The DNA double-strand break (DSB) is one of the most severe types of DNA damage and is most often repaired by homologous recombination (HR) or canonical non-homologous end joining (C-NHEJ) which is known as precise NHEJ. There are, however, several other minor pathways for DSB repair, some of which generate serious rearrangements of DNA structure. It is thought that an incorrect choice among these repair pathways promotes genomic instability, which compromises biological activity and can with time, promote tumorigenesis in higher eukaryotes [1]. The Mre11-Rad50-Xrs2/Nbs1 (MRX/N) complex has many roles in the initial steps of DSB repair, whether by C-NHEJ or HR, and also in the recovery from stalled replication forks, in telomere maintenance, in meiotic recombination and in the Tel1/ATM-related DNA damage response (DDR) signaling [2–6]. Thus, MRX/N acts as an integrating hub of DDR pathways.
In budding yeast, Saccharomyces cerevisiae, repair by C-NHEJ requires several multi-subunit complexes, namely MRX, Yku70-Yku80 (Ku) and Dnl4-Lif1-Nej1 (DNA ligase IV). First, Ku binds to free double–stranded DNA (dsDNA) ends, without needing a specific DNA structure, and then is able to translocate to an internal region of the DNA molecule, including the single-stranded DNA (ssDNA) region [7, 8]. Then, C-NHEJ is completed by DNA ligase IV to rejoin the broken ends. To function effectively, these complexes rely on physical interactions between components, for example, Ku80 binds Dnl4 and Mre11 and Xrs2 binds Lif1 [9, 10]. Alternative non-homologous end joining (A-NHEJ), also known as microhomology-mediated end joining (MMEJ), is an auxiliary pathway for the repair of DSBs that occurs after end processing. MMEJ requires the MRX complex, Sae2, Tel1 and Rad1/Rad10/Slx4, but not Ku nor DNA ligase IV complexes [11, 12]. Although the process of MMEJ is quite similar to the single-strand annealing (SSA) reaction, MMEJ in yeast is genetically distinguishable from SSA by its requirement for Rad52, a protein that plays a key role in HR [11]. A further, minor pathway of NHEJ, often considered as a variant of C-NHEJ, is that of Ku-dependent imprecise end joining [13, 14]. This pathway can rejoin broken ends with or without microhomology after limited (<50 bases) resection [12]. It is unclear what molecular complexes or events distinguish C-NHEJ from the Ku-dependent imprecise–end joining reaction. Thus, MMEJ and Ku-dependent imprecise end joining are classified as imprecise NHEJ.
Mre11 has endo- and exonuclease activities, and Rad50 is a structural maintenance of chromosome (SMC)-like protein [15–18]. The Mre11-Rad50 sub-complex holds the two ends of a DSB together and facilitates their subsequent processing [19]. Mre11 and Rad50 are conserved from prokaryotes to mammals where they bind a third component called Nbs1 [20]. Xrs2 is the yeast ortholog of the Nbs1 subunit. Xrs2/Nbs1 is thus a eukaryote-specific multi-functional regulatory subunit of the MRX/N complex. The protein consists of a fork-head associated (FHA) domain, a pair of BRCA1 C terminus (BRCT) or BRCT-like domains, an Mre11-binding domain and a Tel1-binding domain [21–24]. The FHA domain is conserved in most of the orthologs in the N-terminal domain [21, 22, 25] and the motif generally has a well-known phospho-protein recognition function important for the DNA damage–related signaling pathway [26–28]. FHA domains are thus implicated in the recruitment of appropriate targets to sites of DNA damage through protein-protein interactions in the phosphorylation-transducing pathways, such as Rad53-Rad9 in budding yeast, Nbs1-Ctp1 in fission yeast and Nbs1-MDC1, or RNF8-MDC1, in humans [27–29].
In addition to the N-terminal FHA domain, the C-terminal region of Xrs2/Nbs1 harbors a Tel1/ATM-binding domain that is essential for Tel1/ATM recruitment to sites of DNA damage and to telomeres, as well as for its activity [22, 23, 30, 31]. In addition to Tel1/ATM, Mec1/ATR also has a role in the DDR reaction. Whereas Tel1/ATM recruitment depends on Xrs2/Nbs1, Mec1/ATR requires a replication protein A (RPA)-coated ssDNA stretch that results from Ddc2/ATRIP activity at the site of DNA damage [32]. Recent work suggests a role for RPA in MRX recruitment as well (Seeber A. et al., personal communication). In humans, truncation mutations of N-terminus region including FHA domain of Nbs1 have been identified as a causative factor for Nijmegen breakage syndrome (NBS), which confers a high risk of cancer and immunodeficiency and was, originally identified as an ataxia telangiectasia–like disorder [27, 33–36]. Consequently, cells in which the N-terminal region of NBS1 is truncated have abnormal cell cycle checkpoints, including the S-phase checkpoint, which is manifested as radio-resistant DNA synthesis [36, 37].
Here we report that dysfunction of the Xrs2 FHA domain, as with the loss of Tel1 kinase activity, leads to the accumulation of the Ku complex at DSB ends, which leads to an abnormal increase in imprecise end joining. Moreover, the Xrs2 FHA domain is required for robust activation of Tel1/ATM kinase at DSB ends both during mitosis and meiosis. Our findings reveal a genetic relationship between the Xrs2 FHA domain and Tel1 kinase activity in the maintenance of DSB repair fidelity and provide insights relevant to the human disease NBS.
The FHA domain of Xrs2 is involved in NHEJ [9, 10, 22]. To learn more about the function of the FHA domain of Xrs2 in various NHEJ pathways, especially in imprecise end joining, we analyzed the effect of xrs2 mutations on the repair of two HO endonuclease–induced non-complementary DSBs at the MAT locus in budding yeast [11]. In this system, two HO cleavage sites with opposite orientations were inserted on either side of URA3 to allow repair by both Ku-dependent and Ku-independent pathways [14] (Fig 1A). The survival rate of Ura− prototrophs corresponds to the frequency of repair of non-complementary DSB ends by imprecise end joining, which includes both MMEJ and Ku-dependent imprecise NHEJ (Fig 1A)[11, 12]. In contrast, the survival rate of Ura+ prototrophs, which are caused by re-ligation of two HO-induced complementary ends, corresponds to the frequency of precise NHEJ [11](Fig 1A). In an xrs2Δ mutant, as in mre11Δ and rad50Δ mutants [11], both imprecise end joining and precise NHEJ are compromised (Fig 1C).
Exonuclease defective mutations of MRE11, which encodes a protein that is in the same complex with Xrs2, show a completely different phenotype from that of mre11Δ in the various NHEJ pathways [14]. Thus, this assay is useful for revealing different functions within a given polypeptide. To determine the different roles played by Xrs2, we used two mutant alleles that compromise the FHA function of Xrs2: two point mutations in the FHA domain, xrs2-SH (S47A, H50A) and a truncation mutant that lacks 313 amino acids of the N-terminal domain, xrs2-314M, which eliminates both the FHA and the BRCT-like motifs (Fig 1B) [22]. These xrs2 FHA mutants do not show any defects in MRX complex formation or, gamma ionizing radiation (γIR) sensitivity, but they do show a defect in NHEJ [10, 22]. We analyzed these mutants for their effects on various types of DSB repair and observed a significant 2.3- and 2.2-fold increase in the frequency of imprecise end joining in xrs2-SH and xrs2-314M mutants, respectively, relative to wild-type cells (Fig 1C). In contrast, the frequency of precise NHEJ, which was determined by the ratio of Ura+ prototrophs, was lower in these FHA domain mutants as compared with wild-type cells (Fig 1C), which is due to impaired interaction with Lif1 [10]. Total loss of Xrs2, in contrast, compromised both types of repair. This result indicated that the FHA domain of Xrs2 helps promote precise NHEJ and this or the FHA domain itself leads to the suppression of imprecise end-joining.
The tel1Δ and sae2Δ mutants were reported to show an increase in both precise NHEJ and imprecise end joining [12, 14], probably because of reduced resection at the break. Consistently, a truncation of xrs2 (xrs2-664), which eliminates the Tel1-binding domain [22] (Fig 1B), results in an increase in the re-ligation of linearized plasmids in vivo [10]. Using our assay, we confirmed that the tel1Δ mutant had imprecise–end joining activity, which was 2.2-fold higher than that of wild-type cells (Fig 1D). Because the tel1Δ mutant was quite similar to these xrs2 FHA mutants with respect to the frequency of imprecise end joining, we then examined the relationship between the two deficiencies by scoring imprecise end-joining in the double mutants. The frequencies of imprecise end-joining of tel1Δ xrs2-314M and tel1Δ xrs2-SH mutants were indistinguishable from that of the tel1Δ single mutant (Fig 1D), indicating that the increase in imprecise end joining in the xrs2 FHA and tel1Δ mutants reflects the loss of a single pathway. Interestingly, however, the drop in precise NHEJ frequency in the xrs2 FHA mutants was dominant over the increase observed in the tel1Δ mutant (Ura+, Fig 1D).
We observed a similar result upon loss of Sae2, which indirectly promotes Ku disassembly from DSBs through its activity in DSB end resection [38]. The imprecise–end joining frequency increased 10-fold over wild type, and the sae2Δ xrs2 FHA double mutants showed a higher level of imprecise–end joining activity than did the xrs2 FHA single mutants (Fig 1C and 1E). The effect on precise NHEJ (Ura+) frequencies showed a dominance similar to that of the xrs2 mutants relative to tel1Δ (Fig 1D). However, the FHA domain of Xrs2 and Sae2 were not entirely epistatic to one another with respect to their effects on imprecise end joining (Fig 1E).
We further tested the effects of the yku70Δ mutation in these assays. yku70Δ was dominant over xrs2 FHA mutations and suppressed the abnormal increase in imprecise end joining (Ura−) observed in the xrs2 FHA mutants (Fig 1F). In addition, we confirmed that most of the residual imprecise end joining in the yku70Δ xrs2 FHA double mutants was caused by the Ku-independent MMEJ pathway (S1 Fig). yku mutations are also dominant over tel1Δ mutations in imprecise end joining [14]. These results indicate that the increase in imprecise end joining in the xrs2 FHA mutant is probably achieved by the Ku-dependent imprecise–end joining pathway, which also regulates events in the tel1Δ mutant.
In our previous study, we showed that the Xrs2 FHA domain functions in NHEJ through an interaction with Lif1, a component of DNA ligase IV in budding yeast [9, 10]. This was also confirmed in this assay as a reduced frequency of Ura+ prototrophs, as described above. Interestingly, the xrs2 FHA mutation also suppressed precise NHEJ in the tel1Δ or sae2Δ background, which would most likely have been caused by an interaction defect with Lif1. To confirm this, we identified two Xrs2-interacting domains in Lif1 and constructed lif1 mutations in each domain, named lif1-SST and -T113A, both of which lose the ability to interact with Xrs2 and compromise C-NHEJ [10]. We checked whether these lif1 mutations allow an increase in imprecise end joining, as observed for the xrs2 FHA-deficient mutant (Fig 1G). However, both lif1-SST and lif1-T113A mutants showed a slight decrease in the frequency of imprecise end joining, as compared with wild type, along with the expected drop in precise NHEJ (Fig 1G). In addition, both lif1-SST xrs2–SH and lif1-T113A xrs2-SH double mutants showed indistinguishable increases in imprecise end joining as compared with the xrs2-SH single mutant (Fig 1G). This confirms that the increase in imprecise end joining detected in the FHA-deficient mutants reflects its interaction with Tel1, rather than with Lif1. In contrast, lif1 mutations were dominant over the xrs2-SH mutation in the suppression of precise NHEJ. This indicates that precise NHEJ activity in the FHA-deficient mutants depends on the Lif1 interaction.
We next characterized DSB-repair events observed after the induction of non-complementary DSBs. As shown previously [12], repair products can be classified into five categories (Fig 2A). To distinguish the categories, we determined the junctions of repaired products amplified from Ura− prototrophs after induction of the non-complementary DSBs (Fig 1A). Products in category-A are produced by a Ku- and a DNA ligase IV–independent imprecise pathway [13, 39]. As expected, 95% of the products recovered in the yku70 mutant belong to category A (Fig 2B and S1 and S2 Tables). This pathway is also called MMEJ in yeast [14] or A-NHEJ in mammalian cells [13]. In contrast, products in categories B–E are Ku dependent- and DNA ligase IV dependent [12] (S1 Fig and S2 Table). In addition, DSB repair products in categories B and C are associated with DSB end-resection, whereas those of categories B and E are mediated by microhomology-dependent annealing to repair non-complementary DSBs (Fig 2A and 2B, right).
We analyzed these repair events in xrs2-SH, tel1 kinase-defective (tel1-KN), xrs2-664, xrs2Δ, and yku70Δ mutants for their effects in these end-joining events. First, we detected an increase in total frequencies for imprecise end joining in tel1-KN and xrs2-664 (which truncates the Tel1-binding domain of Xrs2) mutants as well as in the xrs2-SH mutant (Fig 2B and S3 Table). It is remarkable that the increase in imprecise end joining in these mutants is associated with an increase in Ku- and DNA ligase IV–dependent repair, leading to products of categories B–E (Fig 2B, left). Only the xrs2-SH mutant maintained a high level of repair in category A, a Ku-independent pathway, as did wild-type cells (33.3 and 31.8%, respectively) (Fig 2B and S2 Table). In contrast, tel1-KN and xrs2-664 mutants showed a substantial reduction in this category to 0.8% and 2.1%, respectively (S2 Table). This is consistent with a previous report showing that Tel1 function is essential for the Ku-independent MMEJ pathway [11]. In addition, category C products, which would be produced by simple re-ligation between processed ends, were not observed in the xrs2-SH mutant (<0.58%). This phenomenon is distinct from that of the xrs2-664 and xrs2Δ mutants. Based on experiments with double mutants, the elevated level of imprecise end joining in xrs2 FHA mutants was sensitive to the yku70 mutation (Fig 1F, gray bar). Sequence analysis revealed that the yku70Δ xrs2 FHA double mutant showed almost the same distribution for each category with the yku70Δ single mutant (S1 Fig). This also indicates that the xrs2 FHA and tel1 mutants promote an unusual Ku-dependent–imprecise end joining pathway (categories B–E).
To assess which mechanisms were at work in the different mutants, we next analyzed the assembly of Xrs2, yKu70, Tel1 and Sae2 proteins on the HO-induced non-complementary DSB ends at the MAT locus on chromosome III by chromatin immunoprecipitation (ChIP). Quantitative real-time PCR (qPCR) was carried out with a primer pair that is 100 base pairs from the DSB site (Fig 3A). First, we examined assembly of mutant Xrs2 proteins at the DSB. We tagged wild-type Xrs2 and Xrs2–SH and Xrs2–314M mutant proteins with 13Myc epitopes at the C terminus, and confirmed normal recruitment of each mutant form following DSB induction (Fig 3B). Next, we examined assembly of the Ku complex at the DSB by using FLAG-tagged yKu70 in wild-type, tel1Δ, xrs2-SH, or xrs2-314M cells. Although FLAG-tagged yKu70 showed a moderate defect both in imprecise end joining and C-NHEJ, these levels of end-joining activity were substantially higher than in the yku70Δ mutant (S2A Fig). We observed a significant increase in yKu70 binding to the DSB ends in the tel1Δ, xrs2-SH and xrs2-314M mutants, compared with wild-type cells at 120 min after DSB-induction (Figs 3C and S2B). In addition, in the xrs2-SH and tel1Δ, Ku-binding signals were detected from 30 min after DSB induction as a same level with that in wild type, and then, they were gradually increased than wild type in accordance of time elapsed (Fig 3C). This indicates persistent binding of Ku at DSBs in the mutants. To determine whether a defect in the DSB resection indirectly affects the Ku removal from the DSB ends, we examined DSB resection in the xrs2 FHA and tel1Δ mutants. We measured DSB end resection at 1600 base pairs from the DSB site (Fig 3A) by quantitative amplification of single-stranded DNA (QAOS) [40, 41] (Fig 3D). First, we confirmed that the sae2Δ mutant showed significantly reduced ssDNA production at the DSB ends, as reported [42]. Then we showed that tel1Δ had DSB resection with the same kinetics as did wild type. In contrast, the xrs2-SH mutant showed a slight decrease in ssDNA end production in the initial phase (15–120 min) but showed almost the same amount of resected DSB ends with wild type at 150 min, which is when accumulated yKu70 was observed at the ends (Fig 3C). This argues that Xrs2 works together with Tel1 to evict Ku from DSB ends during DSB resection.
As Xrs2 is involved in Tel1 recruitment to the DSB site though its C-terminal region [22] [23], we examined Tel1 binding at the DSB using FLAG-tagged Tel1. Interestingly, we detected Tel1 assembly at the DSB in xrs2-SH at wild-type levels and a significant increase in Tel1 binding in the xrs2-314M mutant cells (Fig 3E). This was distinct from the effect of the xrs2-664 mutant, which lost Tel1 binding (Fig 3E), as previously reported [23]. This result indicates that the xrs2 FHA mutation does not affect Tel1 recruitment to DSBs.
As it is known that Tel1/Mec1 phosphorylation is required for Sae2 function [43–45], we examined Sae2 recruitment using C terminally HA-tagged Sae2, but there was no difference between the wild-type recruitment and that in the xrs2 FHA mutants (Fig 3F). This argues that FHA function of Xrs2 is not necessary for Sae2 binding at DSB sites. We conclude that Xrs2 FHA function does not affect the initial recruitment of Tel1 kinase to DSBs yet is responsible for removing Ku from the DSB ends. This was also observed in tel1 mutant cells. We proposed that the Xrs2-FHA domain specifically supports Ku-removal by maintaining high Tel1 activity at DSB ends.
We demonstrated that the FHA domain of Xrs2 is related to the function of Tel1 in imprecise NHEJ suppression but not to its recruitment to DSB sites. To understand the function of the FHA domain of Xrs2 in a Tel1-dependent DDR pathway, we analyzed γIR sensitivity in the xrs2 mutant cells. As reported previously [4], a mec1Δ mutant showed severe sensitivity to γIR because of defects in the DDR. In contrast, a Tel1-defective mutant did not show similar sensitivity (Fig 4A). This is because the Mec1-dependent pathway is the major pathway for survival in budding yeast after irradiation rather than the Tel1-dependent pathway [4]. rad50S and sae2Δ mutations, which have a defect in the initiation of DSB resection, can change the biased dependency on Mec1 by activating the Tel1 pathway through suppression of ssDNA production at broken ends [4, 42]. Thus, the rad50S mutation suppresses the radiation sensitivity of mec1 mutant cells [4]. Consistently, the mec1 rad50S double mutant was 150-fold more resistant than mec1Δ alone to 500 Gy of γIR (Fig 4B).
We then examined whether the Xrs2 FHA domain plays a role in the DDR relative to γIR sensitivity. xrs2-SH is not sensitive to γIR [22], which we confirmed here; we also confirmed that rad50S mutant cells were only slightly sensitive to high doses of γIR (Fig 4A). Like rad50S mec1Δ, the mec1Δ xrs2-SH double mutant was 5.5-fold more resistant than mec1Δ alone to 500 Gy of γIR, indicating that the FHA mutation partially suppresses the DDR defect in mec1Δ. This suppression by xrs2-SH was less than that achieved with rad50S, however (Fig 2A and 2B). This suggests that xrs2 FHA mutations may be able to change the biased dependency on Mec1 although activation of Tel1 might be incomplete.
MEC1 is an essential gene in budding yeast, but mec1Δ cells grow in the presence of the sml1 mutation [46]. Interestingly, rad50S also suppresses mec1Δ lethality presumably by activating the Tel1 pathway, even in the absence of sml1Δ [4] (Fig 4C, upper). Similarly, we found that the xrs2-SH mec1Δ double mutant was able to grow in the absence of the sml1Δ mutation, forming smaller but viable colonies (Fig 4C, lower). These results indicated that loss of the Xrs2 FHA domain suppresses γIR sensitivity in mec1Δ sml1Δ mutants and suppresses the lethality of the mec1Δ mutation. It remains to be tested whether this was through Tel1 recruitment or activation or through another pathway.
Tel1 activity in vivo depends largely on Xrs2, because of the physical interaction of Tel1 with a domain located in the C terminus of Xrs2 [23, 47]. Consistently, xrs2-664 [22] did not show sensitivity to γIR even in a tel1Δ background (Fig 4D). Eliminating the Mec1 pathway in xrs2-664 mutants, in contrast to xrs2-664 tel1Δ mutant cells, showed a severe loss of viability after irradiation, much like the mec1Δ tel1Δ double mutant (Fig 4D). In contrast to xrs2-664, the xrs2-SH mutation partially suppressed the mec1Δ defect, as described above (Fig 4B). Like the mec1Δ rad50S xrs2-664 triple mutant, the mec1Δ rad50S xrs2-SH triple mutant was more sensitive to γIR than the mec1Δ rad50S double mutant (Fig 4B and 4D). Moreover, the rad50S mutation was not able to suppress the hypersensitivity of the mec1Δ xrs2-SH double mutant, especially at high doses of γIR (Fig 4B). This indicates that the rad50S-dependent suppression of mec1Δ hypersensitivity to γIR requires the FHA function of Xrs2, again acting most likely through Tel1 activation but not through the suppression of DSB processing in the rad50S mutation. We conclude that the Xrs2 FHA domain activates Tel1, although its function is distinct from that of the Tel1-binding domain of Xrs2.
The Mec1-dependent pathway dominates the Tel1-dependent pathway in wild-type cells, but the Tel1 pathway will be equally activated in the rad50S mutant [4]. We hypothesized that this activation requires the Xrs2 FHA domain (Fig 4E) [40]. To test this hypothesis, we next examined DDR activity in the xrs2 FHA mutant cells by monitoring Rad53 activation. Rad53 is a downstream mediator kinase of the yeast DDR pathway and is a phospho-target of Mec1 and Tel1 [48, 49]. We treated yeast haploid cells in vegetative growth with the DSB-inducing compound phleomycin and analyzed the Rad53 phosphorylation status by Western blotting at the indicated time points (Fig 5A). We detected step-wise accumulation of multiple slower-migrating signals, which correspond to different phosphorylated forms of Rad53, with robust phosphorylation achieved by 120 min in wild-type, xrs2-SH and rad50S cells (Fig 5A). We also detected an initial phosphorylation of Rad53 at 15 minutes after phleomycin addition and secondary phosphorylations events after 60 min in mec1Δ rad50S mutant cells. In the mec1Δ rad50S xrs2-SH triple mutant, the initial phosphorylation was also observed with the same timing as in the mec1Δ rad50S double mutant, but accumulation of secondary phosphorylation was delayed (Fig 5A). This indicates that the FHA domain of Xrs2 is not required for the initial step of Rad53 phosphorylation, although it is required for its robust activation in DDR signaling.
To clarify the FHA function in Tel1 activation, we analyzed the recombination checkpoint during meiosis in the xrs2 FHA mutant. During meiosis, the rad50S mutation causes a complete block of DSB end resection because of the inability to remove covalent bonds between Spo11 and the DSB end; this leads to a Tel1-dependent delay in meiosis I entry [4, 50, 51]. In contrast, highly resected meiotic DSBs accumulate in dmc1Δ cells, provoking a Mec1-dependent arrest in prophase I [4, 52]. First, we examined the effects of the xrs2 FHA mutation on progression meiosis I in the rad50S background. We observed that only 4.7% of the cells passed meiosis I in rad50S mutant cells after a 6-hr incubation in sporulation medium (SPM), which was substantially lower than that in wild type (49.1%, Fig 5B, left). The delayed entry into meiosis I in rad50S mutant cells was suppressed by the tel1Δ mutation, as reported previously [4], and a similar effect was observed with the xrs2-664 mutation, reflecting the loss of Tel1-binding (Fig 5B, right). The rad50S xrs2-SH mutant cells also showed progression through meiosis I, with 39% having completed meiosis I after 6 hr in SPM, almost like wild-type cells (Fig 5B, left). This suppression was also observed in other FHA truncation mutants, namely xrs2-314M, xrs2-228M and xrs2–84M (S3A Fig), even with the different level of accumulation of un-resected DSBs in the rad50S background, which is caused by different amounts of Xrs2 protein [22]. Moreover, the xrs2-664/xrs2-SH heterozygotic diploid suppressed the delayed entry into meiosis I in the rad50S background (S3B Fig), indicating that the two alleles in xrs2 do not complement each other through an intermolecular interaction. In contrast, the xrs2-SH mutation did not suppress prophase arrest in dmc1Δ (Fig 5C), which is Mec1-dependent, indicating that the function of the FHA domain of Xrs2 is not required for Mec1 activation.
A meiosis-specific axis component of the synaptonemal complex, Hop1, is phosphorylated at T318 by both Mec1 and Tel1 [53]. To monitor Tel1 activity, we thus analyzed Hop1 phosphorylation at Hop1-T318 by using an antibody specific for phospho-T318 in the rad50S background (Fig 5D and 5E). On Western blots, disappearance of Hop1 phosphorylation was delayed in rad50S (Fig 5D, pT318 asterisk), and the timing of T318 de-phosphorylation corresponded to the timing of the meiosis I transition (Fig 5B). The xrs2-SH mutation compromised Hop1-T318 phosphorylation when it was combined with rad50S, yet the level of phosphorylation was higher than that with the tel1Δ or xrs2-664 mutations (Fig 5D, pT318 asterisk). Then we examined the localization of Hop1-pT318 on meiotic nuclear spreads after 4 hr in meiosis. In the rad50S single mutant, Hop1-T318 phosphorylation was observed as punctate foci on elongated Hop1 structures (Fig 5E). Although we observed normal Hop1 staining, only a few Hop1-T318 phosphorylation signals were observed in the rad50S xrs2-SH double mutant, similar to the staining in rad50S tel1Δ and rad50S xrs2-664 cells. In contrast, in the xrs2-SH single mutant, Hop1-pT318 staining was indistinguishable from that of the wild-type cells (Fig 5E). This result suggests that FHA domain of Xrs2 is required especially for phosphorylation of chromatin bound Hop1 or maintains phosphorylated Hop1 at DSB sites. Finally, these results indicated that the FHA domain of Xrs2 is dispensable for the initial activation of Tel1 in the presence of DSB ends but it is required for its robust, prolonged activation, during both mitosis and meiosis.
We previously demonstrated that mutations in the FHA domain of XRS2 cause defects in the rejoining of the ends of a linearized plasmid and in the repair of HO-induced (complementary) DSBs by binding the C-terminal region of Lif1 [10]. These events correspond to precise NHEJ in vivo. Here we showed that the Xrs2 FHA domain is involved in the suppression of imprecise end joining and, some extent, in the removal of Ku. Taken together, these results indicate that the fidelity of end-joining reactions is compromised by mutation of the Xrs2 FHA domain. Interestingly, mutation of the FHA domain leads to a defect in C-NHEJ [9, 10] but did not affect Ku-dependent imprecise end joining (Fig 2B). This indicates that Ku-dependent imprecise end joining is genetically distinguishable from C-NHEJ.
Nijmegen breakage syndrome in humans is caused by truncation of the N-terminal region, which contains FHA domain, of Nbs1, an ortholog of Xrs2. These patients exhibit a high risk of cancer, as well as immunodeficiency, which often results from a defect in the class switch recombination pathway [54]. In addition, activation of ATM, the human ortholog of Tel1, is required for AID/APE1 activation during class switch recombination [55]. Our work thus contributes to the understanding of this lethal human disease, which arises from alternative end-joining reactions. We have demonstrated that function of the Xrs2 FHA domain is needed for robust activation of Tel1 and for maintaining this activity during the DDR (Fig 5). However, the proper recruitment of Tel1 to a DSB site requires the Tel1-binding domain in the C terminus of Xrs2 [22, 23, 47]. We note that loss of the Xrs2 FHA domain does not impair Tel1 protein recruitment to an HO-induced DSB site, unlike the C-terminal truncation xrs2-664 (Fig 3E). Interestingly, accumulation of Tel1 binding was observed in xrs2-314M but not in the xrs2-SH mutant. The xrs2-314M mutant lacks not only the FHA domain but also the BRCT-like domain (Fig 1B); it might thus be possible that the BRCT-like domain is involved in regulation of Tel1 stability at DSB ends. In addition, the xrs2-314M mutation produces a high amount of Xrs2 protein [22]. Although recruitment of excessive Xrs2 does not occur because Mre11 limits this step [22] (Fig 3B), free Xrs2 protein might affect Tel1 stability at the DSB ends. Moreover, the FHA domain of Xrs2 is required not only for activation, but also for suppression of Tel1 activity in the mec1Δ background (Fig 4C). Thus, Xrs2 is needed for regulation of Tel1 activation in multiple ways. The Xrs2-Tel1 interaction through the C-terminal domain of Xrs2 is not sufficient for robust activation and maintenance of Tel1 activity; the FHA domain of Xrs2 is required for a second step in Tel1 activation after recruitment.
FHA domains are phospho-protein recognition sites [26]. In budding yeast, the FHA domain of Xrs2 interacts with Lif1 and is involved in recruitment of DNA ligase IV complex through the interaction [10]. In the fission yeast, Schizosaccharomyces pombe, the FHA domain of Nbs1, an ortholog of Xrs2, interacts with Ctp1, an ortholog of Sae2 [56]. Similarly, phosphorylation at T90 of Sae2 is involved in its interaction with the Xrs2 FHA domain [57]. We analyzed the mutations of sae2 phosphoacceptor-site (S4A Fig) [43, 58] with respect to the frequency of imprecise end joining. The resulting mutants showed phenotypes that were indistinguishable from that of the sae2Δ mutant, but were quite different from those of the tel1Δ and xrs2 FHA mutants (S4B Fig, compare with Fig 1C and 1D). Thus Sae2 phosphorylation may create an interaction domain for Xrs2-FHA, although it probably also has other roles in repair. We note many of the human FHA domains that are associated with the BRCT domain, including that of Nbs1, recognize poly (ADP-ribose) and are involved in the DNA damage repair process [59]; poly-ADP ribosylation polymerases (PARPs), however, are absent in budding yeast [60].
The Xrs2 FHA domain was also required for robust activation of the Tel1 pathway during meiosis. We note that the initial Mec1 and Tel1 activation is shared between the mitotic DDR pathway and meiotic recombination checkpoint activation, yet the downstream signal transduction partners are quite different [61]. Xrs2 FHA function was not required for the Mec1-dependent pathway in meiosis (Fig 5C). Thus our results indicate that the Xrs2 FHA ligand may be a Tel1-specific target that is shared by the mitotic DDR and meiotic recombination checkpoint process. This could be Sae2, but there may be other candidates as well.
We showed that the tel1Δ mutation was epistatic to the xrs2 FHA mutation with respect to an increase in imprecise end joining. In contrast, dysfunction of precise NHEJ in the FHA mutant results from a defect in the interaction of the DNA ligase IV complex with Lif1 [10]. Lif1 binding by Xrs2 FHA was not, however needed for suppression of imprecise end joining (Fig 1G). Collectively, these results argue that the Xrs2 FHA domain is multi-functional. Thus, in xrs2 FHA mutant cells, addition to the defects in C-NHEJ through DNA ligase IV recruitment, compromised Tel1 activity would cause the abnormal increase of imprecise end joining.
We conclude that the enhancement of imprecise NHEJ in the xrs2 FHA mutant is due to a partial defect in Tel1 function. Repair junction sequence analysis revealed that category A, which corresponds to Ku-independent A-NHEJ, was suppressed in xrs2-664 mutant cells as in the tel1-KN cells, indicating that Tel1 recruitment to the DSB site and its kinase activity are essential for A-NHEJ at the DSB ends. The xrs2 FHA mutant, however, only showed only a slight reduction in the frequency of A-NHEJ. This result suggests that robust activation of Tel1 is not essential for A-NHEJ formation.
We found that the increase in imprecise end joining in the xrs2 FHA mutant was caused by an increase in Ku-dependent products, corresponding to categories B–E (Fig 2B, S2 and S3 Tables). In addition, we showed that Ku accumulates at HO-induced DSB ends in the xrs2 FHA mutants and also, more importantly, in the tel1Δ mutant, which did not show any defect in DSB end resection (Fig 3D). Therefore, the function of the Xrs2 FHA domain, acting through Tel1 activity, might promote Ku removal from DSB ends during end resection. Ku protein is first recruited to ds-DNA ends and possibly is then translocated to internal sites [7]. The abnormally high persistence of Ku protein or the accumulation of Ku protein at an inner region relative to processed DSB ends may activate incorrect end joining in the mutants through an interaction with Dnl4 [9] (Fig 6A). As previously noted, Tel1 activity is required not only for suppression of imprecise end joining but also for suppression of precise NHEJ [14](Fig 1D). In contrast, the xrs2 FHA mutant had a defect in precise NHEJ because of a defect in the interaction between Xrs2 and Lif1. Ku removal thus may be an important function of Tel1 in the initial steps of DDR pathway choice. In vertebrate, another DDR sensor kinase, DNA-PKcs is involved in C-NHEJ and recruited to DSB site with Ku [62]. DNA-PKcs might take charge of the yeast Tel1 function specific to Ku regulation at DSB ends.
We envision the DSB repair process as follows: First, Tel1 is recruited to unresected DSB ends in an Xrs2 Tel1-binding domain–dependent manner (Fig 6B). Then, robust activation of Tel1 through the Xrs2 FHA function is promoted at DSB ends (Fig 6C). Phosphorylation of Sae2 by Tel1 and/or Mec1 plays an important role in DSB end resection as an initial step of HR [43, 44] (Fig 6D). Then, Sae2 and Mre11 promote subsequent Exo1 activity, which is required for extensive resection of DSB ends to facilitate efficient HR [63]. In contrast, Ku complexes compete with Exo1 at this step [63]. The robust activation of Tel1, which is dependent on the Xrs2 FHA domain, is needed to remove Ku from the processed ends to prevent Ku-mediated end-bridging as well as to allow efficient and extensive resection of DSB ends. If there is no resection needed, the FHA domain of Xrs2 can promote precise NHEJ by interacting with DNA ligase IV and then again it promotes Ku removal to ensure faithful DSB repair (Fig 6B). Given that NHEJ is the dominant mechanism of repair in mammalian cells, this latter pathway may be relevant for understanding how human NBS1 mutations in the FHA domain predispose cells to genomic instability and cancer.
All yeast strains and their genotypes are shown in S1 Table. We used isogenic Saccharomyces cerevisiae W303-1A [64] derivatives for γIR sensitivity determination, SLY19 [11] derivatives for the HO-induced non-complementary DSB rejoining assay and chromatin immunoprecipitation (ChIP) assay and the SK1 background NKY1551 [65] derivative for meiotic analysis.
Rad53 and Hop1 antibodies were raised against purified recombinant proteins tagged with hexahistidine from Escherichia coli. Anti-Hop1-pT318 phosphospecific antibody was raised against synthesized polypeptide, ASIQPpTQFVSNC, and then, post-immune IgG was affinity purified by using this phosphopeptide and then titrated using a non-phosphorylated peptide, ASIQPTQFVSNC (custom-made by MBL Co., Ltd.).
End-joining activities to repair noncomplementary DSBs and repaired junction sequences were analyzed using SLY19 [11] and its derivatives as described [12]. Briefly, for each assay, a single colony was grown to log phase, and that culture was inoculated into YP-raffinose medium and incubated for 14 hr at 30°C to a concentration of 2.5 x 106 cells/ml. Then, galactose was added to a final concentration of 2% (w/v). After an additional 2.5 hr incubation, cells were plated on YP-galactose. To quantify the total number of cells, cells were plated in parallel on YPAD plates after appropriate dilutions. For DNA sequencing analysis of repaired junctions, genomic DNA, purified from the cells grown on YP-galactose and shown to be Ura−, was analyzed as described [12].
Each strain was analyzed for γIR sensitivity as described [22]. A Shimadzu Isostron RTGS-21 was used with 60Co as the ionizing radiation source (Research Institute for Radiation Biology and Medicine, Hiroshima University).
The ChIP assay was performed as described [12] with minor modifications. Cells grown to mid-log phase in YP-raffinose medium were collected before galactose addition (DSB–) and at 150 min after galactose addition (DSB +), and were treated as described [12]. The following antibodies were used for immunoprecipitation: anti-DYKDDDDK tag (1E6, Wako) for yKu70-3FLAG and 3FLAG-Tel1, anti-HA (16B12, Covance) for Sae2-3HA and anti-Myc (MC045, Nacalai Tesque) for Xrs2-13Myc detection. qPCR was performed using the SYBR green system (SsoFast EvaGreen super mix and Chromo4, Bio-Rad) with primer sets at for PHO5 on chromosome II (control) and for HO cutting sites (DSB) as follows: SLY19 DSB ChIP-f (5’-GGCCTTATAGAGTGTGGTCG-3’) and SLY19 DSB ChIP-r (5’-CAAAAGAGGCAAGTAGATAAGGG-3’). The specific recruitment of protein to HO-induced DSB ends was indicated as the relative ratio to the non-DSB locus (PHO5) control (DSB/control). The y-axis values (normalized DSB/control) are the relative ratios of the immunoprecipitation value (IP) to the input value (WCE; whole cell extract) as follows:
Normalized DSB/control = ((IPDSB/WCEDSB)/(IPControl/WCEControl)).
QAOS was performed at non-complementary DSB ends as described [40, 41]. Genomic DNA samples, which were purified at the indicated time points after HO induction, were used as the template. Primer extension reaction was carried out using a native genomic DNA sample and a heat-denatured DNA sample at 72°C with Taq DNA polymerase (Ex Taq, Takara Bio). qPCR was performed using the SYBR green system (Fast SYBR green master mix and Step one plus, Applied Biosystems) with primers for HO1 used in a previous study [41] after ExoSAP-IT (Affymetrix) treatment to remove primers from the previous reaction. The percent of ssDNA at the DSB site was calculated as follows:
ssDNA % = QAOSnative/QAOSdenatured.
Western blotting was performed as described [10]. Primary antibody binding was visualized with Alexa Fluor 680–labeled secondary antibodies (Molecular Probes) or IR dye 800–labeled secondary antibodies (Rockland) using an Odyssey infrared imaging system (LI-COR Biosciences). Antibodies used in these assays were anti-Rad53 (this study; rabbit, 1:1000), anti-Hop1 (this study, guinea pig, 1:1000), anti-Hop1-pT318 (this study; rabbit, 1:1000) and anti-α-tubulin (MCA77G; AbD Serotec).
Meiosis time course experiments were performed as described [66]. Meiotic progression was analyzed by counting the number of nuclei in each ascus under an epifluorescence microscope (Zeiss Axioskop 2) after staining with 4', 6-diamidino-2-phenylindole, dihydrochloride (DAPI). The frequencies of post–meiosis I cells containing two, three and four DAPI-stained bodies were determined. More than 200 nuclei were analyzed for each time point.
Immunostaining of yeast meiotic nuclear spreads was performed as described [10]. Stained samples were observed using an epifluorescence microscope (Zeiss Axioskop 2) with LED fluorescence light sources (X-Cite; Excelitas Technologies) and a 100× objective (Zeiss AxioPlan, NA1.4). Images were captured with a CCD camera (Retiga; Qimaging) and processed using IP lab (Silicon) and Photoshop (Adobe). Antibodies used in these assays were anti-Hop1 (1:1000) and anti-Hop1-pT318 (1:500).
|
10.1371/journal.pntd.0001575 | Ebola GP-Specific Monoclonal Antibodies Protect Mice and Guinea Pigs from Lethal Ebola Virus Infection | Ebola virus (EBOV) causes acute hemorrhagic fever in humans and non-human primates with mortality rates up to 90%. So far there are no effective treatments available. This study evaluates the protective efficacy of 8 monoclonal antibodies (MAbs) against Ebola glycoprotein in mice and guinea pigs. Immunocompetent mice or guinea pigs were given MAbs i.p. in various doses individually or as pools of 3–4 MAbs to test their protection against a lethal challenge with mouse- or guinea pig-adapted EBOV. Each of the 8 MAbs (100 µg) protected mice from a lethal EBOV challenge when administered 1 day before or after challenge. Seven MAbs were effective 2 days post-infection (dpi), with 1 MAb demonstrating partial protection 3 dpi. In the guinea pigs each MAb showed partial protection at 1 dpi, however the mean time to death was significantly prolonged compared to the control group. Moreover, treatment with pools of 3–4 MAbs completely protected the majority of animals, while administration at 2–3 dpi achieved 50–100% protection. This data suggests that the MAbs generated are capable of protecting both animal species against lethal Ebola virus challenge. These results indicate that MAbs particularly when used as an oligoclonal set are a potential therapeutic for post-exposure treatment of EBOV infection.
| Ebola virus (EBOV) causes acute hemorrhagic fever in humans and non-human primates with mortality rates up to 90%. So far there are no effective treatments available. This study evaluates the protective efficacy of 8 monoclonal antibodies (MAbs) against the Ebola virus surface glycoprotein, in mice and guinea pigs. Various combinations and doses of the neutralizing and non-neutralizing MAbs were tested, and a post-exposure treatment protocol was determined. There was 100% survival when guinea pigs received a mix of 3 neutralizing MAbs two days after a challenge with 1,000 LD50 of guinea pig-adapted EBOV. This data suggests that the MAbs generated are effective as a post-exposure therapeutic for a lethal Ebola virus infection. Development of a post-exposure therapeutic for an Ebola virus infection is vital due to the high lethality of the disease, the relative speed in which it kills, and the fact that no vaccine has been approved for human use. Additionally, is it unlikely that preventative vaccines will be employed, because Ebola virus infections occur primarily in Africa, and to date have only killed approximately 2,300 people making it financially unfeasible for a mass vaccination. Therefore, having an effective therapy in the event of an outbreak would be extremely beneficial.
| Ebola virus (EBOV) is a filovirus causing severe viral haemorrhagic fever in humans and non-human primates (NHPs) [1]. There are five species of EBOV: Zaire ebolavirus (ZEBOV), Sudan ebolavirus (SEBOV), Cote d'Ivoire ebolavirus (CIEBOV), Reston ebolavirus (REBOV), and Bundibugyo ebolavirus (BEBOV) [2]. ZEBOV has the highest virulence with a case fatality rate of 60–90% [1], [3]. Although several attempts have been made to treat EBOV infections [4]–[8], there are currently no commercially approved vaccines or effective therapies, therefore new treatments are needed. Several studies have been conducted to determine the immune correlates of protection in EBOV infections either by following natural infections, or in in vivo animal models [9]–[15]. Both T and B cell immunity was analysed and it was believed that a strong early humoral immune response may have been a factor in survival [11], [16], [17]. Additionally, in fatally infected patients EBOV-specific IgG was absent, and IgM levels were low in comparison to the survivors [16]. The passive transfer of immune sera or whole blood was tested but its effectiveness is still controversial as it has not consistently provided protection [10]. However, in mice experiments EBOV-specific sera was sufficient for improving survival [10], [18], [19].
The key target for developing effective neutralizing antibodies (NAb) is suspected to be the surface glycoprotein (GP) [20]. EBOV GP is the only protein on the surface of the virus and is responsible for receptor binding, viral entry, and cellular tropism [20]–[24]. GP-specific NAb generated in several species were protective in some animal models, however, the NAb titres are low in natural infections and their effectiveness in humans remains to be confirmed [10], [25]–[27]. Antibodies blocking viral entry, by binding the receptor or preventing viral fusion would be ideal candidates for improving survival. Additionally, the primary pathology of EBOV haemorrhagic fever is vascular injury and coagulation abnormalities, and GP has been shown to cause cytotoxicity and vascular permeability [28], [29]. In fact GP-induced cytotoxicity has been correlated with mortality rates in the different EBOV viral species [28], [30]. Taken together this suggests that prophylactic and post-exposure treatment strategies involving antibodies specific for the EBOV GP would be an effective intervention for an Ebola infection.
Monoclonal antibodies (MAbs) against ZEBOV GP have been created previously and tested in several animal models as a post-exposure therapeutic [26], [27], [31]–[34]. However, the ability of each of the MAbs to improve survival in a lethally infected animal varied considerably. Some MAbs were able to protect mice completely yet guinea pigs partially [32], [33]. One neutralizing MAb KZ52 was 100% efficacious in guinea pigs, but did not protect NHPs [26], [35]. Overall, there are a variety of mechanisms employed by MAbs to improve survival, and the ability of the MAb to neutralize the virus is not essential. The MAbs tested so far are not 100% efficacious in all animal models therefore further research is needed for more effective antibodies. The goal of this study was to test a panel of MAbs specific for the ZEBOV GP for their efficacy in protecting mice and guinea pigs from a lethal ZEBOV infection. Previously, 8 ZEBOV GP-specific MAbs had been generated using the VSVΔG/ZEBOVGP vaccine as the immunogen [36]. A preliminary study characterizing the MAbs found they all improved survival in mice infected with a high dose of mouse adapted-ZEBOV (MA-ZEBOV) [36]. As the MAbs were effective in the mouse model it is possible that these MAbs could be used as a post-exposure therapeutic for a ZEBOV infection. In this study optimization of a post-exposure protocol is undertaken in both the mouse and guinea pig model in order to determine the various treatment parameters, including the dose, treatment time, and MAb combination, that are required to provide complete protection.
All infectious animal work was performed in the biosafety level 4 biocontainment laboratory at the Public Health Agency of Canada, and approved by the Canadian Science Centre for Human and Animal Health Animal Care Committee following the guidelines of the Canadian Council on Animal Care. Animals were acclimatized for 10 days prior to the start of the experiment, and fed and monitored daily pre- and post-infection.
The recombinant virus VSVΔG/ZEBOVGP containing the Zaire ebolavirus, strain Mayinga, glycoprotein (GP) in place of the VSV glycoprotein (G) has been described previously [37]. The mouse-adapted (MA-ZEBOV) and guinea pig-adapted (GA-ZEBOV) ZEBOV strain Mayinga viruses were described previously [38], [39].
The creation of 8 MAbs (1H3, 2G4, 4G7, 5D2, 5E6, 7C9, 7G4, 10C8) has been described previously [36]. Briefly, 6–8 week old Balb/C mice were immunized with 107 pfu VSVΔG/ZEBOVGP intraperitoneally (ip), at 0, 4, and 8 weeks. A final boost with Zaire ebolavirus like particles (eVLPs) was performed before harvesting spleen cells and fusing with SP2/0 myeloma cells according to Kohler and Milstein [40]. The generation of the ZEBOV GP/VP40 eVLPs have been described previously [41].
The hybridomas were grown in Hybridoma SFM (Invitrogen), 1 mM L-Glutamine, 1× Antibiotic-Antimycotic (Invitrogen), in roller bottles at 37°C, 5% CO2. Supernatant was cleared by centrifugation and concentrated ten times using an Amicon Stirred Cell system with a 30 kDA MWCO filter (Millipore). The antibodies were purified on a HiTrap Protein G HP column(GE Healthcare) using Protein A Binding Buffer and IgG Elution Buffer (Thermo Scientific) according to manufacturers' instructions. Positive fractions were pooled, concentrated, then buffer exchanged into PBS using a 10 kDa MWCO Centriprep unit (Millipore). Antibody purity, assessed by gel electrophoresis and coomassie blue staining was >98%.
The 5–6 week old female Balb/C mice from Charles River, (Quebec, Canada), were injected ip with the indicated amount of ZEBOV GP-specific MAbs in 100 µl PBS at the times indicated either before or after i.p. infection with 1,000 LD50 of MA-ZEBOV. Female guinea pigs (Hartley strain), approximately 250 g, from Charles River, were challenged with 1,000 LD50 of GA-EBOV i.p.. At the indicated times post-infection the guinea pigs were treated i.p. with 1 ml of the MAb diluted in PBS. Naive control animals received PBS only. Clinical signs of infection and body weight were monitored for two weeks after challenge and survivors were followed three times longer than the death of the last control animal.
ZEBOVGP-specific MAbs were serially diluted from 1/100–1/12,800 in DMEM. Starting concentrations were 3.75, 3.46, 4.34 mg/ml for 1H3, 2G4, and 4G7, respectively. The MAbs were added to an equal volume of 104 pfu/ml VSVΔG/ZEBOVGP, diluted with DMEM, in order to provide 200 pfu/well. The virus-antibody combination was incubated at 37°C for 1 hour before adding 150 ul/well to a confluent 12 well tissue culture plate seeded with Vero E6 cells. After a 1 hour incubation, 1 ml of MEM 2% FBS, 1% low melting point agarose was added per well. Plates were incubated at 37°C 5% CO2 for 48 hours before adding 1 ml of 0.2% w/v crystal violet, 3.7% Formaldehyde, 2% Ethanol to each well for visualization of the plaques. The assay was performed in triplicate, and a positive control (virus with no antibody) and a negative control (no virus) incorporated. The percent reduction was calculated by averaging the count of the triplicate wells and comparing the number of plaques in the test sample against the number of plaques in the positive control (1−(Test Sample plaques/Positive control plaques))×100 = % reduction.
The log rank statistical test was performed for the Survival curve using the GraphPad Prism 4 software program. The survival curve for the MAb treated animals were compared to the survival curve for the PBS control group.
Previously, 8 MAbs specific for the glycoprotein (GP) of ZEBOV had been generated [36]. An initial characterization demonstrated they bound to a variety of GP segments, and that all 8 MAbs were able to pull down ZEBOV GP1,2 in an immunoprecipitation assay. In the current study, we further characterized the MAbs using a plaque reduction neutralization assay (PRNT50). The PRNT50 demonstrated that MAbs 1H3, 2G4, and 4G7 were neutralizing with a PRNT50 at a 1/200, 1/800, and 1/6,400 dilution, respectively (Figure 1). All of the other MAbs were non-neutralizing with 5D2 showing the highest degree of neutralization at 38% (data not shown).
Preliminary experiments in mice suggested these MAbs would be effective as a therapeutic for a MA-ZEBOV infection [36]. Therefore, a variety of parameters were assessed in order to establish the most effective treatment protocol. An in vivo mouse model was utilized to determine the protective efficacy of the individual MAbs (Table 1). Each MAb was injected either 1 day before (−1) or after (+1) a MA-ZEBOV infection (1,000 LD50) in Balb/C mice. All control mice receiving PBS only had a mean time to death of 6.6 and 5.0 days for the −1 and +1 day treatment, respectively. In contrast, mice treated with MAbs (100 µg) demonstrated either partial or complete protection. For the −1 day protocol, the MAbs 5D2, 5E6, 7C9 were most effective with a 73–87% survival rate, in comparison to the other 5 MAbs (1H3, 2G4, 4G7, 7G4, 10C8) where survival rates ranged from 0–7%. Alternatively, every MAb performed better when given at 1 day post-infection (dpi), with survival rates ranging from 40–100%. Overall, the level of protection against lethality varied with each MAb, and it appears that, in general, the MAbs are more effective when given 1 day after a lethal MA-ZEBOV infection.
Since the MAbs were most effective when given 1 day after the lethal MA-ZEBOV infection, this treatment protocol was used to determine the most effective dose for protection (Table 2). A dose response was observed, and some MAbs were more potent than others for a given dose. The lowest doses providing complete protection from lethality for 5D2, 5E6, 7C9, and 7G4 were 12.5, 25, 50, and 100 µg, respectively. MAbs 4G7 and 10C8 demonstrated an 83% survival rate at the highest dose of 100 µg. MAbs 1H3 and 2G4 were not included as they were not very effective at the highest dose in the first experiment (Table 1). In the partially protected groups of mice, the mean time to death ranged from 6.40 to 8.20 days in comparison to the control mice (5.80 days). Overall, the various MAbs varied in their potency in providing protection against a lethal MA-ZEBOV infection in mice.
Using the most effective MAb dose of 100 µg, the treatment time was extended in both directions in order to determine the optimal time for treatment, and to see how late treatment can be given before the survival rate declines (Table 3). A single dose of 100 µg for each MAb was injected either 1 or 4 days before a lethal MA-ZEBOV infection, or at 1, 2, or 3 dpi, and survival followed. Pre-treatment of the mice 4 days before infection with 1H3, 2G4, or 7G4 did not result in survival, whereas the other MAbs provided 30–90% protection. In the majority of cases, treatment 1 day before infection resulted in lower survival rates than 4 days before infection. Of the 8 MAbs, the most effective MAbs for pre-treatment were 5D2, 5E6, and 7C9. They had the highest survival rates (73–90%) and worked almost equally well on both days −4 and +1.
Extending the start of treatment after infection also had noticeable effects upon survival. Treating the mice on 1 or 2 dpi was the most effective treatment time, with some MAbs (5D2; 100%, 5E6; 93%, 7G4; 100%, 10C8; 87%) being more protective when given at 1 dpi, and the others (1H3; 50%, 2G4; 70%, 4G7; 100%, 7C9; 90%) when given at 2 dpi. Delaying the treatment to 3 dpi resulted in no survival for all MAbs except for 4G7 (10% survival rate). In general post-exposure treatment worked better than prophylactic treatment for the majority of the MAbs, except 5D2, 5E6, 7C9 which were highly effective at improving survival in mice both before and after the infection.
All MAbs were once again tested individually in the guinea pig model. The MAbs were given i.p. at 1 dpi with 1,000 LD50 of GA-ZEBOV, and survival followed (Table 4). The PBS controls all died with a mean time to death of 7.7 days. In those groups receiving treatment, with the exception of 2G4 and 4G7, none of the guinea pigs survived, but the mean time to death was significantly extended (range of 9.4–11.7 days, p<0.050). For 2G4 or 4G7, the survival rate was 60%, demonstrating that the MAbs can provide levels of protective efficacy individually in the more stringent guinea pig model.
Since individual MAbs were partially protective in the guinea pig, a second injection of the 3 neutralizing MAbs (1H3, 2G4, and 4G7) was included on 2 dpi (Figure 2). The guinea pigs were divided into 6 groups (n = 6), with one control group receiving only PBS, and 5 groups each receiving one of the non-neutralizing MAbs at 1 dpi, followed by the neutralizing MAb combination at 2 dpi. The PBS control treated animals all died with a mean time to death of 7 days. In contrast, all of the MAb treated groups demonstrated complete survival, except for 10C8 (83%). The treatment also improved morbidity as the MAb-treated groups maintained their weight in contrast to the controls that lost 6–7% of their weight by 4 and 5 dpi. This demonstrates that a combination of MAbs is an effective post-exposure treatment in guinea pigs.
As two of the neutralizing antibodies were shown to be more effective at improving survival in guinea pigs (Table 4), the 3 ZEBOV GP-specific neutralizing MAbs were delivered as a combination alone to see if they would be sufficient as a therapy for an EBOV infection (Table 5). The combination of neutralizing MAbs (1.5 mg 1H3+1.5 mg 2G4+2 mg 4G7) was given to guinea pigs either 1 day before, or 1, 2 or 3 days after a 1000 LD50 GA-ZEBOV infection, and survival followed. The PBS control group all died, with a mean time to death of 6.58±0.59 days. In contrast all animals receiving the neutralizing MAb combination at 2 dpi. survived. When the treatment was given on 3 dpi the percent survival dropped to 66.7% with a mean time to death of 11.17±3.09 days. Receiving the combination either one day before or after resulted in a survival rate of 50%, with the meantime to death of 11.17±3.09 and 7.92±0.42, respectively. Overall, the neutralizing MAb combination improved survival in all treatment protocols with the 2 dpi treatment protocol being the most effective.
In this study 8 ZEBOV GP-specific MAbs were tested for their efficacy in protecting against a ZEBOV infection in both a mouse and guinea pig model; and a post-exposure protocol for guinea pigs was optimized. Individually, each MAb extended survival partially, or completely after a lethal dose of MA-ZEBOV in mice, whereas only the 2G4 or 4G7 treated groups demonstrated a 60% survival rate against a GA-ZEBOV infection in guinea pigs. The dose response in the mouse experiment suggests some MAbs were more potent than others at improving survival, with 100% protection with 12.5 µg 5D2 in comparison to an 83% survival rate with 100 µg of 4G7 or 10C8 (Table 2). In general, the MAbs worked best for both animal models when given after the start of the infection, particularly at 1 and 2 dpi, before efficacy started to decrease at 3 dpi This suggests that there may be a limited time period in which to begin treatment after becoming infected. Within this 2 day time span, some MAbs were more effective when given at 1 dpi (5D2, 5E6, 7G4 and 10C8), and others at 2 dpi (1H3, 2G4, 4G7, and 7C9) in the mouse model. It is possible that since the MAbs and virus were both injected ip that the MAbs might inhibit ZEBOV infection of cells and extend life. As the MAbs were only partially effective when given individually in the guinea pig model, a combination of 3 neutralizing MAb (1H3, 2G4, 4G7) at 2 dpi was tested in guinea pigs and found to provide complete protection, and prevent morbidity. Each of these MAbs binds to different regions of GP1,2. 1H3 and 4G7 bind to the N- and C-terminus of GP1, respectively, whereas 2G4 binds to GP2 [36]. Targeting multiple regions of GP appears to be a successful strategy. It is possible that a variety of mechanisms for preventing infection are employed by the 8 MAbs that is reflected in the differing amounts of MAbs needed, and the time of treatment in which they are most effective. Some MAbs may have more affinity for their epitope, or the epitope may be more accessible in the natural conformation. There is precedence for this as two ZEBOV-specific neutralizing MAbs KZ52 and JP3K11 were found to neutralize ZEBOV by distinct mechanisms [42]. Based on the various studies using ZEBOV-specific MAbs as a therapy for EBOV infections, it appears that there is no way of predicting which MAb would provide complete protection. However, initial testing must begin in mice and guinea pigs in order to make the initial determination about what therapeutic approach might be best to test in NHPs and humans.
There have been several attempts at producing MAbs against ZEBOV GP however no clear pattern has emerged suggesting which primary sequence domain of GP is most immunogenic or whether neutralizing antibodies are more successful [26], [27], [31]–[34]. To date only neutralizing MAbs 133/3.16, 226/8.1 and KZ52 have shown the capacity to improve survival in guinea pigs [32], [33], [35]. MAbs 133/3.16 and 226/8.1 only provided partial protection, while 25 mg/kg of KZ52 was completely protective in guinea pigs when given at −1 or +1 hours, but had later failed to protect NHPs [26]. Individually, none of the MAbs in this study were as effective as KZ52 in the guinea pig model. However, a combination of MAbs was more effective, and the treatment could be delivered as late as 2 days after infection. This is a significant extension from the 1 hour post-infection required for KZ52. This is an important consideration as it is often difficult to begin treatment as early as 1 hour after an infection.
All 8 MAbs in this study were originally selected by screening for their SSS coating antigen [36]. Theoretically, the ability to bind to the natural conformation may be more advantageous as it is more likely that the epitopes would be available and not hidden, or that the MAbs would be able to interfere with events required for viral entry such as receptor binding, and membrane fusion. Preventing entry into the cell would decrease the overall infection levels thereby giving the immune system a better chance at controlling the infection. There are many characteristics that make an antibody effective and it may not be the same mechanism for any two MAbs. Overall, the MAbs generated in this study and the optimized protocols demonstrate their potential as a post-exposure therapeutic against a ZEBOV infection. Because previous KZ52 antibody treatments that proved effective in guinea pigs later failed to protect NHPs, it is vital that further evaluation of the neutralizing MAb combination protocol should be conducted in NHPs.
|
10.1371/journal.ppat.1004278 | A Tick Gut Protein with Fibronectin III Domains Aids Borrelia burgdorferi Congregation to the Gut during Transmission | Borrelia burgdorferi transmission to the vertebrate host commences with growth of the spirochete in the tick gut and migration from the gut to the salivary glands. This complex process, involving intimate interactions of the spirochete with the gut epithelium, is pivotal to transmission. We utilized a yeast surface display library of tick gut proteins to perform a global screen for tick gut proteins that might interact with Borrelia membrane proteins. A putative fibronectin type III domain-containing tick gut protein (Ixofin3D) was most frequently identified from this screen and prioritized for further analysis. Immunization against Ixofin3D and RNA interference-mediated reduction in expression of Ixofin3D resulted in decreased spirochete burden in tick salivary glands and in the murine host. Microscopic examination showed decreased aggregation of spirochetes on the gut epithelium concomitant with reduced expression of Ixofin3D. Our observations suggest that the interaction between Borrelia and Ixofin3D facilitates spirochete congregation to the gut during transmission, and provides a “molecular exit” direction for spirochete egress from the gut.
| Lyme borreliosis, the most common vector-borne illness in Northeastern parts of USA, is caused by Borrelia burgdorferi sensu lato spirochetes, and transmitted by the Ixodes scapularis ticks. Currently there is no vaccine available to prevent Lyme borreliosis. A better understanding of tick proteins that interact with Borrelia to facilitate spirochete transmission could identify new targets for the development of a tick-based vaccine to prevent Lyme borreliosis. Spirochete growth and exit from the gut is central to transmission, and might involve intimate interactions between the spirochete and the tick gut. We therefore performed a global screen to identify Borrelia-interacting tick gut proteins. One of the four Borrelia-interacting tick proteins, referred to as Ixofin3D, was further characterized. RNA-interference-mediated down-regulation of Ixofin3D resulted in decreased spirochete numbers in the salivary glands and consequently decreased transmission to the host during tick feeding. We demonstrate that Ixofin3D aids spirochete congregation to the gut epithelium, a critical first step that might direct spirochete exit from the gut.
| Ixodes scapularis is the predominant vector of several human pathogens including Borrelia burgdorferi, the agent of Lyme borreliosis in North America [1], [2]. There is currently no commercial vaccine to prevent Lyme borreliosis in humans [3], although recent efforts have been made in that direction [4]. An increased molecular understanding of how the tick acquires, sustains and transmits Borrelia would be conducive to the development of novel strategies, including anti-tick vaccines [5]–[7], to control Lyme borreliosis. Borrelia resides in the unfed tick gut anchored to a gut protein, TROSPA [8]. Transmission begins with the growth of the spirochetes in the gut when the Borrelia-infected tick begins to take a blood meal. Somewhere between 24 and 36 hours of tick feeding [9], coincident perhaps with an optimal quorum of spirochetes or its gene expression profile [10], [11], the spirochete travels from the gut to the salivary glands from where it exits the tick along with tick saliva into the host skin. Growth and migration from the tick gut is therefore an essential prelude to Borrelia transmission.
The spirochete proteome changes dramatically during tick feeding to facilitate migration from the gut [10]. Rudenko et al [12] showed that Borrelia infection alters the transcriptome of the Ixodes ricinus gut during feeding, suggesting a dynamic interaction between the tick gut and the growing spirochete. Consistent with this, Dunham-Ems et al [13] showed by live imaging of spirochetes in feeding I. scapularis guts that the spirochete engages intimately with the epithelial cells of the tick gut during transmission, moving away from the gut lumen towards the basal lamina of the gut. It is likely that Borrelia-gut interactions provide molecular signals that direct the movement of the spirochete from the luminal side of the gut epithelium to the basal lamina of the gut to facilitate egress from the gut. Further, Zhang et al [14] showed that interaction between a secreted tick gut protein, TRE31, and a spirochete outer surface protein BBE31 enables migration of the spirochete through the hemolymph to the salivary glands by mechanisms that remain to be understood. These observations highlight a thematic strategy of the spirochete to interact with gut proteins during growth and migration, which we are only just beginning to understand. In order to delineate the complex molecular interactions of spirochetes with the gut epithelium we screened for Borrelia-interacting tick gut proteins by probing a tick gut yeast surface display (YSD) library with Borrelia outer surface proteins. The YSD approach has traditionally utilized specific proteins individually to probe libraries of single chain antibodies to identify and characterize protein-protein interactions [15]. Work by Cho and Shusta [16] demonstrated that biotinylated whole cell lysates of mammalian cell lines or plasma membrane proteins can be used to screen a YSD library expressing human single chain antibody fragments and identify specific antigen-antibody interactions without a priori knowledge of the candidate antigens [17]. Building on this work, we have, in this study, extended the utility of YSD to examine tick gut-B. burgdorferi interactions without a priori knowledge of either interactants.
We screened 107 tick gut YSD clones with total Borrelia membrane extracts derived from in vitro-grown B. burgdorferi N40 and identified four potential Borrelia-interacting gut proteins from the initial screen. One of the predominant clones encoded a surface exposed tick gut protein with four putative fibronectin type III domains. In this report, we present our observations that suggest that the fibronectin type III domain-containing tick gut protein helps congregation of spirochetes to the gut epithelium during transmission. These observations invoke a functional role for spirochete “clustering” in spirochete egress from the gut.
B. burgdorferi membrane protein extracts were prepared as described [18] from in vitro grown B. burgdorferi (N40) temperature-shifted to 37°C for 24 hours. A YSD expression library of I. scapularis gut cDNAs [14] was probed with biotin-labeled B. burgdorferi membrane protein extracts as described in Materials and Methods. Four rounds of magnetic-activated cell sorting (MACS) screens provided a 40-fold enrichment of YSD clones expressing gut proteins that interacted with B. burgdorferi membrane proteins (Fig. 1A–B). Cells from the 4th sort were plated and one hundred colonies were individually tested for their ability to bind to B. burgdorferi membrane protein extracts by fluorescence-activated cell sorting (FACS) analysis using Alexa488-labeled B. burgdorferi membrane protein extracts. Recombinant plasmids were isolated from colonies that showed greater than 15% binding (40 clones) (Fig. 1C) and insert sizes compared by restriction digestion analysis. Clones with identical insert sizes were grouped (four groups) and two representative clones from each group were sequenced. Four unique clones encoding partial fragments of tick gut proteins were identified and provided a unique identifier based on their in silico predicted function (Table 1).
Clone 1 (identified 15 times in this screen, ∼29% binding) contained an in-frame insert that showed 100% identity to ISCW008121 (www.vectorbase.org) and encoded a protein that contained four putative fibronectin type III-domains. PSORT (http://psort.hgc.jp/) protein localization analysis suggested that it was likely a surface exposed transmembrane protein (Table 1). Clone 2 (identified 10 times in this screen, ∼30% binding) showed 100% identity to ISCW015135. ISCW015135 encoded a putative signal peptide indicative of a secreted protein as seen by Signal P analysis (www.cbs.dtu.dk/services/SignalP). No known functions or domains were identified, and no paralogs or orthologs of Clone 2 were observed by BLAST analysis. Clone 3 (identified 10 times in this screen, ∼37% binding) showed 100% identity to ISCW015049 and encoded a protein with two dystroglycan-like cadherin domains. The cellular location could not be predicted by PSORT analysis. Orthologs of clone 3 were found in eight other invertebrate vectors. Clone 4 (identified three times, 26% binding) showed 100% identity to ISCW016197 and encoded a nuclear membrane localization signal and a putative Guanylate-kinase associated protein domain, and orthologs were identified in 13 other invertebrate vectors (Table 1). B. burgdorferi is an extracellular pathogen, hence the physiological relevance of interactions between B. burgdorferi extracellular proteins and a tick nuclear protein (clone 4) is unclear and hence not prioritized for further analysis.
The expression profiles of the genes contained in the three prioritized clones were assessed by quantitative RT-PCR (qRT-PCR) in the salivary glands and guts of I. scapularis nymphs (Fig. 1D) during feeding. While ISCW008121 (Clone 1) was expressed preferentially in the gut, ISCW015135 and ISCW015049 (clones 2 and 3) were expressed both in the salivary glands and guts. Furthermore, B. burgdorferi infection increased the expressions of Clones 1 and 3 significantly (∼2-fold) in the guts after 72 hours of tick feeding (Fig. 1D).
We addressed Clone 1 in further detail because it was identified most frequently in this screen and because Clone 1 encoded a tick protein with fibronectin type III domains. While B. burgdorferi has been shown to encode lipoproteins that facilitate Borrelia adhesion to fibronectin in the mammalian host to promote infection [19], it is not known if similar interactions occur in the tick. Further, fibronectin type III domains, originally identified in the extracellular matrix protein fibronectin [20], have also been identified in several receptor-like proteins and shown to play a critical role in cell signaling [21]. Therefore, we reasoned that understanding the physiological significance of Clone 1-Borrelia interaction might provide new insights into tick-Borrelia interactions.
The protein encoded by YSD Clone 1 in this screen encompasses amino acids 97 to 449 of the protein encoded by ISCW008121 (Figure 2A). To confirm the start site of the annotated ISCW008121 (www.vectorbase.org), we performed a 5′end RLM RACE (RNA Ligase-Mediated Rapid Amplification of cDNA Ends) using fed I. scapularis gut total RNA and gene specific primers complementary to the first 411 bp of the annotated ISCW008121 gene transcript. Further, YSD Clone 1 sequence and the annotated ISCW008121 transcript did not contain a stop codon, suggesting that the annotated gene sequence did not contain a full-length 3′ sequence. Therefore, we performed a 3′end RLM RACE using fed I. scapularis gut total RNA and gene-specific primers, and identified the transcript from bp 1350 to 1836, complete with stop codon. The full-length sequence was shown to encode a ∼66 kDa protein containing four fibronectin type III domains and a transmembrane domain (Fig. 2B) and is henceforth referred to as Ixodes scapularis Fibronectin 3 Domain-containing gut protein (Ixofin3D) and is assigned the GenBank accession number KF709698.
We were unable to express the full-length protein transcript of Ixofin3D in the Drosophila expression system (DES), but succeeded in expressing a partial fragment of Ixofin3D protein encompassing amino acids 104 to 319 (rIxofin3D-PF) that is also contained in the protein fragment expressed in YSD Clone 1 (Fig. 2A). The 37 kDa rIxofin3D-PF generated in Drosophila S2 cells was glycosylated as seen by Periodic-Acid Schiff's staining (Fig. 3A). Polyclonal rabbit antibodies against rIxofin3D-PF bound to uninfected and B. burgdorferi-infected fed guts (Fig. 3B) as seen by confocal microscopy, suggesting that Ixofin3D is expressed on the surface of the gut. Consistent with the qRT-PCR analysis, quantification of pixel intensity in the TRITC channel (representing anti-rIxofin3D-PF serum binding to native Ixofin3D) using the ImageJ software showed significantly increased binding of rIxofin3D-PF antibodies to the tick gut in 24 h and 72 h fed ticks upon B. burgdorferi infection compared to that in 24 h and 72 h fed uninfected guts (Fig. 3C), and in B. burgdorferi-infected 72 h fed tick guts when compared to B. burgdorferi-infected 24 h fed guts (Fig. 3C) suggesting that Ixofin3D expression was increased in B. burgdorferi infected guts during feeding. rIxofin3D-PF incubated with in vitro grown PFA-fixed non-permeabilized B. burgdorferi N40 showed binding of rIxofin3D-PF to spirochetes as seen by indirect immunofluorescence using rabbit polyclonal antibodies against purified rIxofin3D-PF (Fig. 4A) indicating a potential interaction between Ixofin3D and an exposed B. burgdorferi protein ligand. Under similar conditions, rIxophilin, a tick gut thrombin inhibitor protein [22], not known to engage directly with spirochetes, did not show binding to in vitro grown spirochetes (Fig. 4A). Furthermore, in an ELISA assay using microplates coated with B. burgdorferi membrane protein extract, a dose-dependent increase in rIxofin3D-PF binding to B. burgdorferi membrane protein extracts was observed, whereas dose-dependent binding of rIxophilin was not observed (Fig. 4B).
To determine the role of Ixofin3D in spirochete transmission, we passively transferred purified rabbit IgG against rIxofin3D-PF into eight C3H/HeN mice and challenged these mice with Borrelia-infected I. scapularis nymphs. Control mice received purified rabbit IgG against ovalbumin (Ova). Ticks fed to repletion and engorged comparably on both control and experimental mice (Fig. 5A). Guts and salivary glands were dissected from engorged nymphs and Borrelia burden assessed by qRT-PCR. While the spirochete burden in the guts were comparable in both groups, Borrelia burden in the salivary glands was reduced in nymphs that fed on mice that received anti-rIxofin3D-PF antibodies (Fig. 5B) when compared to that in salivary glands of nymphs fed on mice that received anti-Ovalbumin antibodies, however, the decrease was not statistically significant. Borrelia burden in the skin of mice that received anti-rIxofin3D-PF antibodies (Fig. 5C) was also reduced at 7 days post tick feeding when compared to that in the skin of mice that received anti-Ovalbumin antibodies, however, the decrease was not statistically significant.
As seen with passive immunization, active immunization against rIxofin3D-PF did not impact the engorgement weights of nymphal ticks, and Borrelia burden in the nymphal guts (Fig. 5D–E). Active immunization against rIxofin3D-PF decreased Borrelia burden in the salivary glands of fed nymphs, although, the decrease was not statistically significant (Fig. 5E). However, Borrelia burden in the skin of mice at 7 days post tick feeding was significantly reduced when compared to that in the skin of mice that were immunized against ovalbumin (Fig. 5F).
To circumvent the possibility that antibodies against partial Ixofin3D might not efficiently abrogate Ixofin3D function in vivo, and to clarify the role of Ixofin3D in Borrelia transmission, we decreased the expression of Ixofin3D by RNA interference (RNAi) as described earlier [23]. Four to five double stranded (ds) ixofin3D RNA-injected nymphs or ds gfp RNA-injected were allowed to engorge on each mouse (8 mice/group). Nymphs injected with ds ixofin3D RNA engorged comparably to control nymphs injected with ds gfp RNA (Fig. 6A) despite a significant decrease in the expression of ixofin3D RNA in the guts as seen by qRT-PCR (Fig. 6B). While Borrelia burden in the guts was comparable in ds gfp and ds ixofin3D-injected nymphs (Fig. 6C), Borrelia burden in the salivary glands of fed ds ixofin3D-injected nymphs when compared to that in the salivary glands of ds gfp RNA-injected nymphs was significantly decreased (Fig. 6C). Borrelia burden in the skin of mice fed upon by ds ixofin3D RNA-injected nymphs (experimental group) at 7 and 14 days post tick feeding was significantly decreased when compared to that in the skin of mice that were fed upon by ds gfp RNA-injected nymphs (Fig. 6D).
Borrelia replicates in the gut in preparation for transmission, and adheres tightly to the gut epithelium in order to migrate towards the basal lamina of the gut epithelium, moving away from the lumen [13]. RNAi-mediated decrease in ixofin3D expression did not demonstrate alteration in Borrelia burden in the tick gut as seen by qRT-PCR (Fig. 5B). However, this assessment cannot distinguish gut epithelium-bound Borrelia from those that are not bound to the gut epithelium. Therefore, we assessed by confocal microscopy, if Ixofin3D might enhance Borrelia adherence to the gut epithelium. RNAi-mediated decrease in ixofin3D expression resulted in significantly decreased Borrelia clustering to the gut epithelium as seen by confocal microscopy (Fig. 7A) and quantification of the pixel intensity in the FITC channel (representing binding of anti-B. burgdorferi serum to spirochetes) using the ImajeJ software (Fig. 7B). Consistent with the qRT-PCR observations, the numbers of spirochetes was also significantly reduced in the salivary glands of ds ixofin3D RNA-injected nymphs (Fig. 7C–D). To further assess if Ixofin3D might facilitate spirochete aggregation to the tick gut, fed tick guts were washed to remove luminal blood-meal contents and spirochetes loosely adhering to the gut epithelium. Borrelia burden assessed in the washed gut epithelium by confocal microscopy and quantification of the pixel intensity in the FITC channel using the ImageJ software and by qRT-PCR showed decreased gut-bound Borrelia burden in ds ixofin3D RNA-injected nymphal guts when compared to that in ds gfp RNA-injected nymphal guts (Fig. 7E–G).
Understanding the biology of B. burgdorferi and its pivotal interactions with the host and the vector remains a key area of B. burgdorferi research [24]. In this study we focused on the tick gut, a tissue central to spirochete growth and transmission [11], [25], [26]. Little is known of the molecular interactions between the tick gut and the spirochete that facilitate exit from the gut for transmission to occur. We utilized a yeast surface display (YSD) approach and identified four tick gut proteins that might engage with the spirochete during transmission (Table 1). Zhang et al [14] showed that BBE31, a Borrelia outer surface protein binds to TRE31, a secreted tick gut protein. Our initial screening of the YSD library with in vitro-grown spirochetes did not identify TRE31, presumably, due to the very low expression levels of BBE31 in in vitro grown spirochetes [14].
Clone 1, the most frequently identified clone in this YSD screen, encoded a partial fragment of the gene ISCW008121. Three paralogs of ISCW008121 are represented in the Ixodes scapularis genome. BLAST analysis did not reveal orthologs of ISCW008121 in other tick species. The full-length ∼66 kDa protein referred to as Ixofin3D contained four putative fibronectin type III domains. The fibronectin type III domain was originally identified within the protein fibronectin [27]. B. burgdorferi encodes at least two proteins that bind to fibronectin, RevA [28] and BBK32 [29], [30], and this binding has been invoked in the infection of the murine host, but not the tick. ELISA assessment did not demonstrate the binding of RevA or BBK32 to Ixofin3D-PF (data not shown), suggesting that Ixofin3D might not be a fibronectin-like protein, and that the fibronectin III domains might provide a novel function that remain to be elucidated.
Although active and passive immunizations against rIxofin3D-PF showed a consistent trend towards decreased spirochete burden in the salivary glands, the decrease was not statistically significant. However, we observed a significant decrease in the skin seven days post-tick challenge upon active immunization against Ixofin3D. We expect that active immunization achieved higher levels of circulating antibodies and likely provided more efficient impairment of spirochete migration from the gut to the salivary glands when compared to that observed upon passive immunization. While a threshold of spirochete numbers critical for effective tick transmission is not defined, when lesser numbers of spirochetes are deposited in the skin they might be more vulnerable to the host immune responses. Similarly, RNAi-mediated silencing of ixofin3D expression provided a significant decrease in spirochete burden in salivary glands and consequently significantly decreased spirochete burden in the murine host skin at 7 and 14 days post-feeding. However, burden in the distal organs assessed 21 days post feeding was not different upon active immunization or RNAi-mediated knockdown of ixofin3D expression. This suggests that spirochetes that escape the initial host immune response, replicate, and disseminate successfully with time. Tick challenge experiments described herein utilized 4–5 Borrelia-infected ticks, which provides a combined inoculum of immunomodulatory tick proteins and spirochetes from 4–5 ticks, and thus potentially deflates the significance of Ixofin3D in spirochete transmission. Challenge experiments using smaller numbers of ticks might be more reflective of tick bites on humans and could be viable in studies assessing the vaccine potential of tick and Borrelia antigens.
Immunization or RNAi-mediated interruption of Ixofin3D-Borrelia interaction decreased spirochete burden in the salivary glands without any significant change in the Borrelia burden in the tick gut suggesting that Ixofin3D-spirochete interaction might facilitate spirochete entry into salivary glands or exit from the gut. Ixofin3D is not a secreted protein and is expressed preferentially in the gut, and RNAi-mediated decrease in Ixofin3D was specific to the gut. Therefore, Ixofin3D is more likely to provide a function to the spirochete in the gut. Work by Dunham-Ems [13] has shown that spirochetes migrate through the gut as sheets of spirochete aggregates. RNAi-mediated decrease in ixofin3D resulted in decreased gut epithelium-bound spirochetes. The aggregation of spirochetes on the tick gut might provide critical signals essential for spirochete migration through the gut. Coincident with tick feeding, there is a large increase in spirochete numbers [26] and residence in the luminal space would not be conducive to transmission [13]. Ixofin3D might serve as a sticky mat to facilitate spirochete congregation to the gut epithelium. The clustering of spirochetes to Ixofin3D on the tick gut might provide a molecular direction to aid spirochete exit from the gut.
Ixofin3D is expressed in uninfected nymphal ticks fed on uninfected mice, and likely serves a physiological function in the tick gut. Although tick feeding was not altered upon immunization against rIxofin3D-PF or RNAi-mediated decrease in ixofin3D expression, we cannot rule out the possibility that Ixofin3D might play a role in gut functions unrelated to feeding efficiency, and this might also modulate spirochete-gut interactions critical for transmission. The observation that Ixofin3D expression was significantly increased in B. burgdorferi-infected tick guts suggested that a specific spirochete ligand might be responsible for this increase, or it might represent a tick gut response to the spirochete. In future efforts, successful identification of the Borrelia surface protein that binds to Ixofin3D might illuminate a mechanistic understanding of Ixofin3D and its interaction with Borrelia. This study provides a new insight into tick-spirochete interactions in the gut and offers a molecular handle to unravel the biological significance of spirochete aggregation and the functional consequence on egress from the gut. Exit from the tick gut is fundamental to transmission and a molecular understanding of this event could provide new targets to prevent Borrelia transmission.
Animals were housed and handled under the Guide for the Care and Use of Laboratory Animals of the National Institutes of Health. The animal experimental protocol was approved by the Yale University's Institutional Animal Care & Use Committee (protocol number 2008-07941, approval date: 3/31/2014). All animal infection experiments were performed in a Bio-safety Level 2 animal facility, according to the regulations of Yale University.
I. scapularis nymphs and larvae were obtained from a tick colony at the Connecticut Agricultural Experiment Station in New Haven CT, USA and ticks maintained as described earlier [22].
cDNAs prepared from uninfected I. scapularis nymphs that were fed on uninfected C3H/HeN mice for 72 hours were directionally cloned by ligation into the NotI-EcoRI sites of the pYD1 yeast display vector (Invitrogen, Carlsbad, CA) to obtain a primary unamplified titre of 4×106 CFU/ml with 98% recombination efficiency [14]. Total plasmid DNA was prepared from the primary library and transformed into EBY100 Saccharomyces cerevisiae strain as described by Chao et al [31] and about 1×107 individual YSD clones were utilized for the screening as detailed below. Borrelia burgderfori N40 membrane protein extracts were purified as described by Nowalk et al [18], and biotin labeled using the EZ-Link Sulfo-NHS-Biotinylation Kit (Thermoscientific, Rockford, IL). An overnight culture of 108 yeast cells were induced as detailed by Chao et al [31], washed 3 times with cold PBS, 0.5% BSA, 2 mM EDTA (MACS) buffer, and incubated with 30 µg of biotinylated B. burgdorferi membrane proteins for 1 h at 4°C. Cells were then washed three times and incubated with anti-biotin microbeads (Miltenyi Biotec, Auburn, CA). Cells were washed three times, resuspended in 30 ml of MACS buffer and subjected to magnetic separation to enrich for B. burgdorferi-membrane protein bound YSD clones as described before [32]. The magnetically sorted cells were grown in SDCAA medium for 24 hours at 30°C and Borrelia-interacting clones were enriched by four rounds of MidiMACS sorting under the same conditions as described above. At each round of sorting, an aliquot of the induced cells was incubated with 10 µg of Alexa-488 conjugated-B. burgdorferi membrane protein extracts for 1 h at 4°C, washed three times and analyzed on a FACS Calibur flow cytometer (Beckton Dickinson, Franklin Lakes, NJ) to assess binding. Induction of the YSD library and surface expression of clones was verified by indirect immunostaining with anti-Xpress-epitope [31]. Ten thousand cells were examined on a FACS Calibur flow cytometer and data analyzed using the FlowJo software (Tree Star, Ashland, OR). For screening of individual clones, individual yeast clones were grown overnight, induced and binding to Alexa-488 conjugated-B. burgdorferi membrane protein extracts assessed as described above. YSD clones that demonstrated 15% or more binding were selected for further analysis. Plasmid DNA isolation, insert size assessment and prioritization for sequencing was performed as described earlier [32].
Ticks were allowed to feed for 24 h, for 72 h, or to repletion (between 80 and 96 h after initiation of tick feeding) and RNA isolated from guts and salivary glands using Trizol (Invitrogen, CA) as described earlier [22]. cDNA was synthesized using the iScript RT-PCR kit (Bio-Rad, CA) and analyzed by quantitative PCR for the expression of tick actin and B. burgdorferi and also ixofin3D, clone 2, 3 and 4 using gene-specific primers (listed in table S1) and the iQ SYBR Green Supermix (Bio-Rad, Hercules, CA) on a Opticon Engine MJ cycler (Bio-Rad, CA).
The RLM-RACE kit was used to identify the sequence at the 3′-end and 5′end according to the manufacturer's instructions (Invitrogen, CA). First strand cDNA was synthesized from total I. scapularis gut RNA using the 3′-RACE Adapter. The cDNA was then subjected to a PCR using the outer 3′-RACE primer, which is complementary to the distal part of the anchored adapter and an Ixofin3D specific primer Ixofin3D_882FW (Table S1) complementary to Nt 882- 901. The PCR product was then subjected to a nested PCR on an inner 3′-RACE primer and an Ixofin3D specific primer complementary to Nt 975-993, Ixofin3D_975FW (Table S1). The identification of the sequence at the 5′end was performed using Ixofin3D-specific primers Ixofin3D_595RV and Ixofin3D_475RV as described before [33]. The full-length sequence was assembled using the web-based software SMART [34] (http://smart.embl-heidelberg.de).
Ixofin3D was cloned in-frame into the pMT-Bip-V5-His tag vector (Invitrogen, CA) using ixofin3D_DESF and ixofin3D_DESR primers listed in Table S1 and recombinant protein expressed and purified using the Drosophila Expression System (Invitrogen, CA) as described earlier [22]. Recombinant protein purity was assessed by SDS-PAGE, and quantified using the BCA protein estimation kit (Thermoscientific, IL).
B. burgdorferi (strain N40) was cultured at 33°C and washed 3 times with PBS, resuspended in 500 µl PBS at 105 spirochetes/500 µl and fixed with 4% PFA for 20 min at RT and washed 3 times with PBS. Spirochetes were blocked with 5% FCS in PBS for 1 hour followed by incubation overnight with 1 µg of rIxofin3D or rIxophilin, a tick gut thrombin inhibitor [22], at 4°C. After 3 washes in PBS, spirochetes were incubated with purified polyclonal rabbit IgG against rIxofin3D-PF or polyclonal mouse IgG against rIxofilin, washed 3 times and incubated with 1∶2000 diluted anti-rabbit or anti-mouse TRITC respectively. The spirochetes were then washed 3 times and stained with FITC-conjugated goat anti B. burgdorferi antibodies (KPL, MD) for 1 hour, washed and mounted in Antifade Gold (Biorad, CA) and visualized under a fluorescence microscope (Axiovert 200M; Zeiss, Jena, Germany) at 20× magnification.
B. burgdorferi membrane extract purified as described above was coated (1 µg/ml) on high binding microtiter plates (Microlon, Greiner, Germany) overnight at RT. Wells were blocked with PBS/1% BSA at RT for 1 h and incubated with rIxofin3D-PF or rIxophilin (3–100 pmol/ml) diluted in PBS/0.05%Tween20/1% BSA for 1 h. Wells were washed and incubated with 1∶500 diluted rabbit anti rIxofin3D or mouse anti-Ixophilin IgG and bound antibody detected using HRP-conjugated anti-rabbit and anti-mouse antibody respectively (Sigma, MO) and TMB as substrate (Thermoscientific, IL).
New Zealand white rabbits 4–6 weeks old were immunized subcutaneously with 30 µg of rIxofin3D-PF or ovalbumin in complete Freund's adjuvant and boosted twice with 30 µg of rIxofin3D-PF or ovalbumin once every 3 weeks in incomplete Freund's adjuvant. Test bleeds were obtained from ear veins 2 weeks after the final boost and reactivity to recombinant rIxofin3D-PF and ovalbumin assessed by western blot. Rabbits were euthanized and serum was obtained by cardiac puncture. Polyclonal IgG was purified from the sera using the Melon Gel IgG purification kit (Thermoscientific, IL). For passive immunization, mice were passively immunized 24 h prior to tick placement by intraperitoneal inoculation with 100 µg of purified rabbit IgG against rIxofin3D-PF or ovalbumin. For active immunization, mice were immunized with 10 µg of rIxofin3D-PF or ovalbumin as described for rabbits. To address the role of rIxofin3D-PF in B. burgdorferi transmission, four B. burgdorferi N40 infected nymphs were placed on each immunized mouse. Nymphs were allowed to feed to repletion. Salivary glands and guts were dissected and combined in pools of 2–3 ticks for quantitative RT-PCR as described above. DNA was isolated from skin punch-biopsies at 7, 14 and 21 days and from heart and joints 21 days post tick-detachment and Borrelia burden assessed by quantitative PCR as described [35].
RNAi silencing of ixofin3D in ticks was performed as described before [35] using primers, specific for ixofin3D with an T7 promoter sequence, Ixofin3D_dsRNAFW and Ixofin3D_dsRNARV (Table S1). ds ixophin3D dsRNA was synthesized using the MEGAscript RNAi kit (Ambion/Invitrogen, CA). ds ixophin3D RNA or ds gfp RNA (5 nl, 3×1012 molecules/ml) was injected into the anal pore of Borrelia-infected nymphs as described earlier [35]. dsRNA-injected ticks were allowed to feed until repletion and weighed to assess feeding efficiency, and guts and salivary glands dissected for mRNA isolation and quantitative RT-PCR as described above. B. burgdorferi burden in mice was assessed by quantitative PCR as described earlier [35].
Guts from nymphal ticks (B. burgdorferi-infected or uninfected) were dissected and fixed in 4% PFA for 20 minutes, washed in PBS/0.5% Tween20 (three times) and blocked in PBS/0.5%Tween20, 5% fetal calf serum prior to sequential incubation with rabbit anti-B. burgdorferi N40 antibody and bound antibodies detected using FITC-labeled affinity purified goat anti rabbit IgG antibody (Sigma, MO) and nuclei stained with propidium idodide or with TOPRO-3 iodide (Invitrogen, CA). In experiments where Ixofin3D were visualized, tick guts were fixed as described above incubated with IgG purified from rabbit anti-Ixofin3D-PF sera. Control guts were incubated with Ig purified from rabbit anti-ovalbumin sera. Bound antibodies were detected using TRITC-labeled affinity purified mouse anti rabbit IgG antibody (Sigma, MO). All incubations were conducted in moist chambers at room temperature for 1 hour. All washes were done 3 times in PBS. Stained guts were visualized under a Zeiss LSM510 Confocal microscope.
Pixel intensities in the TRITC channel (as a measure of anti-Ixofin3D-PF serum binding to tick gut Ixofin3D) or in the FITC channel (as a measure of anti-B. burgdorferi serum binding to spirochetes) of confocal images were quantified using ImageJ 1.47t software available in the public domain (http://imagej.nih.gov/ij). Confocal images of 5–6 individual guts were examined in each control and experimental group and mean pixel intensities representing the average intensity of pixels in the region of interest were obtained in at least 5 different regions of each tick gut.
Nymphal ticks fed for 72 h were carefully dissected and the tips of each gut diverticulum nicked with a razor blade to let the luminal contents discharge out, placed in 500 µl of cold PBS and allowed to stand for 5 minutes. The supernatant was aspirated and PBS wash repeated three more times. The guts were then fixed in PFA as described above for confocal microscopy to visualize spirochetes using rabbit anti-B. burgdorferi (N40) antisera or suspended in Trizol for RNA preparation as described above.
DNA sequences obtained by Sanger sequencing were trimmed and translated to protein sequence using the Lasergene 7 DNA analysis tool (DNASTAR Inc, WI) and homology to DNA and protein sequences in the NCBI database determined by BLAST analysis [36]. Assembly of DNA sequences obtained by 5′ and 3′ RACE was performed using CodonCode Aligner 4.2.3. Predicted proteins encoded by the Borrelia-interacting YSD clones were analysed for the presence of secretory signal sequences using Signal P4.1 software (www.cbs.dtu.dk/services/SignalP), cellular localization assessed using the PSORT software (http://wolfpsort.org), protein domains using the Simple Modular Architecture Research Tool available at http://smart.embl-heidelberg.de and theoretical molecular weight (MW) and isoelectric point (pI) using ExPASy proteomics server (http://web.expasy.org/compute_pi/).
The significance of the difference between the mean values of the groups was analyzed using a non-parametric two-tailed Mann-Whitney test or a two-tailed student t test with Prism 5.0 software (GraphPad Software, San Diego, CA), and p<0.05 was considered significant. One-way ANOVA (Analysis of Variance) with Tukey's multiple comparison test was utilized when the mean values of more than 2 groups were compared.
The GenBank accession numbers and VectorBase accession numbers for genes/proteins related with this study: TROSPA: AY189148.1, Tre31: HQ998856, BBE31: NP_045436.1, Clone1: ISCW008121, Ixofin-3D: KF709698. Clone2: ISCW015135, Clone3: ISCW015049, Clone4: ISCW016197, RevA (BBM27): NP_051318.1, BBK32: AAL84596.1, Ixophilin: ISCW003862.
|
10.1371/journal.pbio.1001977 | A DNA Damage-Induced, SOS-Independent Checkpoint Regulates Cell Division in Caulobacter crescentus | Cells must coordinate DNA replication with cell division, especially during episodes of DNA damage. The paradigm for cell division control following DNA damage in bacteria involves the SOS response where cleavage of the transcriptional repressor LexA induces a division inhibitor. However, in Caulobacter crescentus, cells lacking the primary SOS-regulated inhibitor, sidA, can often still delay division post-damage. Here we identify didA, a second cell division inhibitor that is induced by DNA damage, but in an SOS-independent manner. Together, DidA and SidA inhibit division, such that cells lacking both inhibitors divide prematurely following DNA damage, with lethal consequences. We show that DidA does not disrupt assembly of the division machinery and instead binds the essential division protein FtsN to block cytokinesis. Intriguingly, mutations in FtsW and FtsI, which drive the synthesis of septal cell wall material, can suppress the activity of both SidA and DidA, likely by causing the FtsW/I/N complex to hyperactively initiate cell division. Finally, we identify a transcription factor, DriD, that drives the SOS-independent transcription of didA following DNA damage.
| Cells have evolved sophisticated mechanisms for repairing their DNA and maintaining genome integrity. A critical aspect of the repair process is an arrest of cell cycle progression, thereby ensuring that cell division is not attempted before the genome has been repaired and fully duplicated. Our paper explores the molecular mechanisms that underlie the inhibition of cell division following DNA damage in the bacterium Caulobacter crescentus. For most bacteria, the primary, and only mechanism previously described involves the SOS response, in which DNA damage induces cleavage of the transcriptional repressor LexA, driving induction of a battery of genes that includes an inhibitor of cell division (sulA in E. coli and sidA in Caulobacter). Here, we report that Caulobacter cells have a second, SOS-independent damage response pathway that induces another division inhibitor, didA, which works together with sidA to block cell division following DNA damage. We also identify the damage-sensitive transcription factor responsible for inducing DidA. Finally, our study demonstrates that DidA and SidA inhibit cell division in an atypical manner. Many division inhibitors in bacteria appear to inhibit the protein FtsZ, which forms a ring at the site of cell division. DidA and SidA, however, target a trio of proteins, FtsW/I/N, that help synthesize the new cell wall that will separate the daughter cells (the septum). In sum, our work expands our understanding of how bacterial cells respond to DNA damage and the mechanisms by which they regulate cell division.
| Progress through the cell cycle requires the sequential execution of three fundamental processes: DNA replication, chromosome segregation, and cell division. Maintaining the precise order of these events is crucial to preserving genomic integrity, as any attempt to divide before completing DNA replication or chromosome segregation could result in the scission of DNA and a failure to endow each daughter cell with a complete genome. Coordinating DNA replication and cell division is particularly challenging when cells encounter DNA damaging agents that necessitate lengthy periods of chromosome repair. To ensure the order of cell cycle events and preserve genome integrity, many cells employ checkpoints that actively halt cell cycle progression until DNA damage has been repaired. While checkpoints are prevalent and well characterized in eukaryotes [1], their role and significance in governing the bacterial cell cycle is less clear.
The α-proteobacterium C. crescentus is an excellent system for understanding the bacterial cell cycle. Cells are easily synchronized and DNA replication initiates once and only once per cell division, resulting in distinguishable G1, S, and G2 phases. As with most bacteria, cell division in Caulobacter involves the assembly of a large multiprotein complex at mid-cell that drives constriction of the cell envelope and separation of daughter cells [2]. The position of the division machinery, known as the “divisome,” is established by the tubulin homolog FtsZ, which forms a ring-like structure at mid-cell and subsequently recruits other essential cell division proteins [2]–[4]. Once assembled, how these proteins coordinate the various steps of cytokinesis is unclear and the factor(s) that ultimately trigger cytokinesis are unknown.
Like eukaryotes, bacteria can inhibit cell division following DNA damage. The best studied mechanism involves the “SOS response” [5],[6] in which DNA damage stimulates the recombinase RecA to trigger an autocatalytic cleavage of the transcriptional repressor LexA. This cleavage leads to induction of SOS genes, many of which are involved in DNA recombination and repair [6],[7]. The SOS regulon also typically includes a cell division inhibitor that can delay cytokinesis until after damage is cleared. The best characterized SOS-induced division inhibitor, Escherichia coli SulA, disrupts polymerization of FtsZ and thus inhibits assembly of the divisome [8],[9]. However, sulA is not widely conserved beyond the γ-proteobacteria and recent studies have indicated that the SOS-induced division inhibitors from several Gram-positive species do not target FtsZ, although in most cases the direct target remains unknown [10]–[12].
In Caulobacter the primary SOS-induced division inhibitor is a 29 amino acid inner membrane protein called SidA that inhibits division by interacting with the late-arriving division protein FtsW [13]. Although sidA is the primary SOS-induced division inhibitor in Caulobacter, cells lacking sidA can still arrest division when grown in the presence of the DNA damaging agent mitomycin C (MMC). An SOS-regulated endonuclease called BapE may indirectly contribute to inhibiting division [14], but we conjectured that Caulobacter encodes another direct cell division inhibitor that is induced by DNA damage but in an SOS-independent manner. Here, we identify such an inhibitor, now named didA. As with sidA, the overexpression of didA in undamaged cells is sufficient to prevent cell division. Cells lacking both inhibitors divide prematurely following DNA damage, leading to a significant viability defect. DidA does not disrupt FtsZ ring formation or divisome assembly and instead likely inhibits division through an interaction with the divisome component FtsN. Intriguingly, point mutations in FtsW and FtsI, which help drive septal cell wall synthesis, suppress the lethality that results from overproducing either SidA or DidA. Our results suggest that these mutations hyperactivate the cell division process and implicate the protein complex FtsW/I/N in the triggering of cytokinesis. Finally, we identify a transcription factor, DriD, that activates didA expression, thus revealing the basis of a damage-inducible, but SOS-independent pathway in Caulobacter.
Our previous work demonstrated that sidA is the primary SOS-induced division inhibitor in Caulobacter. However, many ΔsidA and ΔrecA cells exposed to the DNA damaging agent MMC still become filamentous suggesting that an SOS-independent inhibitor may also prevent division following DNA damage (Figure S1) [13]. To identify candidate inhibitors, we examined global gene expression changes following MMC treatment of a ΔrecA strain, which cannot induce SOS genes. Wild-type and ΔrecA cells were grown to mid-exponential phase in rich medium and exposed to MMC for 30 minutes. RNA was then isolated and compared to mock treated cells on whole genome DNA microarrays (Data S1A).
Of the 50 most upregulated genes following MMC treatment in wild-type cells, 44 were recA-dependent, including 31 that are directly regulated by LexA (Figure 1A and S2A) [13],[15]. The remaining six damage-regulated genes showed similar induction levels in both wild-type and ΔrecA backgrounds (Figure 1A) and are thus likely controlled by an SOS-independent mechanism. One of these genes, CCNA03212 in the NA1000 (CB15N) genome, encodes a previously uncharacterized 71 amino acid protein with a single predicted transmembrane helix flanked by short cytoplasmic and periplasmic domains (Figure 1B). The open reading frame of CCNA03212 overlaps with the C-terminus of the open reading frame of CC3114, annotated in the closely related strain CB15. In our expression profiling experiments, only those probes lying within the CCNA03212 coding sequence were significantly upregulated in wild-type cells treated with MMC (Figure S2B and S2C), suggesting that the NA1000 annotation is correct. Based on the studies described below, we named this gene didA (for damage-induced cell division inhibitor A).
To confirm that didA encodes a damage-inducible protein, we created a strain in which the chromosomal didA gene was fused to the coding region of the 3×M2 epitope. This C-terminal fusion, DidA-3×M2, was barely detectable in the absence of DNA damage, but was strongly induced following MMC treatment with protein levels increasing nearly 20-fold after 1 hour (Figure 1C). Western blotting indicated a band at the size predicted for DidA-3×M2 (∼11 kDa) and not CC3114-3×M2 (∼25 kDa) indicating that the larger gene product annotated in CB15 is not produced at significant levels in these conditions. To test the SOS-dependence of DidA-3×M2 synthesis following MMC treatment, we examined DidA-3×M2 production in a ΔrecA strain and in a strain harboring lexA(K203A), which encodes a noncleavable form of LexA that blocks the induction of SOS genes. In each case, DidA-3×M2 was slightly elevated in untreated cells, likely due to increased basal levels of damage in the absence of SOS-mediated repair (Figure 1D). Following MMC treatment, DidA-3×M2 was strongly induced in all strains (Figure 1D), consistent with an SOS-independent mode of regulation.
To test whether DidA can inhibit cell division, we fused the didA coding sequence to the vanillate-inducible promoter Pvan and cloned this construct into both low- and medium-copy plasmids. We transformed wild-type cells with each plasmid and then grew cells in the presence of vanillate to induce didA in the absence of a DNA damaging agent. Synthesis of DidA from the low-copy plasmid resulted in mild cellular filamentation and a modest growth defect, while overproduction from the medium-copy plasmid caused a more pronounced division defect with nearly all cells demonstrating severe filamentation after 6 hours (Figure 2A and 2B). Thus, DidA, like SidA, is sufficient to inhibit cell division in the absence of DNA damage.
To assess the level of DidA accumulation during our overproduction experiments, we fused the coding region for a 3×M2 tag to the 5′ end of didA and expressed this construct from its native promoter on the chromosome or from the Pvan promoter on a low- or medium-copy plasmid. After 3 hours of induction, cells producing DidA from either plasmid became filamentous indicating that 3×M2-DidA is functional (Figure S3A). As expected, cells expressing 3×M2-didA from the native chromosomal locus also became filamentous following treatment with MMC. Importantly, the levels of 3×M2-DidA that led to filamentation when produced from either plasmid were slightly lower than that seen when produced from the native locus during MMC exposure (Figure S3B), indicating that the phenotypes observed in Figure 2 are not the result of artificially high DidA levels. Taken together, our results suggest that following DNA damage, DidA accumulates in an SOS-independent fashion to help prevent cell division.
To test whether DidA is necessary to block cell division following DNA damage, we constructed a strain in which all but the first and last three amino acids of didA were deleted. As with a sidA deletion strain, ΔdidA cells grown on plates containing MMC showed no major viability defect (Figure 3A). However, a strain lacking both sidA and didA showed a pronounced defect, with a nearly 100-fold decrease in plating efficiency (Figure 3A). This decreased viability was rescued by the presence of either inhibitor on a low-copy plasmid (Figure 3B). These results indicate that SidA and DidA are, to some extent, functionally redundant in blocking cell division following MMC-induced DNA damage.
To better understand the DNA damage sensitivity of ΔsidAΔdidA cells, we used time-lapse microscopy to examine synchronous populations of swarmer cells during growth on agarose pads containing MMC. Wild-type swarmer cells did not divide for ∼5 hours on average (Figure 3C), which is significantly longer than the average time to first division of 1.9 hours for wild-type swarmer cells grown on MMC-free pads. On MMC pads, roughly 5% of wild-type cells arrested growth following a cell division event (Figure 3D-3E and Data S2), indicating that division may have been premature or inappropriately executed and was, consequently, lethal. The single deletion strains, ΔsidA and ΔdidA, also delayed cell division in the presence of MMC; the average time to division was not significantly different than for wild-type cells. These single deletion strains had 1.5–2 times as many growth arrested cells following division events compared to the wild type, although these defects were apparently insufficient to produce a gross viability defect (Figure 3A and 3D). In contrast to the single mutants, ΔsidAΔdidA cells lacking both inhibitors divided ∼1.25 hours earlier than wild-type (p = 6.9×10−10), and four times as many cells exhibited growth defects following a division event (Figure 3C–3E; Data S2). Taken together, our data suggest that the lethality experienced by ΔsidAΔdidA cells in the presence of MMC results from an inability to appropriately delay cell division.
We next sought to investigate how DidA disrupts cell division. We first asked whether DidA interferes with cell division directly, through an interaction with the divisome, or indirectly by inducing the SOS regulon or inhibiting the cell cycle regulator CtrA. To investigate the possibility of indirect mechanisms, we isolated RNA from cells overproducing DidA from a medium-copy plasmid for 45 minutes and compared it on DNA microarrays to RNA from similarly treated cells grown in the absence of inducer. No significant gene expression changes were observed in the SOS or CtrA regulons (Data S1B) suggesting that DidA acts post-transcriptionally, and possibly directly, to inhibit cell division.
To further explore how DidA inhibits cell division, we examined its subcellular localization. In predivisional cells, the major components of the cell division machinery are located at mid-cell [2] where they synthesize a septum and drive invagination of the cell envelope. To assess DidA localization, we transformed wild-type cells with a low-copy plasmid harboring an M2-yfp-didA fusion under the control of a xylose-inducible promoter. After induction for 3 hours, cells became filamentous indicating that the YFP-DidA fusion inhibits cell division (Figure 4A). Notably, YFP-DidA foci were frequently observed at pinch sites near mid-cell (Figure 4A) placing it in close proximity to the cell division machinery. Further, fractionation of cells overproducing 3×M2-DidA indicated that DidA is strongly enriched in the membrane where many of the middle- and late-arriving cell division components also reside (Figure 4B). These data are consistent with a model whereby DidA inhibits division through an interaction with a component of the divisome.
To test for interactions of DidA with the known set of critical Caulobacter cell division components [2], we performed a bacterial two-hybrid analysis as used previously with SidA [13],[16]. Briefly, proteins were fused to either the T18 or T25 subunit of adenylate cyclase and co-expressed in E. coli; a protein-protein interaction reconstitutes adenylate cyclase and drives synthesis of cyclic-AMP, causing colonies to appear red on MacConkey agar plates. When expressed from the low-copy plasmid pKT25, a T25-DidA fusion interacted almost exclusively with the late-arriving cell division protein fusion T18-FtsN (Figures 4C and S4A). Identical results were obtained in the reciprocal orientation, with a T18-DidA fusion on the high-copy plasmid pUT18C and individual division proteins produced from pKT25 (Figure S4B). SidA, whose primary target is likely FtsW, also interacts, to some extent, with FtsN (Figure 4C) [13]. In sum, our data suggest that DidA is an integral membrane protein that localizes to mid-cell where it may disrupt cell division through an interaction with FtsN.
FtsN is among the last cell division proteins to arrive at mid-cell prior to cytokinesis. Although its precise function is unknown, FtsN interacts with multiple division proteins and may help stabilize the assembled divisome [16]–[18]. To ask whether DidA destabilizes or blocks assembly of the divisome, we examined the localization of early- and late-arriving division proteins during DidA overproduction. Cells producing fluorescently tagged FtsZ, FtsW, FtsI, or FtsN were transformed with a plasmid for overexpressing didA and then grown in the presence of vanillate to induce DidA synthesis. After 4.5 hours of induction, cells expressing ftsZ-yfp, venus-ftsW, or gfp-ftsI were inhibited for cell division, but 89%, 95%, and 85% of cells, respectively, contained fluorescent foci at or near visible pinch sites (Figure 4D). These results indicate that DidA likely does not disrupt the localization of cell division proteins or drive the disassembly of division protein complexes. Additionally, we noted that many cells displayed multiple foci of the FtsZ, FtsW, or FtsI fluorescent fusions suggesting that DidA also does not prevent the formation of new division assemblies.
Intriguingly, cells expressing gfp-ftsN were noticeably shorter (12.6±0.65 µm standard error of the mean [SEM]) and more pinched than those expressing ftsZ-yfp, venus-ftsW, or gfp-ftsI (22.8±0.79, 24.7±1.05, 26.1±0.74 µm, respectively) (Figure 4D). Further, cells expressing gfp-ftsN robustly formed colonies despite DidA overproduction, in contrast to cells expressing the other fluorescent fusions (Figure 4E), indicating that gfp-ftsN functions as a DidA suppressor, possibly by decreasing its affinity for DidA or by stabilizing FtsN and thereby increasing FtsN levels. In either case, these data further support a model in which DidA interacts with FtsN to block cell division, but without disrupting assembly of an intact divisome.
We next sought to determine whether point mutations in FtsN can also suppress the lethality of overproducing DidA. We first constructed a low-copy plasmid on which 3×M2-didA was transcribed from the IPTG-inducible promoter Plac. We then used mutagenic PCR to create a library of ftsN mutants containing, on average, one nucleotide substitution per coding sequence; these ftsN mutants were cloned into a medium-copy plasmid with expression driven by Pxyl. The didA expression vector and ftsN plasmid library were co-transformed into an ftsN depletion strain in which the only chromosomal copy of ftsN is transcribed from the Pvan locus [19]. Cells were plated in the presence of IPTG to induce 3×M2-DidA, but without vanillate such that only plasmid-produced, mutant FtsN accumulated. From ∼168,000 cells plated, two candidate ftsN suppressors were isolated that suppressed the lethality of overproducing DidA. Plasmid sequencing indicated that one clone contained a single mutation, ftsN(L202P), while the other contained two mutations, ftsN(P156S) and ftsN(F252L).
Each mutation was introduced into an otherwise wild-type chromosome and tested for its ability to suppress 3×M2-DidA overproduction. Only those cells harboring the ftsN(L202P) or ftsN(F252L) mutation maintained 3×M2-DidA suppression (Figure 5A and 5B), indicating that ftsN(P156S) was likely a passenger mutation with ftsN(F252L). Intriguingly, both bona fide suppressor mutations reside within the periplasmic, C-terminal “SPOR” domain of FtsN, which may bind peptidoglycan structures within the actively dividing, septal cell wall [19]–[21].
To further explore the regions of FtsN that bind DidA, we tested a series of FtsN truncations and chimeras in the bacterial two-hybrid system (Figure 5C). T25-DidA still interacted with an FtsN construct whose cytoplasmic and transmembrane domains were replaced with the transmembrane domain of the E. coli permease MalF, but not with a MalF fusion to the divisome component FtsA. In contrast, the DidA-FtsN interaction was significantly weakened when FtsN constructs lacked either its entire periplasmic portion or the periplasmic SPOR domain alone. We also noted that DidA still interacted robustly with an FtsN construct in which the only known essential domain, located within the periplasmic linker region and denoted “H1” [19], was replaced with an unstructured region of the Caulobacter protein SpmX. Collectively, these results suggest that DidA binds the periplasmic SPOR domain of FtsN where the suppressor mutations L202P and F252L reside. Moreover, we found that, when introduced into T18-FtsN, each suppressor mutation strongly reduced the interaction with DidA compared to wild-type FtsN or FtsN(P156S) which, as noted, does not suppress DidA lethality (Figure 5C). Importantly, each of the FtsN mutants tested interacted with FtsW as well as the wild-type FtsN did, indicating that the mutants were properly expressed and folded. In summary, our results suggest that DidA binds the SPOR domain of the late-arriving divisome component FtsN, and the substitutions L202P and F252L in this domain suppress the lethality of overproducing DidA by reducing its affinity for FtsN.
To further explore the mechanism by which DidA inhibits division, we also screened for spontaneous mutations that suppress the lethality of overproducing DidA. Wild-type cells carrying a medium-copy plasmid expressing 3×M2-didA from Pvan were grown on plates containing vanillate to induce 3×M2-DidA. Because wild-type cells overproducing 3×M2-DidA cannot form colonies (Figure 5B), those rare colonies arising on plates containing vanillate represent strains harboring putative suppressor mutations. From roughly 3×107 plated cells, 34 suppressors were identified, although only one strain retained high levels of functional 3×M2-DidA. Whole genome resequencing identified a putative suppressor mutation in ftsW, which would produce the substitution A246T in the predicted large periplasmic loop of FtsW (Figure 5A). This mutation was created de novo in a wild-type background and confirmed to suppress the lethality of overproducing DidA (Figure 5B). As noted, no interactions between DidA and FtsW were observed in our two-hybrid analysis. This could be a false negative; alternatively, FtsW(A246T) may suppress DidA overproduction by promoting an activity of FtsW rather than by preventing binding of the inhibitor.
Intriguingly, we had previously found other mutations in ftsW that suppress the lethality of overproducing SidA [13]. We therefore reasoned that SidA and DidA may function similarly to inhibit cell division. To explore this possibility, we asked whether the previously identified suppressors of SidA overproduction could also suppress DidA overproduction, and vice versa (Figure 5A and 5D). Several mutations primarily suppressed the lethality of only one of the inhibitors. For instance, the FtsW(A31K) strain strongly suppressed overproduction of M2-SidA but not DidA, whereas the strains producing FtsN(L202P) or FtsN(F252L) suppressed the activity of DidA but not M2-SidA. These inhibitor-specific suppressors likely prevent binding of their respective inhibitors (Figures 5C and S5) [13]. The other mutations showed varying abilities to suppress the lethality associated with overproducing either inhibitor. In particular, the strains producing FtsW(F145L) or FtsW(A246T) showed robust suppression of both inhibitors.
The ability of these single substitutions, F145L and A246T, to suppress the lethality of overproducing either SidA or DidA could indicate that the inhibitors share a binding site within FtsW that is disrupted by the suppressor mutations. However, this is unlikely given that (1) DidA binds FtsN, but not FtsW, in our bacterial two-hybrid system, (2) DidA-YFP still localizes to the septum in cells producing FtsW(A246T) (Figure S5A), and (3) M2-SidA binds to FtsW(A246T) to the same extent as it does to wild-type FtsW (Figure S5B). Instead, we hypothesized that the subcomplex of late-arriving division components FtsW, FtsI, and FtsN could exist in one of two states: an active state that promotes constriction of the septum and cell division, and an inactive state that is promoted or stabilized by SidA and DidA. In this model, the suppressor mutations in ftsW and ftsI promote the active state and thus enable cell division even in the presence of SidA and DidA.
If the FtsW(F145L) and FtsW(A246T) mutations promote an active state of a subcomplex of cell division proteins, then cells harboring these mutations, but not producing SidA or DidA, may attempt division earlier than wild-type cells, even in the absence of DNA damage. To explore this possibility, we grew strains harboring one of the suppressor mutations in ftsW, ftsI, or ftsN into mid-exponential phase in rich medium and measured cell lengths in a large population of cells. Indeed, several of the suppressor mutations resulted in cells that were significantly shorter on average than wild-type cells even though their growth rates were not substantially different (Figures 6A, 6B, and S6A–S6C). For ftsW(A246T), we verified that all cell types were shorter, indicating that the mutant strains are not trivially enriched for swarmer cells (Figure S6B). The degree of shortening roughly correlated with the ability to suppress both SidA and DidA activity, as cells harboring the mutations ftsW(A246T), ftsW(F145L), and ftsI(I45V) that were best able to suppress both SidA and DidA were also the shortest. Conversely, mutations that only suppressed the activity of one inhibitor were typically not shorter than wild-type. We found that ΔsidAΔdidA cells were also not shorter than wild-type cells. Taken together, these results are consistent with a model in which suppressors exhibiting short cell phenotypes harbor gain-of-activity mutations rather than simply being defective for SidA or DidA binding.
Given that the ftsW(A246T) mutation renders cells insensitive to SidA and DidA, this suppressor strain should also divide earlier than wild-type cells in the presence of MMC like the ΔsidAΔdidA deletion strain. To test this prediction, we grew populations of wild-type and ftsW(A246T) cells on agarose pads containing MMC and measured the time to first division by time-lapse microscopy. The ftsW(A246T) cells divided an average of 35 minutes earlier than wild-type cells and showed a 5-fold increase in the fraction of cells that stopped growing following a division event (Figure S7A and S7B). Accordingly, ftsW(A246T) cells showed a similar sensitivity on MMC plates as observed with the ΔsidAΔdidA strain (Figure 6C).
Although the ftsW(A246T) and ΔsidAΔdidA strains behave similarly in the presence of MMC, only the ftsW(A246T) strain exhibited a short cell phenotype when grown without MMC (Figure 6A and 6B). The ftsW(A246T) cells grew at approximately the same rate as wild-type cells in the absence of MMC; these cells are born shorter than wild-type cells, but also divide when shorter than wild-type cells resulting in nearly identical division cycle times (Figure S6B–6D). The short cell phenotype of this strain in the absence of MMC suggested that FtsW(A246T) harbors increased cell division activity, and has a propensity to divide early, compared to wild-type and ΔsidAΔdidA cells. To further explore this activity, we combined the three suppressor mutations conferring the shortest cell length phenotypes, ftsW(A246T), ftsI(I45V), and ftsW(F145L), engineering each on the chromosome of a single strain. When grown in the absence of MMC, this triple mutant, denoted ftsW**I*, was slightly shorter than the single ftsW(A246T) mutant and exhibited an increased sensitivity to MMC compared to the ftsW(A246T) and ΔsidAΔdidA strains (Figure 6C). These results suggest that the triple mutant likely harbors increased activity relative to the single ftsW(A246T) mutant that alone causes cells to attempt divisions more hyperactively both in the presence and absence of MMC.
We also noticed that the ftsW**I* strain grew more slowly than wild-type or ftsW(A246T) cells in liquid cultures (Figure S6C). Because FtsW and FtsI participate in septal cell wall synthesis, we suspected that this growth phenotype may result from premature or misregulated cell division events that compromise cell wall integrity. To test this possibility, we stained wild-type, ΔsidAΔdidA, ftsW(A246T), and ftsW**I* cells with propidium iodide (PI), a dye that binds nucleic acids, but only if the cell envelope is compromised (Figure 6D). Whereas wild-type, ΔsidAΔdidA, and ftsW(A246T) cells were rarely (0.1%–0.3% of cells) stained by PI, 2.6% of ftsW**I* cells were PI-positive. Given these results, we also tested whether the ftsW(A246T) and ftsW**I* strains were more sensitive than wild type when treated with cephalexin, which interferes with septal cell wall synthesis by blocking the transpeptidase activity of FtsI. Cephalexin does not directly cause DNA damage, and cells treated with cephalexin showed no noticeable induction of sidA or didA (Figure S8). It was thus not surprising that ΔsidAΔdidA cells showed no growth defect compared to wild-type when grown on plates containing a low dose of cephalexin that does not significantly perturb growth or division in wild-type cells (Figure 6C and 6D). In contrast, the ftsW(A246T) and ftsW**I* strains each exhibited cephalexin sensitivity, particularly ftsW**I* (Figure 6C). When grown as liquid cultures with cephalexin, the ftsW(A246T) and ftsW**I* strains had 18- and 48-fold, respectively, more PI-positive cells than wild-type (Figures 6D and S9). By contrast, there was not a similar enrichment of PI-positive cells in the ftsW(A246T) and ftsW**I* strains following an MMC treatment. Furthermore, while the average lengths of cells from the suppressor strains were decreased relative to wild type in MMC, likely due to premature divisions, they were longer in cephalexin, indicating a decreased ability to divide ().
In sum, cells harboring the mutation ftsW(A246T), either alone or in combination with ftsI(I45V) and ftsW(F145L), exhibit cell wall defects and are more sensitive to a cell wall synthesis inhibitor. Importantly, cells lacking sidA and didA do not exhibit these same cell wall defects. These results are consistent with a model in which the mutations identified in ftsW and ftsI do not suppress SidA and DidA by simply preventing the binding of these inhibitors, but instead affect septal cell wall synthesis and increase the propensity of cells to initiate cell division.
Our identification of didA indicates that Caulobacter cells have an SOS-independent mechanism for sensing and responding to DNA damage. To explore this alternative, damage-inducible pathway, we first asked whether didA is induced specifically by DNA damage or more generally by cellular stress. Cells harboring a didA-3×M2 fusion at the native didA locus were treated with a variety of stresses, but the only conditions leading to a significant induction of didA were DNA damaging agents (Figure S10).
To further examine didA induction and compare it to sidA induction, we transformed wild-type cells with plasmids harboring a transcriptional fusion of egfp to either the sidA or didA promoter and then treated each strain with (i) MMC, an alkylating agent that forms single-stranded DNA adducts and double-stranded cross-links, (ii) hydroxyurea, which depletes the dNTP pool by inhibiting ribonucleotide reductase and stalls replication forks, thereby mimicking a consequence of DNA damage, or (iii) zeocin, which directly cleaves DNA, creating double-strand breaks. Western blots for GFP indicated that MMC strongly induced both sidA and didA (Figure 7A). In contrast, hydroxyurea drove induction of PsidA, but not PdidA, even at high doses. Conversely, zeocin strongly induced PdidA, but only weakly induced PsidA. These data indicate that the SOS-independent induction of didA involves a signal or DNA structure that is distinct from the ssDNA-RecA-dependent induction of sidA. In particular, the strong induction of PdidA by zeocin suggests that the signal may be a DNA structure associated with the presence or repair of double strand breaks, which also arise following MMC exposure [22].
We devised a genetic screen to identify factors involved in didA induction. In a ΔdidA background, we fused the didA promoter to lacZ and integrated this reporter construct at the hfaB locus, a region of low transcription. When grown in the presence of X-gal, colonies with high PdidA activity should express lacZ and appear blue while those with low PdidA activity should appear white. We mutagenized this strain using a Tn5 transposon and screened for mutants on X-gal plates containing MMC. We chose a dose of MMC low enough to allow colony formation, but high enough to induce didA induction resulting in blue colonies. We screened ∼26,000 colonies and isolated nine white colonies; five of these colonies had Tn5 insertions in the PdidA-lacZ reporter while the remaining four contained insertions in the coding region of CCNA_01151 (Figure 7B). This gene is annotated as a DeoR-family transcriptional regulator and is predicted to encode an N-terminal DNA-binding domain with a C-terminal ligand-binding domain (“WYL domain,” Pfam domain 13280). Each of the four insertions in CCNA_01151 was unique with one occurring in the DNA-binding domain and the other three in the C-terminal WYL domain. We named CCNA_01151 driD (for DeoR inducer of didA).
To confirm that DriD induces didA, we constructed a strain in which all of driD except the first three and last ten amino acids were deleted. We then transformed wild-type, ΔdriD, and ΔrecA cells with low-copy plasmids harboring PsidA-egfp or PdidA-egfp reporters and monitored the inducibility of each promoter following MMC or zeocin treatment by Western blotting with α-GFP (Figure 7C). As expected, sidA induction by either DNA damaging agent requires the SOS regulator gene recA but is unaffected in cells lacking driD. In contrast, didA induction occurs in ΔrecA cells but not in cells lacking driD. These results confirm the SOS-independent inducibility of didA and indicate that driD is required for didA induction. We also tested whether the driD deletion behaves like a didA deletion with respect to MMC sensitivity (Figure 7D). Indeed, cells lacking both sidA and driD exhibited a roughly 100-fold reduction in viability when grown on MMC plates, compared to the wild type and strains lacking either sidA or driD. A nearly identical defect was observed when combining sidA and didA deletions, further supporting a model whereby DriD drives didA induction.
We next sought to complement our driD deletion by introducing low-copy plasmids containing PdriD fused to wild-type driD or a copy of driD encoding an N- or C-terminal fusion to the 3×M2 epitope; each strain also harbored a chromosomal didA-3×M2 reporter to assess DriD activity. Whereas cells carrying an empty vector were unable to induce didA when treated with zeocin, cells with wild-type or either tagged version of driD were able to induce didA (Figure 7E, bottom panel). Additionally, we noted that the levels of both 3×M2-tagged DriD constructs remained unchanged following zeocin treatment (Figure 7E, top panel) indicating that DriD activity is regulated post-translationally.
Finally, to determine whether DriD directly activates didA, we assessed DriD occupancy at PdidA using chromatin immunoprecipitation (ChIP) followed by quantitative PCR. Cells expressing driD or driD-3×F from a plasmid as the only copy of driD were treated with zeocin for 45 minutes or left untreated and then subjected to ChIP using an α-FLAG/M2 antibody (Figure 7F). PdidA was minimally enriched (normalized IP output/input) in the immunoprecipitate of cells expressing untagged DriD. In cells expressing driD-3×M2, PdidA was enriched roughly 3.5-fold in the absence of zeocin and nearly 30-fold following zeocin treatment. Taken together, our data suggest that DriD is a direct, positive regulator of didA induction that is enriched at the didA promoter following certain types of DNA damage, including double-strand breaks.
During episodes of DNA damage, cells often use checkpoint systems to transiently inhibit the cell cycle and prevent cell division [23]. In bacteria, the regulatory paradigm for responding to DNA damage has long been the E. coli SOS system in which cleavage of the repressor LexA drives the transcription of DNA repair genes and the cell division inhibitor sulA [8],[9],[24]. SOS-induced division inhibitors have subsequently been identified in a range of other bacteria, including sulA homologs in γ-proteobacteria and the unrelated genes yneA, divS, chiZ, and sidA in various other species [10]–[13],[25]. Although these SOS-dependent regulators are often assumed to be the primary, or even sole, mechanism for inhibiting division post-damage, there have been hints of SOS-independent division regulation. For instance, in E. coli, Bacillus subtilis, and Caulobacter, cells lacking their SOS-induced inhibitors or unable to induce an SOS response can still become filamentous following DNA damage indicating an alternative means of blocking cell division [26]–[30]. However, to the best of our knowledge, no damage-induced, SOS-independent division regulators have been previously documented. Here, we identified didA in Caulobacter as one such regulator.
How do Caulobacter cells recognize and respond to DNA damage to induce didA if not through the canonical derepression of SOS genes? DriD is a direct transcriptional activator of didA, but how does DriD sense DNA damage? One possibility is that DriD somehow senses the accumulation of the SOS signal ssDNA, which stimulates RecA to trigger the autocatalytic cleavage of LexA [31]–[33]. Another protein, such as the RecA homolog RadA, could also recognize ssDNA, but ultimately activate DriD. However, this scenario is unlikely given the differential induction of sidA and didA following exposure to DNA damaging agents with distinct mechanisms. Alternatively, a DNA damage sensor unrelated to RecA could recognize a distinct type of DNA damage or DNA structure. For instance, the strong induction of didA following zeocin exposure could indicate that the didA induction machinery recognizes double-strand breaks. In B. subtilis, the diadenylate cyclase DisA monitors genome integrity and may recognize branched DNA structures that arise during the recombination-based repair of double-strand breaks [34]. When paused at such DNA structures, DisA is prevented from synthesizing cyclic-di-AMP (c-di-AMP), a diffusible molecule required for the activation of the transcription factor Spo0A, thereby coupling DNA damage with transcription [34]–[36]. It remains unclear precisely how c-di-AMP affects Spo0A activity in B. subtilis and whether a c-di-AMP-based response to DNA damage extends to other organisms. Nonetheless, didA transcription could follow a similar regulatory strategy that relies on c-di-AMP, or another damage-regulated second messenger. This is a particularly attractive hypothesis since DriD, annotated as a DeoR-family transcription factor has a C-terminal domain predicted to bind a small molecule. Additionally, we found that DriD levels did not change following zeocin treatment, but occupancy and activation of the PdidA promoter by DriD increased significantly. This finding suggests that DriD activity is post-translationally regulated in a DNA damage-dependent manner, so identification of the putative DriD ligand will be a critical next step.
Many cell division inhibitors, including E. coli SulA, block cell division by disrupting FtsZ polymerization. FtsZ is an effective target as it recruits most other cell division proteins. However, neither DidA nor SidA affect the assembly of FtsZ rings in Caulobacter or stimulate Z-ring disassembly, and neither inhibitor prevents the assembly of downstream divisome components. Instead, these inhibitors appear to block cell division by targeting FtsW, FtsI, and FtsN within the assembled divisome. Bacterial two-hybrid studies indicated that DidA interacts with FtsN. Additionally, several point mutations in ftsN diminish the interaction with DidA and suppress the effects of overproducing DidA, supporting a model in which DidA inhibits cell division by binding directly to FtsN, although it remains formally possible that an E. coli divisome protein bridges DidA and FtsN in the two-hybrid analysis. SidA interacts with FtsW and FtsN in the bacterial two-hybrid system, and the lethality of overproducing SidA can be suppressed by mutations in either FtsW or FtsI [13]. Although DidA and SidA bind different proteins, these two inhibitors likely inhibit division in similar ways as two mutations in ftsW, and one in ftsI, can suppress the effects of overproducing either SidA or DidA.
FtsW, FtsI, and FtsN are among the last essential proteins recruited to the cytokinetic ring. These proteins physically interact with each other and likely form a subcomplex within the divisome that drives the synthesis and remodeling of the septal cell wall [2],[37]–[39]. Although its precise biochemical function is unknown, FtsW somehow contributes to septal cell wall synthesis, as does FtsI, which harbors peptidoglycan transpeptidase activity [40],[41]. The function of FtsN is also unclear, although in Caulobacter its essential activity is located within a periplasmic linker domain [19]. In both Caulobacter and E. coli, FtsN recruits proteins involved in cell wall remodeling to the division site [42]–[46], and E. coli FtsN has been suggested to stimulate the transpeptidase activity of PBP1B and could act similarly on FtsI [47].
How do single mutations in FtsW and FtsI prevent the inhibition of cell division by both SidA and DidA? One possibility is that these mutations reduce the affinities of SidA and DidA for their division protein targets. However, SidA binding to FtsW was unaffected by the A246T mutation and DidA binds FtsN, not FtsW or FtsI, in our bacterial two-hybrid system. Another possibility is that SidA and DidA block the recruitment of even later arriving proteins. As noted, FtsN may help recruit cell wall remodeling factors such as the peptidase DipM and the peptidoglycan amidase AmiC [43],[45]. Although the genes encoding such proteins are individually dispensable, it is formally possible that SidA and DidA disrupt the recruitment of multiple peptidoglycan remodeling factors, thereby preventing division. However, given that the inhibitory activity of both SidA and DidA can be suppressed by mutations in FtsW and FtsI, this model seems unlikely.
Instead, we favor a model in which the FtsW/FtsI/FtsN subcomplex exists in two states: an inactive state that is promoted by SidA or DidA, and an active state that drives septal peptidoglycan synthesis and cytokinesis (Figure 8A). We propose that the mutations that suppress both SidA and DidA, such as FtsW(A246T), may lock FtsW/FtsI/FtsN in the active state allowing cells to bypass the block in division normally caused by an accumulation of these inhibitors. On their own, these suppressor mutations cause cells to initiate division hyperactively. In support of this model, cells with the suppressing mutations were reproducibly shorter than wild-type cells (Figure 6A and 6B), likely because they divide at a slightly earlier stage of the cell cycle. Additionally, cells producing FtsW(A246T) or both FtsW(F145L, A246T) and FtsI(I45V) were sensitive to cephalexin, a cell wall synthesis inhibitor, and exhibited compromised cell envelope integrity. Importantly, ΔsidAΔdidA cells did not exhibit increased sensitivity to cephalexin, further supporting the notion that these mutations in FtsW and FtsI do not simply prevent SidA and DidA binding, but rather increase a cell wall synthesis activity.
Taken together, our results suggest that the DNA damage-induced division inhibitors in Caulobacter target the FtsW/FtsI/FtsN subcomplex to block constriction of the division machinery and cell envelope. Precisely how SidA and DidA block division is not yet clear, in part because the execution of cytokinesis remains poorly characterized at a molecular level. The synthesis of septal cell wall material could provide the force and directionality for cellular constriction, with FtsZ required mainly for mid-cell positioning of division proteins. This model is supported by recent data showing that FtsZ often dissociates from the divisome before compartmentalization occurs, indicating that cell wall synthesis may provide the constrictive force for cell division [48]. In such a case, SidA and DidA could prevent division by blocking a critical or rate-limiting peptidoglycan modifying activity of the FtsW/FtsI/FtsN subcomplex. As noted, the suppressor mutants in ftsW such as A246T that bypass both SidA and DidA are, on their own, prone to disruption of cell envelope integrity. Their sensitivity to cephalexin could result from certain cell wall synthesis or remodeling activities continuing without concurrent activation of the FtsI transpeptidase domain. As an alternative to this cell wall-centric model for cytokinesis, GTP hydrolysis by the FtsZ ring may provide the energy for, and directionality of, constriction, effectively pulling the rest of the cytokinetic ring along with it [49]. Assembly or activity of the FtsW/FtsI/FtsN subcomplex could somehow trigger FtsZ constriction, and the inhibitors SidA and DidA may block this step of division. Finally, it is possible that Z-ring constriction and septum synthesis combine to drive cytokinesis. As FtsW, FtsI, and FtsN are transmembrane proteins with cytoplasmic and periplasmic domains, they could coordinate the Z-ring and nascent septum, with SidA and DidA disrupting this coordination. Distinguishing between these various models for cytokinesis and elucidating the precise mechanisms of action for SidA and DidA will ultimately require more detailed studies of the FtsW/I/N subcomplex; the mutants identified here, such as FtsW(A246T), may prove particularly useful in these efforts.
Our results (i) reveal an SOS-independent mechanism for inhibiting cell division in Caulobacter and (ii) highlight the FtsW/FtsI/FtsN subcomplex as an important regulatory node in the control of cell division. Following certain types of DNA damage, DidA and SidA appear to function together to prevent inappropriate cell divisions (Figure 8). Such redundancy may afford cells with a fail-safe survival mechanism. In addition, SidA and DidA are differentially induced following different types of DNA damage, providing independent routes to the inhibition of cell division under different conditions. Also, we note that although cells lacking both sidA and didA divide prematurely during DNA damage, many still filament to some degree, suggesting that yet other mechanisms of division inhibition exist in Caulobacter. Finally, we note that DidA is the latest in a growing class of small, stress-induced membrane proteins that play critical regulatory roles [50],[51]. These proteins are often missed or incorrectly annotated in genome sequences, but many, like SidA and DidA, clearly play critical roles in regulating cellular processes, including cell division.
Strains and plasmids used in this study are listed in Table S1 with construction details and growth conditions provided in Text S1.
Synchronous Caulobacter populations were obtained by centrifugation over a Percoll density gradient as previously described [52]. Following synchronization of the ftsW(A246T) strain, we noticed that 21% of cells (two biological replicates) were unable to form microcolonies on plain PYE agarose pads compared to 2% for wild-type cells. Because of this sensitivity to the synchronization procedure, ftsW(A246T) cells and other suppressors were imaged by time-lapse microscopy following growth in mixed cultures.
RNA expression profiling was done as described [53]. Expression experiments were performed in duplicate and the results for each gene were averaged.
Samples for immunoblots were normalized in sample buffer to 0.5 OD600/50 µl, resolved on 12% sodium dodecyl sulfate-polyacrylamide gels and transferred to polyvinylidene difluoride transfer membrane (Pierce). Membranes were probed with polyclonal rabbit α-CtrA, α–DivL, α–LacZ (Rockland Scientific), and α-GFP (Invitrogen) at a 1∶5,000 dilution and monoclonal mouse α-FLAG (Sigma) at a 1∶3,000 dilution. Secondary HRP-conjugated α-rabbit (Pierce) or α-mouse (Pierce) were used at a 1∶5,000 dilution. Blots were visualized by chemiluminescence; raw black-and-white images were inverted for display. Biochemical fractionation was performed as described [13].
All phase contrast images were acquired on a Zeiss Observer Z1 microscope with a 100×/1.4 oil immersion objective and an LED-based Colibri illumination system. For additional information on image analysis and time-lapse microscopy, see Text S1.
Two-hybrid complementation assays were performed essentially as described [16]. BTH101 cells harboring plasmids with the T25 and T18 fusion constructs were grown to single colonies on LB agar plates and restruck or spotted on MacConkey agar plates supplemented with maltose for imaging.
The ftsN mutagenesis PCR reaction contained 21 µl 3M Betaine, 1 µl DMSO, 5 µl 10× Taq buffer (Invitrogen), 1.5 µl 50 mM MgCl2, 4 µl dNTPs, 0.2 µl primers, 50 ng genomic DNA, 2 µl mutagenesis buffer (100 mM dCTP, 100 mM dTTP, 50 mM MgCl2, 500 mM MnCl2), 0.3 µl Taq polymerase (Invitrogen), and water to 50 µl. The PCR reaction was incubated at 95°C for 5 minutes followed by 35 cycles of 95°C for 1 minute, 58°C for 1 minute, and 72°C for 3 minutes with a final extension of 72°C for 10 minutes. The mutant ftsN library was then cloned into a medium-copy plasmid downstream of the xylose-inducible promoter.
An ftsN depletion strain harboring a low-copy plasmid expressing 3×M2-didA from Plac was transformed with a medium-copy plasmid expressing the mutant ftsN library from Pxyl and grown on plates containing oxytetracycline, kanamycin, and 75 or 100 µM IPTG. The medium-copy kanamycin-resistant plasmids from suppressor colonies were isolated and retested in a clean ftsN depletion background for their ability to suppress 3×M2-didA overexpression from the IPTG-inducible low-copy plasmid. ftsN mutations in suppressor plasmids were identified by Sanger sequencing.
Wild-type cells were transformed with a Pvan:3×M2-didA overproduction plasmid and plated on PYE agar in absence of vanillate to allow colony formation. Single colonies were grown overnight in PYE and plated on PYE agar supplemented with vanillate at roughly 2×106 colony forming units per 10 cm plate. Rare colonies were grown overnight in PYE supplemented with vanillate and samples were taken for immunoblots, plasmid preparations, and archiving. To isolate chromosomal suppressor mutations and eliminate mutations arising in the 3×M2-didA overproduction plasmid, we screened for colonies that met two criteria. (1) We used immunoblotting to check that 3×M2-DidA production in each suppressor strain was similar to that seen in wild-type cells transformed with the same plasmid and grown in vanillate for 1.5 h. (2) Plasmids from the suppressor strains were transformed into wild-type cells and plated on PYE agar supplemented with or without vanillate. The presence of thousands of colonies on plain plates and few colonies on vanillate indicated a functional plasmid. The mutation in the ftsW(A246T) suppressor strain was identified by whole genome resequencing.
Cells expressing lacZ from PdidA at the hfaB locus in a ΔdidA background were mutagenized with the EZ-Tn5 transposome (Epicentre) and grown on plates containing kanamycin and 20 µg/ml X-gal. Colonies appearing white were isolated and tested for low or undetectable levels of full-length LacZ by western blot with α-LacZ antibodies. Transposon insertion mutations were identified as described (Epicentre, TSM08KR protocol) by rescue cloning with pir-116 electrocompetent E. coli cells (Epicentre).
ChIP was performed as detailed in Text S1. Quantitative PCR was performed with the dye SYBR Green (Roche) on a Lightcycler 480 system (Roche). Each reaction contained 5 µl SYBR Green Master, 1 µl DNA (diluted 1∶500 for pre-ChIP input DNA, and 1∶20 for post-ChIP output DNA), 0.5 µl primer mix at 10 µM, and 3.5 µl nuclease-free water. Primers amplifying a product within the ruvA coding sequence were used as a control. Cycle threshold values were calculated using the Lightcycler 480 software and converted to DNA concentrations based on a standard curve generated from 2-fold dilutions of Caulobacter genomic DNA. Fold enrichment values were calculated as ([PdidA−output]/[ruvA−output])/([PdidA−input]/[ruvA−input]). Error bars in Figure 7F were generated from technical triplicates, and the experiment shown is representative of biological duplicates.
|
10.1371/journal.pgen.1004515 | A System for Genome-Wide Histone Variant Dynamics In ES Cells Reveals Dynamic MacroH2A2 Replacement at Promoters | Dynamic exchange of a subset of nucleosomes in vivo plays important roles in epigenetic inheritance of chromatin states, chromatin insulator function, chromosome folding, and the maintenance of the pluripotent state of embryonic stem cells. Here, we extend a pulse-chase strategy for carrying out genome-wide measurements of histone dynamics to several histone variants in murine embryonic stem cells and somatic tissues, recapitulating expected characteristics of the well characterized H3.3 histone variant. We extended this system to the less-studied MacroH2A2 variant, commonly described as a “repressive” histone variant whose accumulation in chromatin is thought to fix the epigenetic state of differentiated cells. Unexpectedly, we found that while large intergenic blocks of MacroH2A2 were stably associated with the genome, promoter-associated peaks of MacroH2A2 exhibited relatively rapid exchange dynamics in ES cells, particularly at highly-transcribed genes. Upon differentiation to embryonic fibroblasts, MacroH2A2 was gained primarily in additional long, stably associated blocks across gene-poor regions, while overall turnover at promoters was greatly dampened. Our results reveal unanticipated dynamic behavior of the MacroH2A2 variant in pluripotent cells, and provide a resource for future studies of tissue-specific histone dynamics in vivo.
| The ability of cells to remember the correct cell fate is at least partly dependent on how the genome is packaged. Embryonic stem (ES) cells, which have the ability to become any cell type in the body, are a particularly well-studied system for understanding how the packaging of the genome – chromatin – controls cell state. One of the more curious aspects of ES cell chromatin is its “hyperdynamic” nature, as the histone proteins that comprise chromatin have been reported to exchange rapidly on and off the DNA in these cells. Here, we report a pulse chase system for studying histone dynamics in mouse ES cells, and report on the dynamics of two histone variants, H3.3 and MacroH2A2. Notably, MacroH2A2 is highly dynamic in ES cells, with rapid exchange occurring over gene promoters, alongside much more stably-bound domains that cover large blocks of the genome. Upon differentiation to fibroblasts MacroH2A2 becomes much more stably-bound to the genome, consistent with the idea that this histone variant plays a role in “locking down” repressed regions the genome. These results provide further evidence for a key role of histone dynamics in control of cell state inheritance.
| All genomic transactions in eukaryotes occur in the context of chromatin. While histones are generally among the most stably-associated DNA-binding proteins known [1], a subset of histones exhibit dynamic replication-independent exchange with the soluble pool of nucleoplasmic histones [2]–[4]. Dynamic histone exchange is intimately linked to a variety of key aspects of chromatin biology.
In all eukaryotes studied, histone H3 exchange is most rapid at promoters [5]–[12], and is generally slowest over heterochromatic regions. In addition, H3 exchange is rapid at boundary elements that block the spread of heterochromatin [5], [7], raising the possibility that rapid histone exchange could function mechanistically to erase laterally spreading chromatin states. These correlations, in which histone exchange is slow over epigenetically-heritable heterochromatin domains but is rapid at boundary elements, raise the question of how histone dynamics contribute to epigenetic inheritance. Interestingly, H3/H4 tetramers carrying the H3.3 variant “split” during replication to a greater extent than do H3.1-containing tetramers [13], consistent with the hypothesis that dynamic regions of chromatin could potentially self-perpetuate through replication [2]. In addition, the rapid histone turnover observed at promoter regions of actively transcribed genes suggests that histone turnover may have an important role in gene regulation, as higher histone turnover rates could provide greater access of regulatory proteins to specific DNA elements. Yet much remains to be learned about the mechanistic basis for, and the biological consequences of, dynamic chromatin states.
Embryonic stem (ES) cells are a key model for mammalian pluripotency and cell state inheritance. ES cells are characterized by unusual chromatin packaging [14], and a wide variety of chromatin regulators have been implicated in control of pluripotency and differentiation [15]–[19]. One curious feature of ES cell chromatin is its “hyperdynamic” state—photobleaching experiments show that many histone variants exchange more rapidly in ES cells than in differentiated cell types [20]. This hyperdynamic state has been proposed to maintain the ES cell genome accessible as a relatively permissive ground state that becomes “locked down” during the process of lineage commitment and subsequent differentiation. Understanding histone exchange dynamics in ES cells, and during differentiation, is therefore of great interest for understanding the roles for chromatin in cell state inheritance.
The histone variant MacroH2A plays a key role in cell state stabilization in mammals. Mammals encode three MacroH2A variants, MacroH2A1.1 and MacroH2A1.2, which are alternatively spliced isoforms of a single gene, and the distinct gene product MacroH2A2. All three MacroH2A variants are distinguished by the presence of the unusual “Macro” domain fused to their relatively well-conserved H2A cores. It has been suggested that MacroH2A plays a role in fixing the epigenetic state of differentiated cells (reviewed in [21]). Support for this notion comes from observations that MacroH2A deposition increases with cellular age and senescence [22], [23], and that epigenetic reprogramming via somatic cell nuclear transfer is accompanied by an active removal of MacroH2A1 from the donor chromatin upon transfer into the ooplasm [24]. More recent studies have indicated that MacroH2A depletion from somatic cells increases their propensity for undergoing epigenetic reprogramming [25]–[28] —in several of these studies, depletion of either MacroH2A1 or MacroH2A2 enhances reprogramming, with depletion of both having an additive effect. These studies suggest that removal of MacroH2A from the somatic genome may be prerequisite for acquisition of pluripotency during epigenetic reprogramming.
MacroH2A may further contribute to fixing the epigenetic state of differentiated female cells due to its accumulation on the inactive X chromosome (Xi) [29]. However, association of MacroH2A1 with the Xi appears to occur after the random inactivation of the X chromosome (XCI) [30], and in conditional Xist deletions gene silencing is maintained despite the loss of MacroH2A1 on the Xi [31]. Nonetheless, while MacroH2A1 appears to be dispensable for XCI, removal of this variant from the Xi could still potentially represent a barrier to epigenetic reprogramming of a differentiated, post-XCI somatic cell to the pre-XCI ground state of pluripotency.
Despite the general characterization of MacroH2A as being a “repressive” histone variant, there are numerous examples where Macro incorporation is associated with increased gene expression, particularly during early lineage specification after embryoid body formation from ES cells [32], and more recently in embryonic fibroblasts where MacroH2A1 is present at high levels in the active Thy1 gene, but nearly completely absent when this gene is silent in pluripotent ES cells [27]. Determining the dynamics of MacroH2A turnover in both pluripotent ES cells and somatic cells is therefore of paramount interest for gaining an in-depth understanding of the epigenetic processes underlying cellular reprogramming.
Three methods are currently used to study histone dynamics [33]. First, the original discovery that the H3.3 variant marks sites of replication-independent histone exchange [3], [4] has enabled many labs to infer histone dynamics simply from steady-state H3.3 localization patterns [6], [7], [9]–[11]. Second, genetically encoded “pulse-chase” systems have been utilized in which an epitope-tagged histone molecule is induced, and mapping of the epitope tag at various times after induction provides a detailed kinetic view of histone exchange dynamics [5], [8], . Finally, a metabolic labeling strategy termed “CATCH-IT” enables kinetic analysis of overall chromatin dynamics [37].
Here, we extend the approach of inducible expression of epitope-tagged histone variants to study chromatin dynamics in murine embryonic stem cells. We generated ES lines carrying doxycycline (“Dox”)-inducible HA-tagged versions of several histone variants, including H3.3 and MacroH2A2. These cells allowed us to monitor the rate of incorporation of HA-tagged variants by ChIP-Seq at varying times following Dox induction. For the well-studied H3.3 variant, we validate our method by recapitulating known aspects of H3.3 localization and dynamics. We also characterized the dynamics of the understudied MacroH2A2 variant in detail in ES cells and in their embryonic fibroblast (MEF) derivatives. MacroH2A2 exhibited broad, likely replication-coupled, incorporation throughout large stretches of the ES cell genome, along with unexpectedly rapid turnover behavior at highly-expressed promoters. In contrast, MacroH2A2 in more differentiated MEFs was additionally associated with a subset of gene-poor genomic loci, and its exchange at promoters slowed considerably. These results reveal surprising aspects of MacroH2A2 localization and dynamics and suggest that the view of MacroH2A2 as simply an indicator and/or mediator of repressed chromatin states is not accurate. Moreover, these studies establish a model system for investigation of histone variant dynamics in tissue culture systems as well as in complex organ systems in vivo.
In order to assay genome-wide histone variant dynamics in embryonic stem cells and cell types derived from them, we generated ES cells based on the murine KH2 ES cell line [38], which harbors a modified reverse tetracycline transactivator (M2rtTA) targeted to the ROSA26 locus and an FRT recombination site targeted into safe-haven chromatin downstream of the Type I Collagen (Col1A1) locus. Introduction of a donor plasmid carrying another FRT recombination site along with HA-tagged cDNA sequences encoding the histone variants of interest under transcriptional control of the tetracycline operator (TetO), along with an additional plasmid encoding the FLP recombinase, allows for site-specific integration of the tetracycline-inducible HA-tagged histone cassette into the genome (Figure 1A). Subsequent addition of the tetracycline analog doxycycline (“Dox”) to these ES cell clones or mice derived from them results in induction of the tagged histone variant (Figure 1A, bottom panel). Cell lines were generated and validated for several different histone variants, including MacroH2A2 (hereafter called Macro in some contexts), and H3.3 (Figure S1A).
For all cell lines analyzed, no expression of tagged histones was detected in the absence of Dox by Western blotting or immunofluorescence staining with an HA antibody. Robust activation of the tagged proteins was detectable within 2–3 hours of Dox addition (Figure 1B, Figure S1B). Several controls show that ectopic expression of tagged histone variants did not significantly perturb ES cell pluripotency. First, even after 12 hours of induction, ectopically expressed histones were far less abundant than the endogenous proteins levels (Figure S1C). Second, after 72 hours of overexpression, ES colonies maintained their pluripotent state as assessed by cell morphology, alkaline phosphatase (AP) staining, and expression of pluripotency markers such as Oct4, Sox2 and Nanog (Figure S2). The only exception was H3.3, where 72 (but not 24) hours of ectopic expression resulted in a slightly reduced proliferation rate, but did not compromise pluripotency based on Oct4 or AP staining (not shown). Finally, as shown below, mapping of total MacroH2A2 (for which a high quality commercial antibody exists) both before and after HA-Macro induction yielded nearly identical results, demonstrating that ectopic expression did not drive nonphysiological incorporation of this histone variant into ectopic sites throughout genome.
To validate our system, we first sought to determine whether a pulse-chase experiment is consistent with steady state mapping of H3.3 localization [4], [6], [7], [9], and what additional information it provides. ES cells carrying doxycycline-inducible HA-H3.3 were treated with Dox, and harvested after 0, 3, or 6 hours. HA-H3.3-containing chromatin was mapped genome-wide by chromatin immunoprecipitation followed by Illumina deep sequencing (ChIP-Seq). Sequencing reads were mapped back to the genome. Importantly, HA mapping at t = 0 (no doxycycline) did not show enrichment over specific loci but rather genome-wide nonspecific background, demonstrating the specificity of the anti-HA antibody (see below). Because H3.3 replacement is strongly associated with the 5′ ends of genes [6], [7], we aligned all annotated genes by their transcription start sites (TSSs), and averaged all mapped reads at each position relative to the TSS (Figure S3). Consistent with studies in flies and murine ES cells [6], [7], [9], we find that H3.3 is localized to two peaks surrounding the TSS, and that H3.3 levels correlate with the mRNA abundance of the associated gene. We also confirmed that the rapid H3.3 dynamics observed at Polycomb-bound regulatory elements in flies [7] are also present in the mouse embryonic stem cell genome at regions occupied by polycomb proteins Rnf2 and Suz12 (Figure S3E). Our results therefore recapitulate major known aspects of histone H3.3 dynamics. As H3.3 replacement has been extensively studied, we therefore turned to the understudied MacroH2A.2 variant.
We next extended our studies to a histone variant with unknown dynamic properties, MacroH2A2. Because MacroH2A2 localization in ES cells has not been characterized, we first carried out genome wide mapping of MacroH2A2 in murine ES cells using a commercially available antibody (Figure 2, Tables S1, S2). MacroH2A2 was broadly localized to large (megabase-scale) blocks across the mouse genome, where it colocalized with regions of high gene density (Figure 2A, Figure S4, Table S3) —the correlation between average MacroH2A2 enrichment and gene density was 0.46 for 100 kb windows, and rose to 0.59 when considering 1 MB windows of the genome (Figure 2B). In addition to broad localization over gene-rich regions, we noted that MacroH2A2 exhibited a tight (∼500 bp) peak on average over promoters (Figure 2C). Counterintuitively, genes lacking MacroH2A2 were generally poorly expressed (Figure 2D and Figure S4B, see Cluster 3), stemming largely from the absence of MacroH2A2 at repressed gene families such as those encoding olfactory receptors or zinc finger transcription factors. Interestingly, among genes associated with promoter MacroH2A2, tighter localization was correlated with higher expression levels (Figure S4B, Clusters 1 and 2). Consistent with the surprising correlation between MacroH2A2 localization and active promoters, we found a moderately positive correlation between our MacroH2A2 dataset and H2A.Z localization [39] in ES cells (Figure S5).
Confidence in these surprising observations comes from four lines of evidence. First, localization datasets obtained before and after HA-MacroH2A2 induction were highly-correlated (Figure 2C). Second, anti-HA ChIP-Seq in uninduced HA-Macro cells yielded a nearly flat genome-wide background (Figure 2A and Figure S4A, top panel, Figure S6, left panel). Third, MacroH2A2 localization obtained using the MacroH2A2 antibody was very highly correlated with the localization pattern observed using anti-HA ChIP-Seq from cells expressing HA-MacroH2A2 (Figure 2A and Figure S6), but not HA-H3.3 (Figure S3). Finally, MacroH2A2 localization patterns were strongly correlated, but not identical, between ES cells and MEFs (see below). Thus, we find MacroH2A2 localizes to large blocks of gene-rich chromatin in ES cells, and within these blocks exhibits strong promoter localization at expressed genes.
We next carried out genome wide mapping of HA-MacroH2A2 at 3 time points (3, 6, and 12 hours) after Dox induction. Reads were mapped back to the genome and genes were aligned by TSS as above. HA mapping in the no Dox control revealed a primarily flat genomic background (Figures 2A, 3A), with trace levels of promoter localization likely resulting from low levels of leaky expression of HA-MacroH2A2 (Figure S6). Data from 3, 6, and 12 hours after HA-Macro induction was strongly correlated with endogenous MacroH2A2 localization (Figure 3A, Figure S6). The strong correlation between all 3 time points and the steady-state localization is to be expected from the fact that ES cells are rapidly cycling, so even at 3 hours of induction a substantial subpopulation of cells will have gone through S phase and carried out any replication-dependent MacroH2A2 incorporation.
In yeast, cell cycle arrest can be used to explicitly assay replication-independent histone replacement dynamics [5], [8]. However, this is impractical in ES cells, and our data come from asynchronously cycling cells. Nonetheless, such data can be used to study histone turnover. Two considerations, one conceptual and the other empirical, will aid in understanding how pulse-chase data obtained from cycling cells can be used to infer turnover dynamics (see Figure S7). Conceptually, we expect that loci exhibiting replication-coupled histone deposition (or slow replication-independent deposition) will gradually accumulate epitope-tagged histone variants over a time course of induction (Figure S7A). In contrast, because replication-independent replacement will initially occur in a greater fraction of cells than the subset of cells that are actively transiting S phase, such loci will exhibit more rapid accumulation of tagged histone. Given that genome-wide measurement methods typically normalize for sequencing depth (with the underlying assumption/hypothesis being equivalent total amounts of material between samples), the end result of this is that loci exhibiting rapid turnover will exhibit high levels of epitope tag enrichment early in a time course, but later in the time course this normalized relative enrichment will decrease as the bulk of cells transit S phase and replication-coupled deposition results in a greater total amount of epitope tag incorporated into the genome. In other words, relative enrichment of the rapidly exchanging population is high at early time points before population-wide assembly of HA-histone into the slower subpopulations, whereas at later time points normalization relative to the extensive HA-histone in cold domains results in a diminishing peak at “hot” loci (Figure S7B). Importantly, the assessment of relatively hot and cold loci is robust to normalization methods (Figure S7B, Methods). This predicted behavior is exactly what we have previously observed [5] empirically in yeast—here, replication-independent H3 turnover was directly measured in G1-arrested yeast. A parallel experiment was carried out using asynchronous cells, and those loci shown to exhibit rapid replication-independent turnover exhibited precisely the above-predicted behavior—rapid enrichment of tagged H3, followed by diminishing tag enrichment as the bulk of the genome was assembled into tagged H3 via replication-coupled assembly.
Consistent with the above considerations, in addition to the genome-wide HA incorporation observed at all 3 time points, we also observe extensive locus-specific variation in HA-Macro dynamics (Figure 3A, red and green-bordered boxes identify regions of rapid and slow HA incorporation, respectively). Regions exhibiting high levels of HA at 3 hours relative to 12 hours were inferred to be “hot” (Figure S7, [5]), and typically occurred in highly delimited peaks associated with promoters (see below), whereas cold regions generally covered broad chromosomal stretches, often in intergenic regions (Figures 3B–D, S8, Table S2). These trends can also be seen in detail when focusing on promoter proximal Macro dynamics (Figure 4). On average, the TSS-proximal peak of MacroH2A2 diminished from 3 hours to 6 hours to 12 hours, consistent with rapid replication-independent replacement. This observation was reproduced in a second HA-Macro induction time course (Figure S9). In contrast, genes associated with broad domains of MacroH2A2 across their promoters (Figures 2B and 4A, Cluster 2) exhibited consistent HA-MacroH2A2 mapping patterns at all three time points, as would be expected if these broad domains were relatively stable and incorporated Macro either via slow replacement or only during replication.
These results are consistent with at least two populations of MacroH2A2-containing chromatin that can be distinguished by their dynamic behavior. We infer that the TSS-proximal MacroH2A2 that is enriched at early time points before diminishing in enrichment represents a rapidly exchanging population of Macro that is present at moderate steady state occupancy, while larger Macro domains undergo either slow turnover or replication-coupled assembly. These larger domains tend to be gene poor, often occurring over gene deserts (Figures 3C–D) but occasionally encompassing individual genes as well (Figure 4A, cluster 2).
To gain further insight into the population of dynamic promoter-proximal Macro, we sorted genes with tight promoter Macro peaks (Cluster 1) according to their relative inferred Macro dynamics (Figure 5A) —note that relative dynamic behavior is completely insensitive to whether data are normalized assuming equivalent levels of Macro, or taking increasing total Macro incorporation over time into account. Genes with rapidly exchanging MacroH2A2 were enriched for GO processes consistent with housekeeping functions such as “translation” or “metabolism” (not shown) that are generally highly expressed, suggesting a potential link to expression level. Indeed, we found a strong correlation (r = 0.46) between MacroH2A2 dynamics and mRNA abundance (Figure 5B), as poorly expressed genes were associated with more stable MacroH2A2 than were highly expressed genes (see Figures 5C–D for examples). This link between promoter Macro dynamics and mRNA abundance supports our hypothesis that a pattern of diminishing HA enrichment over our time course is diagnostic of rapid MacroH2A2 replacement. These results are also consistent with the rapid histone H3 dynamics at promoters observed in a variety of organisms.
What is the function of rapid MacroH2A2 replacement at highly-expressed promoters? Knockdown of MacroH2A2 resulted in extremely modest changes in global mRNA abundance (Table S4), likely reflecting compensatory gene regulation by MacroH2A1, which is present in ES cells at ∼10-fold higher abundance than is MacroH2A2. Nonetheless, mRNA abundance exhibited greater changes at genes associated with slow MacroH2A2 exchange dynamics than at genes with rapid MacroH2A2 replacement (Figure S10).
Given that several histone variants are “hyperdynamic” in ES cells [20] and that the Macro content in somatic cells is considered an epigenetic barrier for epigenetic reprogramming to pluripotency [24], [26], we sought to characterize the changes in MacroH2A2 dynamics between ES cells and mouse embryonic fibroblasts (MEFs). We generated transgenic mouse embryos by injecting TRE-HA-MacroH2A2 ES cells into blastocysts, derived MEFs from E12.5 chimeric embryos, then purified a homogenous population of TRE-HA-MacroH2A2 MEFs after selection against host blastocyst-derived cells. Importantly, HA-Macro protein induction dynamics were similar in ES cells and MEFs (Figure S11), enabling comparisons of Macro dynamics using this system. We first mapped MacroH2A2 in MEFs using an anti-MacroH2A2 antibody. As observed for ES cells, Macro localization patterns were strongly correlated before and after Dox induction (Figure S12), and were strongly correlated with HA mapping data from Dox-induced cells (r = 0.98, see below), providing strong evidence for antibody specificity.
Overall, MacroH2A2 patterns were similar (r = 0.67 using 100 kb bins) between ES cells and MEFs (Figure 6A, Figure S13, Tables S1–S2), supporting prior reports showing good correlations for MacroH2A1 localization between different cell types [40]. There was a general increase in MacroH2A2-enriched regions in MEFs relative to ES cells (Figure S13A), consistent with the fact that MacroH2A2 levels are higher in MEFs than in ES cells. Overall, while MacroH2A2 was generally maintained at gene-rich regions in MEFs as well as ES cells (Figure 6A, S13), we identified a large number of additional regions that gained Macro in MEFs relative to ES cells. Interestingly, MEF-specific Macro domains typically occurred in gene-poor chromosomal regions (Figure S13). In terms of gene categories associated with the sparse genes found in these gene-poor regions, MEFs gained MacroH2A2 at a broad set of genes involved in alternative differentiation programs including neural, leukocyte, muscle, and spermatogenesis programs (Table S5). This broadly supports the idea that the more plastic pluripotent chromatin state becomes progressively restricted during differentiation, with unused genes in each variety of differentiated cell type becoming “locked down” via MacroH2A2 incorporation.
In addition to broad gains of MacroH2A2 over gene-poor regions, we observed widespread changes in Macro enrichment over promoters between MEFs and ES cells. The average peak of MacroH2A2 over promoters exhibited an apparent decrease in MEFs (Figure 6B), although given that genome-wide there is more MacroH2A2 signal distant from promoters in MEFs relative to ES cells, this loss is overestimated as a result of dataset normalization. Accounting for this possibility, we nonetheless noted extensive redistribution of promoter-localized MacroH2A2 between ES cells and MEFs (Figure 6C). Curiously, MacroH2A2 changes between ES cells and MEFs correlated poorly (r = 0.02) with gene expression changes between these cell types, although we did note that exceptionally unpregulated genes characteristic of fibroblasts such as collagen and extracellular matrix factors (Col1a1, Col5a1, Lox, Tgfb2, Fib1, etc.) generally lost MacroH2A2 at their promoters in MEFs (Tables S1–S2). Instead of correlating with gene expression changes, we found that loss of Macro in MEFs tended to occur at promoters exhibiting dynamic Macro turnover in ES cells (Figures 6C–D). In contrast, stably Macro-associated promoters in ES cells preferentially retained Macro in MEFs. Together, these results suggest that dynamic assembly and disassembly of MacroH2A2 at highly expressed promoters is a specific feature of ES cells that is lost upon differentiation. In other words, while ribosomal protein genes (Rpl8, Rpl32, etc.) are highly expressed in both ES cells and MEFs, in ES cells their promoters are associated with rapidly-exchanging Macro, whereas these promoters are depleted of Macro in MEFs.
To explicitly characterize Macro dynamics in MEFs, we carried out HA-Macro mapping at 3, 6, and 12 hours after Dox induction. As with ES cells, HA localization at all 3 time points was highly correlated (r = 0.98 for all three time points using 100 kb windows) with mapping data obtained using the anti-Macro antibody. In contrast to ES cells, however, inspection of genome browser tracks yielded many fewer instances of 3 hour HA peaks that diminished at 6 and 12 hours. More systematically, we found that the average TSS-proximal HA peak was nearly identical at all three time points (compare Figures 7A and B). Not only was the average promoter HA peak nearly identical at all three time points, but there was less variation from t = 3 to t = 12 in our MEF data than in our ES data (Figure 7C). Sorting genes by inferred turnover behavior in MEFs revealed a subtle correlation between promoter turnover kinetics and mRNA abundance in MEFs (Figure 7D), but this relationship was far less robust (r = 0.17 versus r = 0.46) than that observed in ES cells (Figure S14). Taken together, these data show that rapid MacroH2A2 turnover is a specific feature of ES cells, and that upon differentiation to MEFs Macro is lost from dynamic promoters but retained in larger blocks of stably-associated Macro.
Here, we extended a genetically encoded pulse-chase approach to use inducible epitope-tagged histone variants to study chromatin dynamics in mammalian cells. The inclusion of a temporal component in studying histone variant turnover yields information normally lost when analyzing static binding data, even for the well-studied H3.3 histone variant. Our use of non-transformed pluripotent embryonic stem cells and primary mouse embryonic fibroblasts demonstrate the broad utility of this approach in mammals.
We primarily focus here on the relatively unstudied MacroH2A2 variant. Overall, we observe extensive differences in the localization and dynamics of this variant between pluripotent ES cells and committed mouse embryonic fibroblasts. In ES cells, we observed widespread localization of MacroH2A2 across gene rich domains, along with a strong TSS-proximal peak of MacroH2A2. From our time course mapping studies, we infer that MacroH2A2 is rapidly replaced at promoters, and that this replacement is positively correlated with a gene's expression level. It is worth noting that H3.3 replacement is also rapid at promoters and correlates with mRNA abundance, indicating that in ES cells promoters exhibit rapid turnover of multiple histone variants.
Upon differentiation to embryonic fibroblasts, MacroH2A2 is broadly gained over gene poor domains, resulting in increased MacroH2A2 levels over genes associated with alternative differentiation programs such as neural or immune cell differentiation. Intriguingly, MacroH2A2 becomes far less dynamic in MEFs, and moreover MacroH2A2 is generally lost from those promoters where it is most dynamic in ES cells. Among other things, this observation argues that dynamic MacroH2A2 replacement inferred at highly-expressed genes in ES cells does not simply reflect nonspecific association of ectopically expressed histones with “open” promoters, as the highly expressed genes in MEFs exhibit far more subtle Macro dynamics than do the same genes in ES cells. Removal of the X chromosome from all key analyses (Figure S15) does not alter any of the conclusions regarding the change in Macro behavior between ES cells and MEFs, which is unsurprising as both cell types used in this study are male and thus data from the X chromosome reflects only the active X.
Together, these results are broadly consistent with the idea that pluripotent cells are characterized by “hyperdynamic” chromatin [20]. Interestingly, in contrast to the global hyperdynamic state observed by photobleaching for other histone variants, here we observe local, rather than global, dynamic MacroH2A2 behavior at a small fraction of loci—promoters of highly expressed genes. It will be interesting to identify factors contributing to ES-specific promoter MacroH2A2 dynamics in future studies.
Our findings that the dynamics of Macro turnover decrease as pluripotent ES cells become developmentally committed, and that stable MacroH2A2 becomes incorporated in gene poor regions and at genes associated with alternative cell fates in MEFs, have implications for the interpretation of several recent studies suggesting that the MacroH2A content of somatic cells acts as a barrier to epigenetic reprogramming of the genome to a pluripotent state. It is widely appreciated that Macro content increases during cellular differentiation and ageing, and studies employing somatic cell nuclear transfer (SCNT, or cloning), revealed that somatic MacroH2A1 is actively removed from the genome prior to the acquisition of pluripotency [24]. These observations, coupled with the accumulation of MacroH2A on the Xi during the process of X chromosome inactivation in female cells, suggest that removal of MacroH2A from the somatic genome may facilitate, or even be a prerequisite for, reprogramming to pluripotency.
Indeed, a recent study found that depletion of MacroH2A1 and 2 from somatic cells prior to initiation of epigenetic reprogramming via the ectopic expression of Oct4, Sox2, Klf4, and c-Myc (the Yamanaka factors—[41]) greatly improved reprogramming efficiency [27]. This study implicated repression of pluripotency-associated genes (Oct4, Sox2) with high MacroH2A1 content in somatic cells as the epigenetic barrier, such that removal of MacroH2A from pluripotency-associated promoters might allow for the reprogramming factors to more readily activate these genes. While this may be a contributing factor, in general MacroH2A content does not strongly predict gene repression. For example, in MEFs MacroH2A1 is highly enriched at the active Thy1 gene, but in ES and iPS cells, where Thy1 is silent, MacroH2A1 is nearly completely absent [27]. Indeed, in ES cells we find that MacroH2A2 is associated with active promoters (Figure 2), further arguing against a simple model for a universally repressive function of MacroH2A. Instead, we speculate that stable association of MacroH2A (Figure S10), rather than average MacroH2A occupancy per se, is more likely to play a role in gene repression. Consistent with this idea, we observe Macro enrichment over Sox2 in both ES cells and in MEFs, but in ES cells this gene is marked by rapid Macro replacement whereas Macro association is much more stable in MEFs (not shown).
Our findings that (1) dynamic incorporation of MacroH2A2 in gene-rich regions is correlated with highly active promoters, and that (2) stable MacroH2A2 incorporation in gene-poor regions (harboring genes associated with alternative cell fates) in MEFs is correlated with gene silencing, suggests that Macro removal during reprogramming may be most critical at these stable loci for re-establishing the “permissive” chromatin state characteristic of pluripotent cells.
To date, the majority of studies on histone dynamics have been carried out in cell culture systems. However, it will be of great interest to begin understanding the tissue-specific differences in chromatin dynamics in vivo, both under control conditions and in response to environmental perturbations. Thus, we generated a inducible histone variant mouse strain after blastocyst injection of the TRE-HA-H3.3 ES cell line and successful germline transmission of the R26-M2rtTA and TRE-HA-H3.3 alleles (Figure S16). Administration of 2 mg/mL doxycycline in the drinking water of TRE-HA-H3.3 mice resulted in HA-H3.3 induction in liver nuclear extracts (Figure S16C). These animals will therefore provide a unique and exciting resource for characterization of histone dynamics in different tissues and cell types, and provide a proof of principle for the application of our approach in vivo.
All procedures involving mice were reviewed and approved by the Institutional Animal Care and Use Committee of the University of Pennsylvania (Animal Welfare Assurance Reference Number #A3079-01, approved protocol #803415 granted to Dr. Lengner) and were in accordance with the guidelines set forth in the Guide for the Care and Use of Laboratory Animals of the National Research Council of the National Institutes of Health.
cDNA of various histone variants ((H2a-MMM1013-98478233; H2Az-MMM1013-9498090; MacroH2a2-MMM1013-9201250; H3.3-MMM1013-98478016, H1o-MMM1013-65296, Open Biosystems & human H3.1-Kind gift of Eric Campeau)) were initially subcloned in-frame with the HA-tag, then were cloned into the unique EcoRI restriction site of the pBS31 vector containing a PGK promoter followed by an ATG start codon and an FRT recombination site, followed by a splice acceptor-double polyA cassette, the tetracycline operator with a minimal CMV promoter, the unique EcoRI site, and an SV40 polyadenylation signal. The pBS31 vector containing the histone/histone variant cDNA was then electroporated along with a Flpe recombinase-expressing vector into KH2 embryonic stem cells harboring the modified reverse tetracycline transactivator (M2rtTA) targeted to and under transcriptional control of the ROSA26 locus, as well as an FRT-flanked PGK-neomycinR cassette followed by a promoterless, ATG-less hygromycinR cassette targeted downstream of the Collagen1a1 locus [38]. Selection for hygromycin resistance upon flip-in yielded numerous colonies which were verified for proper site-specific recombination at the Coll1a1 locus by digestion of genomic DNA and Southern blotting with a 3′ internal probe, yielding a 6.2 kb wildtype band, a 6.7 kb band for the FRT-containing knock-in allele, and a 4.1 kb band for the successfully flipped-in inducible allele. Together, the components of this system enable tetracycline induction of the epitope-tagged histone variant of choice in embryonic stem cells from a genomically-integrated construct.
Activation of the TetOn HA-tagged histone expression was carried out by addition of 2 µg/mL doxycycline hyclate (Sigma D9891) to the culture media. Cells were collected at different induction time points and induction of HA tagged histone variants in ES cells was assayed via Western blot.
ES cell cultures were fixed in 4% paraformaldehyde for 5 minutes prior to staining for pluripotency markers alkaline phosphatase and Oct4. Alkaline phosphatase was detected by enzymatic reaction using a Vector Red substrate kit (Vector Labs). Immunofluorescence staining for Oct4 was carried out by first permeabilizing and blocking in 5% FBS, 0.1% Triton-X 100 for 15 minutes, then incubating with an anti-Oct4 primary antibody at 1∶100 for 1 hr at room temperature (Rabbit polyclonal H-134, Santa Cruz Biotech). After 3 washes with PBS, cells were incubated with an anti-rabbit secondary antibody labeled with Cy3, washed, stained with DAPI for total DNA, and imaged.
HA-MacroH2A2 or HA-H3.3-inducible ES cells were injected into BDF2 blastocysts and transplanted into pseudopregnant recipient females. For HA-MacroH2A2 MEF isolation, pregnant females were euthanized at E12.5, embryos were dissected followed by removal of internal organs. Embryos were then minced in the presence of 0.25% Trypsin-EDTA and incubated at 37°C for 20 minutes. MEF medium was then added and cell suspension was titrated followed by plating cells onto two 15 cm culture dishes per embryo. Cells were cultured for 12 hours at 37°C, 3% CO2, and 5% O2 after which puromycin was added to the MEF culture media to select against host blastocyst-derived cells (by virtue of a constitutively active puromycin resistance cassette targeted to the ROSA26 locus along with the M2rtTA). After 48 hours of puromycin selection, homogenous populations of HA-MacroH2A2 MEFs were trypsinized and frozen at passage 1. ChIP-Seq experiments on MEFs were carried out after thawing and one additional passage (i.e., p2 MEFs). For generation of HA-H3.3 mice, blastocyst injection was performed as above, but embryos were carried to term. High contribution chimeras (>95% by coat color) were backcrossed to Bl/6 mice to establish an inducible HA-H3.3 mouse colony.
ES cells were grown in standard ES media containing Lif (ES Gro, Millipore) on mitotically inactivated feeder MEFs until approximately 80% confluence. ES cells were then pre-plated on gelatin and incubated for 45 min to deplete feeder MEFs by virtue of their faster adherence than ES cells (roughly 3 hours). ES cells were then split onto three gelatinized plates each of which was induced at different time points by the addition of final 2 µg/mL doxycycline hyclate (Sigma). A similar procedure was used for induction of MEFs at passage 2. All time points were crosslinked with formaldehyde to a final concentration of 1% for 10 minutes, and were quenched with 125 mM glycine. Crosslinked cells were resuspended in 270 µl SDS-Lysis Buffer (1% SDS, 10 mM EDTA and 50 mM Tris-Cl, pH 8.1) including protease inhibitor complex (Sigma) and PMSF (Sigma), and chromatin was sonicated in Bioruptor (UCD-200) to an average size of 150–400 base pairs. 70 µg of chromatin of each time point was immunoprecipitated either with HA antibody (Abcam) or MacroH2A2 antibody (Abcam). Eluted ChIP materials were PCI (Phenol-Chloroform-Isoamylalcohol) extracted, RNAse (Qiagen) and CIP (NEB) treated.
ChIP material was then gel-purified and DNA fragments were blunt-ended and phosphorylated with the End-it-Repair kit (EPICENTRE). Illumina genome sequencing adaptors were ligated using the Fast-Link ligation kit (EPICENTRE) after the addition of adenosine nucleotide, using exo- Klenow. And samples were PCR amplified with Illumina genomic DNA sequencing primers. PCR products (250 to 450 bp in size) were gel purified and sent for Illumina GA2 “Solexa” sequencing at the UMass Worcester deep sequencing core facility.
Data will be available at Gene Expression Omnibus, Accession #GSE57665.
Raw FastQ reads were first collapsed by their sequences while the occurrences were kept. We then mapped reads to the mm9 genome using bowtie allowing at most one mismatch in the alignment. Only one mapping was randomly picked by the -M 1 parameter setting for dealing with multimappers. Each aligned coordinate was extended toward its 3′ end to reach 150 bp length (although extension was clipped if it exceeded the length of the chromosome). We calculated the relative distance to the nearest TSS for all named genes, and for each TSS tallied the sum of read occurrences from 4 kb upstream to 4 kb downstream. The occurrences were normalized to p.p.m. and binned in 20 bp intervals. For TSS-centered averages (as in Figure 3C, for example) data were additionally normalized relative to the average of the first 2 kb (from −4 kb to −2 kb).
Importantly, for turnover analyses, relatively hot and cold regions are insensitive to the normalization method used—if we normalized all datasets to the hottest regions of the genome, rather than observing decreasing HA enrichment at promoters over time, we would observe very slow incorporation across the rest of the genome with increasing enrichment over time. However, in the absence of a true benchmark with known absolute occupancy (eg a set of promoters with 100% occupancy of MacroH2A2 at t = 3 hours), we choose to utilize standard genome-wide normalization and interpret our dataset with these considerations in mind.
The mouse genome was segmented into nonoverlapping 100 kb tiles (eg chromosome 1 1–100,000, chromosome 1 100,001–200,000, etc.). For each tile, total normalized Macro or HA levels were calculated, and number of annotated TSSs was counted (using only TSSs for named genes). Tiles with the top 1% of signal in the anti-HA dataset from uninduced cells were discarded, as these typically covered regions adjacent to extensive repeats that show artifactual “enrichment” in all public datasets examined, including pre-ChIP input sequencing. For computing correlations between datasets, unmappable tiles with zero mapped reads were also removed.
For all named genes, data were aggregated into 20 bp bins from −4 kb to +4 kb surrounding the annotated TSS. These data were used for clustering and visualization throughout. In addition, we calculated a summary statistic based on total enrichment values for the 1.2 kb stretch from −600 to +600 bp—this value was used for analyses such as Figures 5A, 6D, or 7C (and related Supporting Figures). For comparisons between ES cells and MEFs, we used all genes with an average promoter MacroH2A2 enrichment of at least 0.1, in one of the two datasets, for the 1.2 kB surrounding the TSS.
|
10.1371/journal.pcbi.0030162 | Inferring Function Using Patterns of Native Disorder in Proteins | Natively unstructured regions are a common feature of eukaryotic proteomes. Between 30% and 60% of proteins are predicted to contain long stretches of disordered residues, and not only have many of these regions been confirmed experimentally, but they have also been found to be essential for protein function. In this study, we directly address the potential contribution of protein disorder in predicting protein function using standard Gene Ontology (GO) categories. Initially we analyse the occurrence of protein disorder in the human proteome and report ontology categories that are enriched in disordered proteins. Pattern analysis of the distributions of disordered regions in human sequences demonstrated that the functions of intrinsically disordered proteins are both length- and position-dependent. These dependencies were then encoded in feature vectors to quantify the contribution of disorder in human protein function prediction using Support Vector Machine classifiers. The prediction accuracies of 26 GO categories relating to signalling and molecular recognition are improved using the disorder features. The most significant improvements were observed for kinase, phosphorylation, growth factor, and helicase categories. Furthermore, we provide predicted GO term assignments using these classifiers for a set of unannotated and orphan human proteins. In this study, the importance of capturing protein disorder information and its value in function prediction is demonstrated. The GO category classifiers generated can be used to provide more reliable predictions and further insights into the behaviour of orphan and unannotated proteins.
| As a result of high throughput sequencing technologies, there is a growing need to provide fast and accurate computational tools to predict the function of proteins from amino acid sequence. Most methods that attempt to do this rely on transferring function annotations between closely related proteins; however, a large proportion of unannotated proteins are orphans and do not share sufficient similarity to other proteins to be annotated in this way. Methods that target the annotation of these difficult proteins are feature-based methods and utilise relationships between the physical characteristics of proteins and function to make predictions. One important characteristic of proteins that remains unexploited in these feature-based methods is native structural disorder. Disordered regions of proteins are thought to adopt little or no regular structure and have been experimentally linked with the correct functioning of many proteins. Additionally, disordered regions of proteins can be successfully predicted from amino acid sequence. To address the requirement for protein function prediction methods that target the annotation of orphan proteins and explore the use of information describing protein disorder, a machine learning method for predicting protein function from sequence has been implemented. The inclusion of disorder features significantly improves prediction accuracies for many function categories relating to molecular recognition. The practical utility of the method is also demonstrated by providing annotations for a set of orphan and unannotated human proteins.
| One of the challenges of the post-genomic era is to predict the function of a protein given its amino acid sequence. Most automated function prediction methods rely upon identifying well-annotated sequence and structural homologues to transfer annotations to uncharacterised proteins (see [1,2] for a comprehensive review). Sequence similarity–based methods are relatively successful at annotating homologous proteins; however, they are not applicable to annotating orphan proteins or proteins whose relatives are not themselves functionally annotated. Currently, around 35% of proteins cannot be accurately annotated by homology-based transfer methods [3], highlighting the need for function prediction methods that are independent of sequence similarity.
ProtFun [4,5] is an ab initio feature based protein function prediction method that addresses the annotation of orphan proteins and is applicable to any protein whose sequence is known. The method makes use of sequence-based feature descriptors encoded from localisation, secondary structure, and post-translational modification predictions. Function category predictions were made using individual ensembles of neural networks trained to recognise feature patterns associated with particular functions. Similar approaches have been reported using structural properties and sequence information for prediction of enzyme classes [6,7]. One advantage of this type of approach is that features that are important in recognition of different function classes can be easily identified and quantified.
Over the past few years, there has been a growing awareness of the fundamental importance of disordered proteins in many biological functions and processes. Disordered regions of proteins can be predicted from amino acid sequence [8,9], allowing for rapid surveying of the occurrence of disorder in entire proteomes. The prevalence of disordered proteins in higher eukaryotes is thought to reflect the complexity of signalling and regulatory process within these organisms [10–12].
Disordered regions in proteins are defined as those which lack a stable well-defined 3-D structure in their native states [13,14]. Intrinsically disordered proteins may be either entirely disordered or partially disordered, characterised by long stretches of contiguously disordered residues. The presence of protein disorder is thought to confer dynamic flexibility to proteins, allowing transitions between different structural states [15]. This increased flexibility is advantageous to proteins that recognise multiple target molecules with high specificity and low affinity [13,15].
The functions of numerous disordered proteins have been characterised experimentally and include DNA and protein recognition, transcription and translation regulation, and targeted protein degradation [16–18]. The disordered regions of these proteins have been shown to be essential for their function, forcing a re-examination of the classical sequence–structure–function paradigm central to the field of structural biology and at the core of most automated function prediction algorithms [19,20].
The Protein Trinity hypothesis [19] states that protein function can arise from any of three states: ordered, molten globule, random coil, or from transitions between any or all of these states. Wright [17] defined a continuum of protein structures ranging from an unstructured conformational ensemble to mostly structured proteins containing only locally disordered regions. The functions of disordered proteins along the continuum are influenced by the presence and type of the unstructured regions. For example, disordered stretches can be flexible linker regions that allow movement between domains or can be sites of molecular attachment that become ordered on binding and give rise to functional specificity. In other proteins, disordered regions are associated with sites of post-translational modification that regulate protein–target interactions. It is clear that protein disorder is an important determinant of some protein functions; however, the value of this information remains unquantified and unexploited in current protein function prediction methods. To investigate the correlation of disorder with function, we considered the human complement of disordered proteins as predicted by DISOPRED2 [9,21]. Based on pattern analysis between the distributions of protein disorder and different function annotations, an encoding scheme for representing the occurrence of disorder in proteins is proposed. We then assess the direct influence of protein disorder in function prediction using single class Support Vector Machines [22] (SVMs) to predict individual Gene Ontology [23] (GO) categories.
In this analysis, a protein was considered disordered if it contained a contiguous stretch of predicted disordered residues of ≥30 amino acids. GO categories were identified that were over-represented with disordered proteins as a positive control set of categories likely to be associated with protein disorder features. 31 MF categories and 33 BP categories (Figure 1) were significantly enriched in disordered proteins at corrected p-values of <0.001. The cutoff was purposefully stringent to ensure virtually no false positive terms were selected.
“Transcription factor”, “DNA and protein binding”, “kinase signaling”, and “phosphorylation” molecular function (MF) categories were amongst those enriched in disordered proteins indicated by the highest log ratios of observed/expected occurrence of disordered proteins (Figure 1A). Transcription factor categories were most enriched in disordered proteins, followed by Ion channel and phosphorylation related functions. Metal-ion and nucleotide binding functions exhibited smaller yet significant enrichment in disordered proteins. “Transcription regulation”, “kinase signalling”, “RNA metabolism”, and “phosphorylation” featured in the BP categories (Figure 1B) that were enriched in disordered proteins. These categories were consistent with those functions reported both experimentally [24–26] and those reported in similar analyses of other organisms [10].
We examined the distributions of protein disorder within different GO categories to ensure that the disorder features we used captured the trends and patterns relevant for function prediction. We used location descriptors to encode the position of disordered regions in proteins and length-based descriptors to distinguish short from long contiguous stretches of disordered residues. Correlations between location descriptors and GO categories were demonstrated by calculating the average frequency of disordered residues within different location windows for protein sequences annotated by a GO term (see Methods for more detail, and Figure 2). These averaged values were converted to Z-scores individually for each location window. This procedure normalised for the fact that the false positive rates for prediction of disordered residues are higher at the N and C termini of proteins than in the interior regions [10]. The Z-scores emphasized trends and sampling bias of frequencies of disordered residues directly attributable to the annotation categories. Clustering of annotation categories was performed using Ward's hierarchical method [27], which minimizes within-cluster variance measured by sums of squares error.
The location descriptors showed several trends associated with GO categories. “Transcription regulator”, “DNA binding”, and “RNA pol II Transcription factor” functions were associated with disordered residues in the protein interior, rather than at N and C termini (Figure 2A). “Transcription factor activator”, “Transcription factor repressor”, and “Transcription factor” categories showed significant associations with disordered residues toward the C terminus. Disordered residues were over-represented at the N terminus within the set of Ion Channel and more specifically potassium channel annotated proteins. A further weak association was observed between disorder at the C terminus and the ion channel categories. These observations can be confirmed by crystal structure information. For example, it has been reported that the majority of voltage-gated potassium channel proteins contain intrinsically disordered residues at their N and C terminus [28]. At the N terminus, the residues are responsible for channel inactivation [29]. The disordered residues at the C terminus are adjacent to a PDZ motif mediating binding to scaffold proteins that support the assembly of multiple ion channel subunits into a fully functioning complex [28].
Descriptors for the occurrence of different lengths of disordered regions were also constructed. The link between the length of disordered regions and sequence composition has already been described [30]. To investigate whether this observation also corresponded with functional influences, a similar clustering was performed using descriptors derived from the length distributions of disordered regions within each GO category. The region ranges were selected to reflect the shape of the entire distribution of disordered regions in the human proteome and to avoid sparse descriptors at the upper tail of the distribution (see Figure S1).
Clustering the GO categories by the lengths of their disordered regions (Figure 3) revealed a greater degree of function association (more significant Z-scores associated with GO categories) than for the location descriptors. Long regions of more than 500 contiguous disordered residues were over-represented in transcription-related function categories. Shorter regions (50 residues or less) were over-represented in proteins performing metal ion binding, ion channel, and GTPase regulatory functions. Proteins annotated with serine/threonine kinase and phosphatase categories were also over-represented with contiguous stretches of 300–500 disordered residue regions. Again these findings can be supported by structural evidence. Short disordered regions at the mid- to N-terminal regions in small GTPase regulatory proteins mediate a switching mechanism, enabling the protein to interact with multiple binding partners [31,32]. We demonstrate that these correlations are not simply a function of correlations between protein length and GO categories by considering “Ion Channel” and “Transcription factor binding” categories (Figure 3A). We observed a statistically significant association between shorter disordered regions and the Ion Channel GO category, yet the average length of protein within this annotation category is more than 900 amino acids. In contrast, for “Transcription factor binding”, the opposite trend is observed. The average protein length for this class is closer to 700 amino acids, and we have reported an association with long (more than 500 residue) stretches of disorder.
The correlations between function category and disorder region length may be symptomatic of the nature of the structurally disordered region. Tompa [33] described a general set of six functional classes for Intrinsically Unstructured Proteins (IUPs) that reflect their capacity to fluctuate freely in conformational space or their ability to partner molecules either permanently or transiently. It may be that the correlations displayed here between disordered region length and GO class represent the degree of structural malleability required by the protein to perform its function. For example, longer disordered regions observed in transcription regulator categories (Figure 3B) predominantly act as assemblers that are entirely unstructured and require great flexibility to function. GO categories that contain proteins whose disordered regions are predominantly display sites, for example those that are phosphorylated or involved in ubiquitination (Ubiquitin cycle in Figure 3B), require only shorter disordered regions conferring local flexibility within the protein.
The cluster groupings (Figures 2 and 3) were symptomatic of the relationships between annotation terms in the GO graph structure. Specific terms inherit annotations from general parent terms and thus share protein sequences in common. The fact that inherited terms occupied the same or similar clusters provided evidence for the robustness of the observed trends between different annotation categories. Our systematic analysis of disordered regions in the human proteome revealed significant associations between both lengths and locations of disordered regions within proteins and their different GO categories. Many of the observations can be verified by available experimental structure information, highlighting the potential value in using these attributes of disordered proteins as feature descriptors in a method to predict protein function.
Including highly correlated features as inputs to machine learning algorithms often results in little increase in performance, and can sometimes result in decreased performance. To investigate relationships between the disorder features and other features to be used in function prediction, a large set of general feature descriptors was assembled (see Table S1). These were grouped into biological concepts: glycosylation or secondary structure, for example. Redundancy between feature pairs was evaluated using a feature distance matrix (1-Pearson correlation). To represent the important information in the matrix in fewer dimensions, classical Multi-Dimensional Scaling (MDS) was performed. Visualisation of the matrix using the first three dimensions as orthogonal axes (Figure 4) showed three clearly defined groupings. Amino acid composition, phosphorylation, and glycosylation features formed the first group, followed by secondary structure and transmembrane features. Disorder descriptors form a third group less extended from the origin of the plot. The shorter disorder axis reflects the fact that disordered residues are not predicted for all proteins, and, therefore, the information content within these features is comparably less than for amino acid or secondary structure features, which are generic to all proteins.
The feature relationships agreed with biological knowledge. For example, sequence features such as hydrophobicity and charge were related to the frequencies of particular amino acids within proteins. The correlations between predicted phosphorylation sites and frequency of Ser, Thr, and Tyr residues (Pearson correlation ∼0.2) were due to the fact that high frequencies of phosphorylated residues can only be observed when the relevant amino acid types occurred with a high frequency in the protein. Similarly, the frequencies of predicted O and N glycosylation sites displayed correlations with the occurrence of Asn and Ser/Thr residues. The features most closely related to disorder were random coils, PEST, and low-complexity descriptors with correlation values of 0.472, 0.211, and 0.307, respectively, at the residue frequency level. These correlations, although relatively weak, indicated that some of the information within the disorder features is also encoded by these related feature descriptors. Disordered regions in proteins frequently contain residues that are also recognised as low sequence complexity [34]; however, a region of low complexity does not always imply structural disorder. For example, fibrous proteins such as collagens and silks are rigidly structured in their native state yet contain repetitive regions of low complexity [16]. PEST motifs are degradation motifs present in proteins involved in protein phosphorylation, protein–protein interactions, and cell adhesion [35]. These motifs have been shown to be enriched in an experimentally characterised database of disordered proteins [36], and the residues that characterise the motifs represent a subset of those amino acids known to be disorder-promoting [18,37]. However, the correlations observed here between predicted occurrences of these features were small. The general spatial isolation of disorder descriptors in feature space suggested that they contain unique biological information not represented by the other features previously used in function prediction.
Feature importance estimates for all features were collated across all GO categories using a leave-one-out elimination strategy. The histogram columns (Figure 5) represent the average percentage loss in classifier accuracy for all GO categories belonging to MF and BP ontologies, regardless of their individual category performance. Secondary structure features contributed the most to classifier performance for the majority of MF and BP categories. Disorder features were the second most important feature for BP category recognition. Amino acid composition and secondary structure contributions were higher on average for MF categories than for BPs. For all other features, the importance estimates were higher for BP categories.
Our results suggest that disorder patterns are more indicative of the biological process than the molecular activity of the protein. This is striking considering that only one-third of the proteins in the human proteome are predicted to contain significant disordered regions and the information content of the disorder feature set is comparably lower than that for secondary structure or amino acid composition. One possible reason for this observed difference lies in the respective ontology definitions. BP categories describe modules of functions that make up parts of a multi-step process [23], whereas MFs describe a protein's biochemical activity. For example, the receptor tyrosine kinase signalling BP category annotation describes the series of molecular signals generated as a consequence of a transmembrane receptor tyrosine kinase binding to its physiological ligand. Three example proteins annotated by this term are neurterin precursor a neurotrophic growth factor, Rap guanine nucleotide exchange factor, and erb-B2 receptor tyrosine-protein kinase. These proteins are all unrelated at the primary amino acid sequence and secondary structure level, yet each sequence is predicted to contain at least one 30–50 disordered residue stretch (exemplified in Figure 3B). The role of disordered regions in molecular recognition and in hub proteins in protein–protein interaction networks is well-defined [38–40]. Biologically, it would make sense that proteins that are part of the same multi-step process are more likely to co-localise and possess a common interaction surface such as a disordered region without sharing any similar sequence composition or secondary structure.
To evaluate the contribution of disorder features in classification accuracy for individual categories, the performance loss was measured when disorder features were removed from each classifier using the Matthews Correlation Coefficient (MCC). This measure represents the additional value of disorder features in function prediction, accounting for both interaction and compensatory effects between features. Classifier performances were reported for 26 GO categories (Table 1) whose sensitivity at a false positive rate of 10% exceeded 50%. The significance of the improvements in correlation coefficients for individual categories were evaluated using Fisher's Z test, which considers both the magnitude of the performance increase and the strength of the correlation. The improvements that were significant at the 5% level (p < 0.05) were marked in bold (Table 1, column MCC+diso).
Eleven BP categories and 12 MF categories that were identified as enriched in disordered proteins (Figure 1) showed improvements resulting from the addition of disorder features. Several additional GO classes were identified during feature selection that required disorder features for optimal performance. Seven categories: “UDP-glycosyl transferase”, “hormone”, “growth factors”, “transferase”, “hydrolase”, and “carboxylic acid transporters” were added to the MF set of categories, and “G protein signaling” was added to the BP category set of classifiers. The most notable performance gains were observed for “protein tyrosine kinase signaling,” “G protein signaling”, “ubiquitin specific protease”, “transcription”, “protein kinase”, and “helicase” categories. For some categories (“cation-channel”, “ion channel”, “metal ion transport”, “purine-nucleotide binding”, “nucleotide binding”, and “DNA binding”), little or no performance increase resulted from the addition of disorder features. Particularly for Ion channel, Metal Ion transport, and Nucleotide binding categories, other features such as transmembrane regions or secondary structure better characterised the relationship between the primary amino acid sequence of the protein and its function.
The MCC diso–only values (Table 1) showed the correlation observed when classifiers were trained with only disorder features. Some of the BP categories relating to transcription and the Transcription factor MF category could be recognised with sensitivities of >50% at false positive rates of less than 10%, yielding Matthews correlations of ≥0.3. For these categories, the increased performance resulting from the addition of disorder features (difference between MCC+diso and MCC–diso columns in Table 1) was much lower than the correlation obtained from disorder features alone. This result can be explained by the representation of mutual information between random coil, low complexity, or PEST features reducing the magnitude of the effect of the disorder features. Conversely, for “G protein signaling” and “Receptor tyrosine kinase” BP categories and “Growth factor”, “Helicase”, “Hydrolase”, and “Ubiquitin specific protease” MF categories, the improvement resulting from the addition of disorder features was greater than the correlation obtained using disorder features alone. This finding indicates that disorder features interacted cooperatively with other features in the dataset to achieve a greater performance increase.
Throughout this study, classification performance for GO categories has been reported using the MCC. This measure accounts for the imbalanced class frequencies encountered in the GO term classifiers. For completeness, the classification sensitivities obtained at 10%, 5%, and 1% false positives were reported (Table S2 and Figure 6). The number of positive class labels is also included to stress that different error rates are required for comparable performance between these classifiers. This fact is exemplified by the Receiver Operating Characteristic (ROC) curves (Figure 6 and Table S3) which vary according to class size. The curves have been zoomed in to show the sensitivities at false positive rates of below 50%. The majority of reported classifiers were capable of achieving more than 50% sensitivity at false positive rates of less than 10%. Some categories were not recognised as enriched in disordered proteins using statistical tests due to small class frequencies and low occurrences of proteins containing disordered residues. This finding highlights the advantage in using a machine learning–based approach to assess patterns of disordered features over a simple statistical approach using frequency of occurrence in recognising GO categories for which disorder is an important determinant.
In contrast to the finding that disorder features contributed more to BP category recognition, the improvements for MF and BP categories in Table 1 were slightly greater for MF than BP categories. However, these data reflect a subset of the categories for which we were able to produce accurate classifiers. This result highlights the fact that overall more BP categories utilised information from disorder features for classification than MF categories, resulting in a higher feature importance estimate overall. However, for most of these categories, we were not able to produce sufficiently accurate classifiers to be of practical use.
Our method differed from the original ProtFun method [4] in several important ways. Firstly, our predictions for structure, disorder, and transmembrane regions utilised PSI-BLAST profiles rather than single sequence predictions as feature inputs. Encoding information from sequence profiles in this manner increased the accuracy of feature predictions for those proteins that belonged to unannotated families. Second, additional secondary structure features were encoded that recorded the frequencies of helices and sheets of particular length ranges within each protein. Despite these differences, we felt it was important to provide a benchmark comparison between our method and an independent method that did not utilise disorder information. To assess the performance of the ProtFun method, the ProtFun server GO category assignments used the 14,055 annotated proteins used in this study.
Classifier accuracy was reported for eight common categories (Figure 7 and Table S4). The results indicated that our method outperformed the ProtFun server for all tested categories assessed using the MCC. All of these improvements were significant at the 95% level using Fisher's Z test for significance of correlation difference, except for the ion channel category. The performance of our method without disorder features (Table 1) was also reported so that the improvements in accuracy could be attributed to the use of disorder features or to the use of different training datasets and machine learning algorithms. Four of the compared function categories; “Ion Channel”, “Voltage gated ion channel”, “Cation channel”, and “Metal ion transport” did not utilise information from disorder features; therefore, improvements resulted from other methodological differences. For the remaining categories, “transcription”, “regulation of transcription”, “hormone”, and “growth factor”, the source of performance improvements were a mixture of these effects and the addition of disorder features. The greatest accuracy increase resulting directly from the addition of disorder features was observed for the “growth factor” category. For the “hormone” category, the increased accuracy resulted equally from the addition of disorder features and the algorithm and encoding differences. “Transcription” and “Regulation of transcription” accuracies were improved more by the feature encoding and more recent training datasets used than the addition of disorder features. This result was not surprising considering that the ProtFun features included low complexity, PEST regions, and random coils that overlap considerably with disorder features within these categories.
In this benchmark study, it was difficult to provide an unbiased performance measure that was comparable between the two methods. For ProtFun we were restricted to using the server output alone rather than individual neural network output scores, and any testing dataset was likely to have been used at least partially in the training of this method. However, these results indicate that our method represents a significant improvement in predicting protein function from sequence.
The molecular recognition process and function classifiers reported have been used to classify a dataset of unannotated and orphan IPI proteins. A majority rule approach was applied to the annotations such that three of the five classifiers for each GO term must report a positive term assignment. At a confidence cutoff of 0.6 (see Figure S2 for confidence distributions), we were able to assign putative functions to 317 proteins. The majority of high confidence predictions (>0.9) were made by “transcription” and “DNA binding” MF classifiers (Table 2). Additionally, the hierarchical nature of the relationships between the GO classes can be exploited to distinguish more confident predictions. For example, many of the proteins predicted to be “regulators of transcription” also receive independent positive assignments from parent terms “transcription” and “regulation of cellular process”. The annotations have been made publicly available at http://bioinf.cs.ucl.ac.uk/anno/IPI.html.
The aim of this study was to investigate the contribution of protein disorder features in protein function prediction. This work extended numerous survey studies that report the occurrence of protein disorder within entire proteomes by identifying relevant trends and patterns of disordered regions that can be used to predict the function of proteins. Additionally, we have extended and enhanced the repertoire of GO categories that can be recognised in prediction methods by incorporating disorder features.
Disorder features contributed greater overall improvements in recognition of BP categories than MF categories. In fact, the disorder features were the second most informative feature set in BP category recognition whilst amino acid composition features were the least informative. The differences in feature importance were attributed to the differences in the descriptive nature of the two Ontologies. The anticorrelation observed between the importance of disorder features and amino acid composition for BP categories suggested that associations between disordered region length and location and BP category were not a function of similar amino acid compositions of proteins within BP categories. This finding is particularly relevant for methods that attempt to predict function or possibly protein interactions from amino acid sequence without the use of homologous sequence relationships.
The performance of 26 GO category classifiers could be improved using disorder features. Using the disorder features alone, sensitivities above 50% at false positive rates of less than 10% were obtained for some transcription-related BP categories. The results for all other categories were significantly better than random using disorder features as the sole input. These findings were impressive considering that in this study disordered residues were predicted rather than experimentally confirmed. Consequently, the estimates of feature importance were conservative and restricted by the accuracy of the disorder prediction algorithm. DISOPRED2 currently predicts 57% of residues correctly at a false positive rate of 5%. Additionally, whilst structural and compositional subtypes of disordered region have been suggested in the literature [33,41], such classifications have not yet been exploited in a method that predicts disorder from sequence. The potential value of encoding subtypes of disordered region in our function prediction method is indicated by the fact that in most cases the mutual information contained within PEST and low-complexity features was important for recognition of many of our reported GO categories.
Finally, we have demonstrated the practical application of our classifiers in predicting function for orphan and unannotated human proteins. The classifiers are applicable to any protein sequence and are well-suited to predicting putative molecular recognition functions that can then be assayed in vivo for activity, or for the purpose of target prioritisation. For the better performing classifiers, such as DNA binding and transcription related categories, identification of function from sequence can be performed. Overall, our findings reflect the importance of capturing protein disorder information and demonstrate the value of disorder features in human protein function prediction.
We used the International Protein Index (IPI) [42] as a comprehensive human protein dataset and the Gene Ontology Annotation (GOA) [43] for human. 28,057 proteins were annotated with one or more GO categories. The Cd-Hit [44,45] algorithm was used with a threshold of 60% identity to reduce overall sequence redundancy. The remaining 14,055 sequences were partitioned into five equally sized groups for cross-validation and testing. For rigorous cross-validation, the partitioning algorithm ensured that those sequences with significant homology relationships, defined as having a BLAST E-value ≤ 1e-6, were allocated to either the same training set or the same test dataset but never both. This resulted in five equally sized training and testing sets for each GO term where the maximum sequence identity between pairs of training and testing proteins did not exceed 40% sequence identity or a BLAST E-value of 10−6.
Positive and negative training sets for each GO term with at least 50 representative proteins were generated. Positive training examples included those proteins annotated with a particular GO term or any of its child terms in the GO hierarchy. Negative training examples included those proteins not annotated with the particular GO term or any of its children. To avoid potential class labelling errors, proteins annotated with any of the parent or less specific terms in the GO hierarchy were subsequently removed from the negative training sets. These proteins represent incomplete annotations with respect to the GO category under consideration and may belong to either the positive or negative training set for the given term.
Fisher's exact test was performed under the null hypothesis that the occurrence of the GO term annotation and presence of disorder in a protein were independent. The hypothesis was rejected at p-values of <0.001 after applying Bonferroni multiple testing correction. The calculations were performed using the R package for statistical computing [46]. The degree of over-representation for each GO category was compared using the log odds ratio of observed over expected numbers of disordered proteins. The expected number of disordered proteins represents the background frequency, or occurrence, of disordered proteins by random chance within a sample size equivalent to the size of the GO category. This calculation yields a scale whereby values of zero indicate equality between observed and expected numbers of disordered proteins and higher values indicate a larger difference between the observed and expected values.
For a particular GO term, the set of proteins annotated by the term or any of its child terms was considered. For location-based measures, each protein was split into ten segments; N terminus, equally proportioned segments 1 through 8, and C terminus. The frequency of disordered residues within each segment of each protein was calculated. Disordered residues were defined as those residues predicted to be disordered by DISOPRED2 at a threshold of 5% per residue false discovery rate. The set of frequencies of disordered residues within each location window for proteins annotated by each GO term was then averaged. This resulted in a set of ten average frequencies, one for each location region within each GO category. The average frequencies were Z-score normalised independently within each location window to account for the fact that the false positive rate for prediction of disordered residues is greater at the N and C termini than in the protein interior [10].
A similar approach was adopted to assess correlations between disordered region length in proteins and GO term annotations. Disordered regions in proteins were defined as contiguous stretches of ≥30 residues predicted to be disordered. The average frequency of regions that corresponded to each length range across all proteins annotated by the GO term was then calculated and converted to an independent Z-score for each length range.
The Support Vector Machine [47] (SVM) is an efficient classification algorithm suitable for solving binary classification problems in high-dimensional spaces. The algorithm separates positive from negative class data by positioning a linear hyper-plane though the class examples. Often, the input data is not linearly separable, and a kernel function is required to map the data into a higher dimensional space to find the optimal separating hyperplane. The SVM was chosen over other machine learning methods of choice due to its capacity and ability to control error without causing overfitting to the data.
The SVMlight [48] SVM package was used to train binary classifiers for individual BP and MF GO terms using the radial basis function kernel. Kernel parameters C and γ were selected by exhaustive grid searches performed on a 272 processor Linux cluster that maximised the MCC for each classifier. MCC was chosen as a more informative measure of classifier performance than percent accuracy or error as it avoids bias resulting from unbalanced class frequencies. For example, each of the five testing sets for “GO:0045449 regulation of transcription” comprised 356 positive class examples and 1,726 negative class examples. A classification accuracy of more than 82% can be obtained by setting all predicted outcomes to be negative, whereas the MCC balances and controls for the bias in class frequencies. The MCC is similar to the Pearson correlation coefficient where 0 represents random classification and 1 implies perfect classification.
Feature selection was carried out using a recursive elimination strategy. Initially each classifier was trained and tested using all feature inputs. Optimisation of C and γ kernel parameters was performed at this stage. A single feature set was iteratively removed from the input data and the performance measured in terms of MCC. Feature attributes that did not contribute to classification performance or indeed caused improvements to performance when removed were permanently eliminated from the input data. When no further improvements were observed, a second round of parameter optimisation was performed on the final feature sets to produce final classification performance statistics. The results from feature elimination can be found in Table S2.
The features were divided into global (single values per protein) and spatial (multiple descriptors describing feature location within the protein). Global features comprised amino acid composition, sequence features, signal peptides (SignalP 3.0 [49]), and localisation information (psortII [50]). The sequence features described general protein characteristics calculated directly from the protein sequence such as molecular weight, average hydrophobicity, iso-electric point, charge, and atom counts. Local features Disorder, PEST [51] (motifs rich in proline, glutamate, serine, and threonine), coiled coils, and low-complexity residues were predicted using DISOPRED2 [52], epestfind, coils [53], and pfilt [54] algorithms with default parameter settings. Transmembrane and secondary structure content was predicted using Memsat3 [55] and PSI-Pred [56] algorithms. Post-translational modification features phosphorylation and glycosylation were predicted by NetPhos3.0, Net-N-Glyc, and Net-O-Glyc software [57]. A detailed list of descriptors for these features can be found in Table S1. All feature descriptors were scaled to between 0 and 1 before use in classification. Frequency-based descriptors such as the number of transmembrane regions were log-transformed prior to scaling.
DISOPRED2 was used to predict disordered residues for the representative protein sequence set using three iterations of PSI-BLAST [58] against the UNIPROT database release 6.0. Residues were predicted as being disordered at a false positive rate of 5%. Residue predictions were post-filtered for the presence of transmembrane regions predicted using MEMSAT 3.0 [55] set to default parameters. Predicted disordered regions were further filtered for stretches of at least 30 contiguous residues.
A dataset comprising 2,157 orphan and unannotated IPI human proteins was compiled. These proteins contained one or more predicted disordered regions and represent a mixture of proteins that are either members of unannotated families or have no detectable sequence homologues by BLAST similarity searches. To calibrate comparable prediction accuracies between classifiers, the SVM outputs (distances from the separating hyperplane) were converted to posterior probabilities [59]. The probabilities were estimated from the testing datasets so that they reflect the performance of the classifiers on unannotated proteins. The predictions for the unnanotated disordered proteins have been made publicly available at http://bioinf.cs.ucl.ac.uk/anno/IPI.html. |
10.1371/journal.pntd.0004888 | Tourniquet Test for Dengue Diagnosis: Systematic Review and Meta-analysis of Diagnostic Test Accuracy | Dengue fever is a ubiquitous arboviral infection in tropical and sub-tropical regions, whose incidence has increased over recent decades. In the absence of a rapid point of care test, the clinical diagnosis of dengue is complex. The World Health Organisation has outlined diagnostic criteria for making the diagnosis of dengue infection, which includes the use of the tourniquet test (TT).
To assess the quality of the evidence supporting the use of the TT and perform a diagnostic accuracy meta-analysis comparing the TT to antibody response measured by ELISA.
A comprehensive literature search was conducted in the following databases to April, 2016: MEDLINE (PubMed), EMBASE, Cochrane Central Register of Controlled Trials, BIOSIS, Web of Science, SCOPUS.
Studies comparing the diagnostic accuracy of the tourniquet test with ELISA for the diagnosis of dengue were included.
Two independent authors extracted data using a standardized form.
A total of 16 studies with 28,739 participants were included in the meta-analysis. Pooled sensitivity for dengue diagnosis by TT was 58% (95% Confidence Interval (CI), 43%-71%) and the specificity was 71% (95% CI, 60%-80%). In the subgroup analysis sensitivity for non-severe dengue diagnosis was 55% (95% CI, 52%-59%) and the specificity was 63% (95% CI, 60%-66%), whilst sensitivity for dengue hemorrhagic fever diagnosis was 62% (95% CI, 53%-71%) and the specificity was 60% (95% CI, 48%-70%). Receiver-operator characteristics demonstrated a test accuracy (AUC) of 0.70 (95% CI, 0.66–0.74).
The tourniquet test is widely used in resource poor settings despite currently available evidence demonstrating only a marginal benefit in making a diagnosis of dengue infection alone.
The protocol for this systematic review was registered at PROSPERO: CRD42015020323.
| Dengue is an infectious disease transmitted by mosquitoes in the Tropics. There are 2.5 billion people around the world at risk. Dengue presents as an acute febrile illness with symptoms including headache, bone or joint and muscular pains and rash. The objective of this study is to perform a diagnostic accuracy meta-analysis comparing the use of the Tourniquet Test (TT) to a laboratory assay standard (ELISA) for making a diagnosis of dengue infection. A comprehensive literature search (to April, 2016) was conducted to map and assess the quality of the available evidence, using the following databases: MEDLINE (PubMed), EMBASE, Cochrane Central Register of Controlled Trials, BIOSIS, Web of Science, SCOPUS. We included 16 studies with 28,739 participants in the meta-analysis. Pooled sensitivity for dengue diagnosis by TT was 58% (95% Confidence Interval (CI), 43%-71%) and the specificity was 71% (95% CI, 60%-80%). In the pooled subgroup analysis sensitivity for dengue fever diagnosis was 55% (95% CI, 52%-59%) and the specificity was 63% (95% CI, 60%-66%). The tourniquet test is widely used in resource poor settings despite currently available evidence demonstrating only a marginal benefit in making a diagnosis of dengue infection alone.
| Dengue is an arboviral infection ubiquitous to tropical and sub-tropical regions,[1–3] where it is transmitted by domesticated day-biting mosquitoes including Aedes aegypti. After an incubation period of 4–10 days (mean, 7 days), illness onset is abrupt (with headache, fever, myalgia/arthralgia and rash) and can last up to 14 days[4–8]. Four virus serotypes are in circulation around the world (DEN-1, DEN-2, DEN-3, DEN-4)[9,10], with specific “Asian” genotypes within serotypes DEN-2 and DEN-3 being associated with severe dengue infection particularly in secondary infections [11,12]. According to WHO estimates 50–100 million new dengue infections occur annually, resulting in 500,000 cases of DHF and 22,000 deaths[11,12]. It is thought that approximately 2.5 billion people, or 40% of the world’s population are at risk of dengue infection, with important factors including warm and humid climate, overcrowding and residence in major urban centers[11–14].
Virus transmission can cause a spectrum of illness from subclinical to severe dengue infection characterized by plasma leakage, haemorrhage and end-organ impairment. Characterisation of specific phenotypes of infection is complex and has recently changed[11–14]. Clinically, dengue fever presents as an acute febrile disease with symptoms of headache, bone or joint and muscular pains, rash and leukopenia. Traditionally a further two stages were described, consisting of dengue haemorrhagic fever (DHF), characterized by high fever, haemorrhagic phenomena, often with hepatomegaly. In severe cases, further signs of circulatory failure may develop culminating in dengue shock syndrome (DSS), which is associated with poor outcomes. More recent consensus guidance[12] recommends distinction of dengue illness into dengue (with or without warning signs which may precede the development of more severe infection) and severe dengue (encompassing the manifestations of severe plasma leakage, severe bleeding or severe end-organ involvement)[11,12].
The clinical diagnosis of dengue is challenging as the symptoms are non-specific and common to many other infections[10–12], notably malaria and other arboviral infections. To aid diagnosis, specifically during the initial, acute, febrile phase which may last 2–7 days after the development of fever, the WHO recommend the use of the Tourniquet Test (TT, also known as the Rumpel-Leede or Hess test) to support diagnostic decision-making [13,15–21]. As an inexpensive, quick and easy to perform procedure, use of the TT has become widespread in clinical practice globally. The TT is a marker of capillary fragility and can be undertaken by inflating a blood pressure cuff around the upper arm to the point midway between the individual’s systolic and diastolic blood pressures and leaving it inflated for 5 minutes. The cuff is subsequently released and after two minutes the number of petechiae below antecubital fossa are counted. The test is positive if more than 10 petechiae are present within a square inch of skin on the arm[11,12]. The clinical diagnosis of dengue may be confirmed by laboratory testing, which in many settings involves the measurement of an antibody response (IgM or IgG) by ELISA[3], for years considered to be the diagnostic standard[22]. This test is less sensitive in the first 5 days after exposure and frequently relies on testing of paired sera samples. Newer tests available in some centres include reverse-transcriptase PCR (polymerase chain reaction) or direct antigen detection (non-structural protein 1). While these tests are likely to offer an improvement in diagnostic accuracy, the cost and current limitation of not detecting all serotypes limits their application.
The evidence to support the recommendation and widespread use of the TT to aid the diagnosis of dengue fever is mixed with variable sensitivity and specificity being reported previously(15–21). The aim of this study is to map the evidence, assess the quality of the studies and perform a diagnostic accuracy meta-analysis of the diagnosis of dengue using the TT compared to ELISA.
The protocol for this systematic review was registered at PROSPERO (International prospective register of systemic reviews, http://www.crd.york.ac.uk/PROSPERO/display_record.asp?ID=CRD42015020323). We searched Medical Literature databases Analysis and Retrieval System Online (Medline), Excerpta Medical Database (EMBASE), Allied and Complementary Medicine Database (AMED), Global health, Biological Abstracts/Reports, Reviews, Meetings (BIOSIS) altogether through OVID. Latin American and Caribbean Health Sciences (LILACS) and the Cochrane Library through their website for relevant publications until April 2016. Additionally, we searched the WHO ICTRP (International Clinical Trials Registry Platform) and ClinicalTrials.gov for completed and ongoing studies.
The search, performed according to the Cochrane Highly Sensitive Search Strategy, used the following terms: “Rumpel-Leede” OR capillary OR “blood pressure cuff” OR petechiae OR tourniquet OR “Hess” AND dengue. The search was sensitive, we used no study filters and no language or publication restrictions. We checked the reference lists of all primary studies included for additional references. There were no language or publication restrictions on our search.
We included cross-sectional and cohort studies that evaluated the diagnostic accuracy of tourniquet test for dengue infection. Both retrospective and prospective studies that consecutively or randomly selected patients were included, together with studies that used delayed verification for gold standard. We included studies looking at patients presenting with fever who were subsequently tested for dengue using both the TT (index test) and ELISA detection of antibody response (reference standard).
For this review, definitions of dengue were used according to those proposed by the WHO[11,12], as these were the definitions used during the time period from which studies were drawn. For the purposes of this meta-analysis, ‘dengue’ was considered to consist of non-severe ‘dengue fever’ and ‘haemorrhagic dengue fever’, defined as follows. Dengue fever included fever plus 2 or more symptoms of nausea/vomiting, rash, aches and pains. Dengue hemorrhagic fever (DHF) was considered as infection accompanied by haemorrhagic manifestations such as petechiae and mucosal or gastro-intestinal bleeding[11,12].
Three comparisons were performed; TT vs. ELISA to diagnose dengue (i.e. both non-severe dengue fever plus DHF; TT vs. ELISA to diagnose dengue fever and TT vs. ELISA to diagnose DHF.
Two review authors (AJG, HR) independently assessed all studies identified from the database searches by screening titles and abstracts using the Review Management website Covidence (http://www.covidence.org). We separated potential studies for full-text reading. A third review author (ET) resolved any disagreements, and reasons for including and excluding trials were recorded.
Two review authors (AJG, HR) independently extracted data from the included studies using a standard data extraction form. With this form we extracted information of study design, participant description, index test description, reference test description, dengue classification and total number of participants. A 2x2 table was created for each study comparing both tests.
All included studies were assessed for their methodological quality using the quality assessment tool for diagnostic accuracy studies (QUADAS-2)[23]. The tool is composed of 17 items regarding study patient selection, index test, reference standard and flow and timing. For each domain mentioned there are items for risk of bias and applicability. Items were scored as positive (low risk of bias), negative (high risk of bias), or insufficient information (unclear). A description of each assessment was described in the results section.
For each study, a 2x2 contingency table was constructed. We calculated sensitivity, specificity and likelihood ratios (LRs). When the primary study had 0 in a cell of the 2x2 table, the value of 1 was added, so calculations could be done[24], this only happened in one study(17). We planned to exclude primary studies reporting two cells with 0, but this did not occur.
The sensitivity, specificity and LRs were pooled from each study and a forest plot was generated with 95% confidence intervals. Due to the variability in diagnostic data, we logit-transformed sensitivity and specificity for each primary study and for the aggregate result, considering variability within-study and between-study. The output results are random effects estimates of the mean sensitivity and specificity with corresponding 95% CI. The weighing considered the inverse of the standard error, so indirectly to the sample size reported in the studies. Inconsistency (I2) was explored as an indicator of statistical heterogeneity[24]. Summary receiver operating characteristic (ROC) curves were generated with calculation of area under the curve (AUC) as an indicator of test accuracy. To assess for the possibility of publication bias, we constructed funnel plots to visually assess for signs of asymmetry [25].
Statistical analyses were performed with Stata v10.0 (StataCorp LP, Texas, USA) and with RevMan v5.3 (The Nordic Cochrane Centre, Copenhagen, Denmark)[26].
We identified 1610 studies of which 637 were excluded as duplicates (Fig 1). A total of 973 studies were assessed on the basis of the title and abstract, of which 883 were excluded because they did not fulfill inclusion criteria. Full text studies were retrieved for 90 titles, of which 74 were excluded (Table 1): unable to extract absolute numbers of true positives, false positives, false negatives, and true negatives (n = 46), wrong test (n = 15) and wrong study designs (n = 13)[4–8,27–95].
The remaining 16 studies were included in the systematic review and meta-analysis[13–21,96–102] (Table 2), 14 of which were prospective cohorts while two were retrospective cohort studies. The number of participants in each study ranged from 79 to 13,548. Ten studies were from countries in Asia and six studies from Latin America.
All the analysis showed high levels of heterogeneity, represented by an I2 ranging from 75% to 100%. Given this considerable heterogeneity between studies, we performed a random effects meta-analysis presented below.
We used the instrument QUADAS-2, which is composed of four quality categories (patient selection, reference standard, index test, and flow and timing), to critically appraise each included study (Fig 2). Six studies (33%) were considered to have high risk of bias in patient selection due to inclusion of patient data from a database, raising the possibility of bias from multiple assessors, or selection of patients with pre-existing disease. Two studies (17%) had not adequately described their sampling methods, so were classified as unclear risk. Eight studies (50%) were low risk of bias for patient selection.
Considering the Reference standard category (ELISA), all studies were considered low risk of bias. For the Index test category, four studies (25%) had not clearly described the process used to conduct the TT, blind assessors or train assessors.
For the flow and timing category, only two studies (12.5%) were considered at high risk of bias as the TT was repeated multiple times over a period of several days. Four studies (25%) were considered unclear risk due to lack of information of withdrawals and appropriate sequencing of tests.
In this comparison, we included all 16 studies including both non-severe dengue fever and DHF cases. The pooled sensitivity for dengue diagnosis was 0.58 (95% CI, 0.43–0.71) and the specificity was 0.71 (95% CI, 0.60–0.80)(Fig 3). The positive predictive value was 1.63 (95% CI, 1.45–1.82). The negative predictive value was 0.60 (95% CI, 0.51–0.71). The Diagnostic Odds Ratio was 2.88 (95% CI, 2.17–3.83). The area under the curve was 0.70 (95% CI 0.66–0.74)(Fig 4A).
In this comparison, we included six studies. The pooled subgroup analysis sensitivity for dengue fever diagnosis was 0.66% (95% CI, 0.47–0.81) and the specificity was 0.68 (95% CI, 0.55–0.80)(Fig 5). The positive predictive value was 1.81 (95% CI, 1.45–2.25). The negative predictive value was 0.52 (95% CI, 0.36–0.75). The Diagnostic Odds Ratio was 3.80 (95% CI, 2.40–6.00). The area under the curve was 0.73 (95% CI 0.68–0.76)(Fig 4B).
In this comparison, we included seven studies. In the pooled subgroup analysis, sensitivity for dengue haemorrhagic fever diagnosis was 0.63 (95% CI, 0.39–0.82) and the specificity was 0.60 (95% CI, 0.48–0.70). The positive predictive value was 1.54 (95% CI, 1.06–2.24)(Fig 6). The negative predictive value was 0.59 (95% confidence interval, 0.37–0.86). The Diagnostic Odds Ratio was 2.08 (95% CI, 1.15–6.82). The area under the curve was 0.66 (95% CI, 0.62–0.70)(Fig 4C).
None of the included studies reported data comparing TT and ELISA for patients with dengue shock syndrome.
We conducted a subgroup analysis for the included studies considering only children and adolescents aged 6 months to 15 years. No analysis with adults were conducted, since all 16 included studies did not explore only adults’ participants, when they analyzed adults they mixed the data with children and adolescents
In this subgroup analysis, we included eight studies including both non-severe dengue fever and DHF cases. The pooled sensitivity for dengue diagnosis was 0.71 (95% CI, 0.59–0.82) and the specificity was 0.59 (95% CI, 0.47–0.70) (Fig 7). The positive predictive value was 1.66 (95% CI, 1.45–1.91). The negative predictive value was 0.52 (95% CI, 0.43–0.64). The Diagnostic Odds Ratio was 3.44 (95% CI, 2.25–5.25). The area under the curve was 0.69 (95% CI, 0.65–0.73).
We conducted the following sensitivity analyses of the Dengue vs ELISA analysis: 1. In order to analysis the impact of the mix of cut-off points reported by studies (≥10 petechiae per one square inch and ≥20 petechiae per one square inch) we repeated the analysis in just studies using the criteria of ≥20 petechiae per one square inch. 2. We conducted another sensitivity analysis removing all studies with high risk of selection bias.
Thus, we included 12 studies including both non-severe dengue fever and DHF cases. The pooled sensitivity for dengue diagnosis was 0.64 (95% CI, 0.51–0.74) and the specificity was 0.68 (95% CI, 0.55–0.80) (Fig 8). The positive predictive value was 1.68 (95% CI, 1.46–1.93). The negative predictive value was 0.56 (95% CI, 0.47–0.66). The Diagnostic Odds Ratio was 3.37 (95% CI, 2.33–4.86). The area under the curve was 0.71 (95% CI, 0.67–0.75).
We removed six studies with high risk for selection bias. Thus, 10 studies including both non-severe dengue fever and DHF cases were combined. The pooled sensitivity for dengue diagnosis was 0.64 (95% CI, 0.50–0.76) and the specificity was 0.66 (95% CI, 0.56–0.75) (Fig 9). The positive predictive value was 1.74 (95% CI, 1.52–1.98). The negative predictive value was 0.57 (95% CI, 0.48–0.69). The Diagnostic Odds Ratio was 3.37 (95% CI, 2.35–4.85). The area under the curve was 0.70 (95% CI, 0.66–0.74).
Funnel plot asymmetry test revealed evidence of publication bias (Fig 10). The I2 statistics were, as expected in diagnostic meta-analyses, over 95% in all three comparisons made (Figs 3, 5 and 6)[26].
Dengue fever is an infection with significant global public health importance. Increasing urbanization and crowding in endemic areas, coupled with failing vector control programs have resulted in a significant increase in cases and major outbreaks since the 1950’s[11]. This is the first systematic review and meta-analysis to specifically investigate and compare the utility of the tourniquet test to diagnose dengue infection compared to the widely-used standard laboratory ELISA testing.
Overall, our results demonstrate that the TT is a relatively poor diagnostic test for dengue of any severity. When assessed by ROC analysis, low AUCs (≤0.70) suggest that the TT, as an isolated diagnostic test and in comparison to ELISA, should be classified as a “relatively poor” test.[103]
Funnel plot analysis suggests that there may be a major element of publication bias in the previous reporting of tourniquet test usefulness. Many studies were observed to have overly extreme results, i.e. have a large effect for positive TT and large effect against TT. Reasons for this may include small sample size with wide standard errors, or other problems in study design, with resulting overemphasis in reporting positive or confirmatory results. Resulting estimates of TT efficacy may there have demonstrated a significant skew towards overly positive effects.
In our subgroups analysis, we removed the studies that mixed adults and kept only children and adolescents from 6 months to 15 years of age. We did not find any significant change in utility of TT in diagnosing dengue infection. We did further sensitivity analyses considering both a diagnostic cut-off of ≥20 petechiae per one square inch and a repeat analysis after removing studies at high risk of bias for patient selection. Neither of these analyses led to any significant difference in our findings. This is of interest, as using a higher, stricter cutoff would generally reduce sensitivity but increase specificity for diagnosis. Lack of an effect seen when increasing the threshold to ≥20 petechiae here, suggests limited biological correlation of this clinical observation. Additional analyses demonstrated that data from individual studies were scattered and included a wide range of participant age-groups and geographic regions. Ideally, we would also have liked to investigate the performance of the TT in different age-groups using a range of cutoff thresholds, for example, ≥10 petechiae in children, ≥20 petechiae in adults. It was not possible to extract these age-group specific data from many of the primary studies however.
Considering the subgroup analysis, we could hypothesize that increasing the threshold for diagnosing dengue the sensitivity would decrease and specificity increase, however this hypothesis was not confirmed, the lack effect on the outcomes shows the need to use a more rigorous test to diagnose dengue.
Using the GRADE approach to assess the quality of the evidence generated in this study, we can classify it as low, which means that “further research is very likely to have an important impact on our confidence in the estimate of effect and is likely to change the estimate or any estimate of effect is very uncertain”, the evidence was downgraded due to imprecision (wide confidence intervals) and inconsistency (widely different estimates) across the included studies.
Further limitations to our analysis of data currently available in the literature may arise from heterogeneity in the time periods at which the tests were performed, or the number of occasions on which the test was repeated prior to getting a positive result. Additional practical reasons for previous overestimation of the efficacy of the TT may include difficulties in interpreting a positive result in individuals with different skin pigmentation or variation in the virulence or pathogenicity of strains resulting in higher rates of capillary permeability, for example in South East Asian genotypes of the DEN-2/3 serotypes.
Here, we have assessed and presented the best available evidence for use of the TT in making a diagnosis of dengue infection. Clearly the TT should not be used in isolation for making a diagnosis of dengue, however given the evidence available it is doubtful as to whether the test offers any additional benefit over and above careful clinical evaluation. Inconsistencies in data reporting in the primary study datasets also render assessment of whether the TT maybe useful for particular population/disease subgroups currently infeasible.
The clinical diagnosis of dengue is challenging as disease presentation is almost indistinguishable to many other infections commonly found in the tropics[104]. Current WHO recommendations suggest a combination of clinical history, leukopenia and the tourniquet test result to make a diagnosis if ELISA testing is not available or prior to the availability of results. Given the requirement for paired sera samples in many areas where dengue is endemic to demonstrate an increase in antibody titre, reliance on clinical diagnosis will be still greater.
While still widely used, our analyses suggest that data supporting routine use of the tourniquet test is, at best, relatively poor, however it is important to consider that the quality of the evidence is low due to imprecision and inconsistency across the included studies. Furthermore, the data used to underpin current international recommendations likely overestimate its utility. Over reliance on the use of the TT to support a clinical diagnosis of dengue infection may result in misdiagnosis of patients and inaccurate estimates of disease incidence; relatively low sensitivity but higher specificity suggest that disease incidence may be underestimated if the TT is overly relied on. While current recommendations should be re-examined in light of these findings, replacement of the tourniquet test in routine clinical practice will only come once improved point-of-care diagnostics are made more widely available, especially in resource-poor areas.
|
10.1371/journal.pcbi.1002621 | Single Nucleotide Polymorphisms Can Create Alternative Polyadenylation Signals and Affect Gene Expression through Loss of MicroRNA-Regulation | Alternative polyadenylation (APA) can for example occur when a protein-coding gene has several polyadenylation (polyA) signals in its last exon, resulting in messenger RNAs (mRNAs) with different 3′ untranslated region (UTR) lengths. Different 3′UTR lengths can give different microRNA (miRNA) regulation such that shortened transcripts have increased expression. The APA process is part of human cells' natural regulatory processes, but APA also seems to play an important role in many human diseases. Although altered APA in disease can have many causes, we reasoned that mutations in DNA elements that are important for the polyA process, such as the polyA signal and the downstream GU-rich region, can be one important mechanism. To test this hypothesis, we identified single nucleotide polymorphisms (SNPs) that can create or disrupt APA signals (APA-SNPs). By using a data-integrative approach, we show that APA-SNPs can affect 3′UTR length, miRNA regulation, and mRNA expression—both between homozygote individuals and within heterozygote individuals. Furthermore, we show that a significant fraction of the alleles that cause APA are strongly and positively linked with alleles found by genome-wide studies to be associated with disease. Our results confirm that APA-SNPs can give altered gene regulation and that APA alleles that give shortened transcripts and increased gene expression can be important hereditary causes for disease.
| Variants in DNA that affect gene expression—so-called regulatory variants—are thought to play important roles in common complex diseases, such as cancer. In contrast to variants in protein-coding regions, regulatory variants do not affect protein sequence and function. Instead, regulatory variants affect the amount of protein produced. The 3′ untranslated region (UTR) is one gene region that is critically important for gene regulation; cancers for example, often express genes with shortened 3′UTRs that, compared with full-length 3′UTRs, have higher and more stable expression levels. We have investigated one kind of regulatory variant that can affect the 3′UTR length and thereby cause disease. We identified several such variants in different genes and found that these variants affected the genes' expression. Some of these variants were also strongly linked with known markers for disease, suggesting that these regulatory variants are important hereditary causes for disease.
| In protein-coding genes, the polyadenylation process consists of cleaving the end of the 3′ untranslated region (UTR) of precursor messenger RNA (pre-mRNA) and adding a polyadenylation (polyA) tail. Alternative polyadenylation (APA) can occur when several polyadenylation (polyA) signals lie in the last exon of a protein-coding gene. Many APA signals are evolutionary conserved [1], and Expressed Sequence Tag (EST) data suggest that 54% of human genes have alternative polyadenylation signals [1]. The polyA signals themselves are hexamer DNA sequences that usually lie 10 to 30 nucleotides upstream from the cleavage site [2], but a GU-rich region 20 to 40 nucleotides downstream of the cleavage site is also important for the polyA-process [2].
One functional consequence of APA is transcripts with different 3′UTR lengths and different microRNA (miRNA) regulation [3], [4]. Shortened transcripts tend to have increased expression compared with longer transcripts, and the same expression increase can be achieved by deleting miRNA target sites in non-shortened transcripts [5].
Data on APA can be used as an efficient biomarker for distinguishing between cancer subtypes and for prognosis [6], and seems to play an important role in gene deregulation and in many human diseases [7]. One such mechanism for deregulation is mutations in the polyA signal or GU-rich downstream region [7]. A single nucleotide polymorphism (SNP) in the GU-rich region downstream of an alternative polyA signal in the FGG gene has for example been shown to affect the usage of this polyA site, and has been associated with increased risk for deep-venous thrombosis [8]. Similarly, a mutation in the 3′UTR of the CCND1 gene has been shown to create an alternative polyA signal and is associated with increased oncogenic risk in mantle cell lymphoma [9].
Hypothesizing that mutations in DNA elements such as the polyA signal can be an important cause of altered APA, we investigated to what extent SNPs can create or disrupt APA signals (APA-SNPs). Specifically, we tested whether APA-SNPs can give shorter 3′UTRs, increased gene expression through loss of miRNA regulation (Fig. 1), and be associated with disease. Our hypothesis focuses on shorter 3′UTRs rather than longer ones, because the loss of functional miRNA sites in the 3′UTR is more likely than the gain of new sites downstream of the gene.
First, by analysing EST data, we found that SNPs can create polyA motifs and affect 3′UTR length. Second, differential allelic expression from RNA-seq data, as well as mRNA and miRNA microarray expression data revealed an association between alternative polyA site strength (signal and GU-content), loss of miRNA target sites, and transcript expression. Third, based on these analyses we also identified significant APA-SNPs. Fourth, we mapped the identified SNPs to disease-associated SNPs and found that APA alleles were significantly correlated with disease-risk alleles. Together, these results suggest that APA-SNPs can be a significant causative mechanism in disease (Fig. S1).
The distribution of SNPs within 3′UTRs is fairly uniform [10] (Fig. S2A). The main exceptions are microRNA target sites and the start and end of the 3′UTR, which have decreased SNP diversity that is consistent with these regions containing functional elements under selective pressure [10]. Indeed, when specifically investigating the region around the transcription end site, we found that the position containing the polyA signal has a markedly decreased SNP density (Fig. S2B,C), indicating that SNPs arising there could have a high functional impact.
To analyse SNPs in alternative polyadenylation signals, we first identified a set of SNPs that potentially create new APA signals in 3′UTRs. Specifically, we searched for any Hapmap SNP [11] that could create or disrupt one of the 13 known polyA signal hexamers [1] in any coding gene's 3′UTRs (see Methods). We found 1954 SNPs, including 755 SNPs that are mono-allelic in the CEU population from Hapmap [11] (see Datasets). We kept only the APA-SNPs that change from no signal to one signal in the locus, by discarding loci with several signals in the 40 nucleotides around the SNP, discarding SNPs that change one signal into another, and discarding mono-allelic SNPs. After filtering, 412 SNPs that can create or delete potential polyadenylation signals remained. We will from now refer to them as our candidate SNPs.
To investigate whether SNPs can create functional alternative polyA sites, we analysed the EST-based polyA sites from the PolyA_Db database [12], [13]. In the PolyA_Db database, there are several polyA sites which do not have any noticeable polyA signal (according to the reference genome) in the 40, 80, and 100 nucleotides upstream from the reported cleavage site position (Table S1). In those regions, we used different SNP data to look for SNPs that could create a polyA signal with the non-reference allele. When considering regions of 100 nucleotides and SNPs from NCBI dbSNP Build 130 [14], we could identify polyA signals with the alternative allele for more than of the missing signals. Some of the remaining sites can probably be explained by SNPs further upstream, and some other by exon splicing, by alterations in ESTs that are not registered in dbSNP, or as false positive sites.
Since EST-based annotated polyA sites can be affected by SNPs, we wanted to test whether alleles in polyA sites could be associated with EST ending positions. Specifically, we first took the intersection between the polyA signals from our 412 candidate SNPs, and the polyA sites from PolyA_Db database [12], [13]. We identified 18 intersecting polyA sites that have a polyA signal for either the reference or the non-reference allele. These sites corresponded to 18 SNPs in 18 genes. Five SNPs were discarded because they lie within the 20 last nucleotides of the reference 3′UTR. The following 13 genes remained: ABCC4, AKAP13, FANCD2, KY, MIER1, OSTM1, PNN, RASGRP3, RHOJ, SELS, SHMT1, SLBP, and SLC11A2. Second, for each of these genes, we identified and imputed (see Methods) alleles at the SNPs in the EST sequences when possible, and tested if the proportion of alleles with polyA signal (APA alleles) was different for EST sequences ending within the interval nucleotides around the polyA site, compared to EST sequences ending further downstream (see Methods). The two genes MIER1 (SNP rs17497828) and PNN (SNP rs532) were significant (Fig. 2, Table 1). After correcting for multiple testing (Benjamini & Hochberg correction), the genes remained significant when including alleles imputed based on haplotype (Table 1).
For MIER1, 12 of the 16 EST sequences ending near the annotated APA site had the APA allele (including 2 with a clear polyA tail), whereas 3 had the non-APA allele (none of them had a clear polyA tail). Similarly, for PNN, all of the 34 EST sequences ending near the annotated APA site had the APA allele (including 10 with a clear polyA tail). Together, these results suggest that SNPs can create functional APA sites and thereby affect 3′UTR length.
EST data can be used to identify alleles and transcript ending positions (Fig. 1), but EST data seldom have sufficient resolution to quantify transcript expression levels. In contrast, RNA-seq data can both be used to genotype SNPs [15] and to analyse transcript length and expression patterns. The main challenge with RNA-seq data compared with ESTs, however, is the shorter sequence reads, which makes it challenging to distinguish between homozygotes, heterozygotes with strong expression differences between its alleles (allelic imbalance), sequencing errors, and alignment errors.
To explore whether RNA-seq data could reveal whether APA-SNPs affect transcript expression, we therefore developed and validated an RNA-seq-based genotyping approach (see Supporting Text S2). We then used this approach to show that APA-SNPs can affect transcript expression and that this effect is associated with loss of miRNA regulation. Specifically, we first show that homozygous APA-SNPs have significantly shorter 3′UTRs than have heterozygous or homozygous wildtype SNPs. Second, we show an association between allelic imbalance of heterozygous APA-SNPs and the two following important features of polyA sites: signal strength and GU level downstream of the cleavage site. Third, we show that the loss of miRNA target sites can be the missing link in this association. Fourth, we use allelic imbalance to detect potentially functional APA-SNPs. Fifth, we show that APA-SNPs at strong sites (strong APA signal and high GU level) that have a strong predicted effect on miRNA regulation, have higher allelic imbalance and higher transcript expression than have other APA-SNPs.
To confirm the results from the RNA-seq-based allelic imbalance analyses, we turned to gene expression data from the well characterised Hapmap population. We looked at human gene expression profiling of EBV-transformed lymphoblastoid cell lines from 270 unrelated Hapmap individuals [17], and genotypes of the same individuals, from the Hapmap database [11]. Specifically, we first investigated whether genotypes of SNPs in strong polyA sites that affect miRNA targeting in general are associated with increased gene expression. Second, we investigated whether individual APA-SNP genotypes correlate significantly with gene expression.
Since SNPs can alter polyadenylation and affect miRNA target sites and gene expression, we wondered whether they can also play an important role in human diseases. We therefore tested if any of our APA-SNPs were linked to trait-associated SNPs from the NHGRI GWAS catalogue [20], [21], which consists of SNP-trait associations from published genome-wide association studies (GWAS) (accessed Apr. 18, 2011). Specifically, we mapped our 412 APA-SNPs to the 4304 GWAS SNPs, by using the mapping method described in Thomas et al. [22]. The mapping was based on linkage disequilibrium (LD) data from the Hapmap database (CEU population release 27). We identified 135 APA-SNP/GWAS-SNP pairs (consisting of 84 unique APA-SNPs and 123 unique GWAS SNPs) that had available haplotype data in Hapmap and one known and unique risk allele in the GWAS catalogue. For each APA-SNP/GWAS-SNP pair, we computed the correlation between the APA allele and risk allele as the LD value [23], where , , and are respectively the APA allele frequency of the APA-SNP, the risk allele frequency of the GWAS SNP, and the “APA allele risk allele” haplotype frequency in the CEU Hapmap population. For each of the 84 unique APA-SNPs, we computed as the mean of when an APA-SNP was linked to several GWAS SNPs, and similarly as the mean of .
We hypothesised that if APA-SNPs play a role in diseases, then APA alleles would be positively () and strongly (high ) correlated with risk alleles, particularly for the significant APA-SNPs that we identified in the previous sections, as they are more likely to be functional, and particularly those that are linked to GWAS-SNPs from CEU-population-related studies, since the values are based on CEU haplotypes.
Among the 84 APA-SNPs, 60 were paired to GWAS-SNPs that are trait-associated in CEU-related populations. Nine of those SNPs were identified in the previous sections as significant APA-SNPs, and those nine SNPs had a significantly high number of positive (more positive correlations between APA and risk alleles than expected) and a significantly high number of greater than 0.2 (higher number of correlations between APA and risk alleles than expected) (Table 2). In contrast, for computed from CEU haplotypes but for GWAS-SNPs that are trait-associated in non-CEU-related populations, binomial test p-values were not significant, suggesting that GWAS and haplotype data should be matched according to population, to detect potential disease-related APA-SNPs.
Those results show that a significantly high proportion of our candidate SNPs is in LD with trait-associated SNPs and their APA alleles are positively correlated with risk alleles of trait SNPs. This suggests that those APA-SNPs can potentially be the cause of their corresponding disease-association signals measured and registered in the GWAS catalogue.
Our analyses confirmed the hypothesis (presented in Fig. 1) that SNPs can create functional alternative polyadenylation signals and thereby affect miRNA-based gene regulation and give increased gene expression. Both EST and RNA-seq analyses supported our hypothesis, despite some limitations. Additionally, the microarray analysis could further confirm these results and strengthen our hypothesis. Given the results from these three analyses, we estimate the proportion of functional APA-SNPs to be ().
The EST analysis supports our hypothesis but has some limitations. Specifically, we analysed EST data for 13 genes and found that 2 of them had an APA-SNP that could create polyA motifs and affect 3′UTR length. However, the EST analysis does not take into account the presence of a polyA tail in the EST sequence. Moreover, the ESTs came from a mix of tissues, which could also affect the results. Segregating ESTs based on tissue origin or filtering on sequences with clear tails in the “short” group, reduces sample size and affects statistical power. However, when combining sequences from our two significant genes, all of the 12 EST sequences ending at the alternative cleavage site and that have a polyA tail, had the APA allele. This number is significant (binomial test p-value of , where the expected proportion of the APA allele is the combination of weighted allele frequencies of APA alleles for the 2 SNPs), and tells that the shortened transcripts arose from functional APA signals from the APA alleles.
Similarly, RNA-seq data from the Burge Lab, matched to miRNA expression data showed association between alternative polyA site strength (signal and GU-content), loss of miRNA target sites, allelic imbalance, and transcript expression. The Burge dataset was generated by using cDNA fragmentation, which gives a good coverage of 3′UTRs [24]. An increased allelic imbalance towards the APA allele could come from the loss of miRNA target sites, but also from the fragmentation method. This is because cDNA fragmentation gives a good coverage at the end of the transcript, and, in case of alternative polyadenylation, the transcript is shorter for the APA allele, which results in a high coverage at the SNP locus. In contrast, a longer transcript with the non-APA allele could have a higher coverage downstream, but a lower coverage at the SNP locus. Bias from cDNA fragmentation would therefore give an increased allelic ratio towards the APA allele simply because of transcript length differences. Consequently, we cannot exclude that some of the overall RNA-seq trends can be attributed to cDNA fragmentation bias.
The independent microarray data strongly support the EST and RNA-seq results, however. Specifically, the mRNA and miRNA microarray expression data showed association between alternative polyA site strength (signal and GU-content), loss of miRNA target sites, and transcript expression. This microarray analysis had the advantage of directly using genotype data from Hapmap, instead of genotyping SNPs through mapped RNA-seq reads. Furthermore, the microarray analysis focused on transcript expression differences between individuals and therefore required data from a unique cell type, whereas the RNA-seq analysis focused on allelic expression differences within a sample and could therefore involve different cell types. As expected, the microarray analysis showed similar results as the RNA-seq analysis, suggesting that the increased allelic ratios from RNA-seq data did not come from a potential bias due to the cDNA fragmentation method, but from the loss of functional miRNA target sites.
One clear disadvantage of using the RNA-seq data for genotyping and allelic-imbalance-based detection, was false positive homozygotes. We could detect potentially functional candidate SNPs by testing for allelic imbalance, which takes into account the number of reads and their quality, while testing for unusual allele proportion patterns. The difficulty was to find extreme allelic imbalance, as we could miss extreme imbalance by classifying a locus as homozygote because of too few reads () corresponding to the alternative allele. This was a conscious trade-off, however, since we wanted to maximise true positive heterozygotes and avoid false positives (i.e. predicted heterozygotes that were in fact homozygous).
RNA-seq data enabled us to genotype SNPs in expressed genes and compute allelic imbalance. Genotype classification could be checked with known genotypes from the Heap dataset and with mono-allelic SNPs. However the Heap dataset could not be used in the allelic imbalance analysis, because the library was generated by using RNA fragmentation, which gives a good coverage for the coding regions [24], but not for the UTRs. Since we were interested in SNPs in 3′UTRs, and particularly at the end of potentially alternative transcripts, RNA fragmentation would affect allelic imbalance.
The whole analysis is limited to SNPs that can make the reference 3′UTR shorter, lose miRNA sites and upregulate genes, because the loss of functional miRNA sites within the 3′UTR is more likely than the gain of new ones downstream of the annotated 3′UTR. However, it could be interesting to consider the hypothesis where SNPs in the signals at the end of the reference transcript could make 3′UTR longer having more miRNA target sites further downstream, and down-regulate the gene.
Alternative polyadenylation alleles play a role in 3′UTR shortening, gene deregulation, and increased disease risk (Fig. 1). Our analyses confirm that APA is an important factor for miRNA-mediated gene regulation [4]. EST data suggest that SNPs can create polyA motifs and affect 3′UTR length, and allelic imbalance from RNA-seq data coupled to miRNA expression data suggest an association between alternative polyA site strength (signal and GU-content), loss of miRNA target sites, allelic imbalance and transcript expression. Similarly, mRNA microarray expression data and matched genotypes of the same individuals, coupled with miRNA expression data could confirm association between alternative polyA site strength (signal and GU-content), loss of miRNA target sites, genotype and transcript expression.
Each of our analyses could also be used to detect potentially functional APA-SNPs. The detected APA-SNPs could further be linked to GWAS-SNP markers and a significant part of these APA-SNPs had their APA allele positively correlated with disease-risk alleles. We propose that these APA SNPs are potential disease-causative variants.
We used SNP data from the CEU population (CEPH - Utah residents with ancestry from northern and western Europe) from the human haplotype map project (HapMap database [11]), release 22 for haplotype data, and release 27 for the genotype, allele, frequency, and linkage disequilibrium data. We used the human genome assembly version 18 (hg18) [25], RefSeq gene annotations (hg18 version), and EST sequences from the UCSC Genome browser [26]. We used human APA sites from PolyA_Db [12], [13]. We used disease-associated SNPs from the NHGRI GWAS catalogue [20], [21]. RNA-seq data came from Heap et al. [15] and from the Burge Lab [27]. Human miRNA profiles came from Landgraf et al. [16] (their Table S5) and from Wang et al. [18]. MicroRNA data came from the MirBase database release 16 [28].
Thirteen polyA signal motifs are known in human genes: AAUAAA, AUUAAA, UAUAAA, AGUAAA, AAGAAA, AAUAUA, AAUACA, CAUAAA, GAUAAA, AAUGAA, UUUAAA, ACUAAA, and AAUAGA [1] (ordered by strength ranks). We detected SNPs in potential APA signals, by a motif search that looks if any CEU Hapmap SNP in the 3′UTR of any coding gene would create/disrupt one of those 13 motifs. For a given SNP, the motif search looks for a given motif in an mRNA sub-sequence consisting of the SNP and its flanking sequences (6 nucleotides up/downstream), for each allele.
We downloaded the 28.857 APA sites (human) from PolyA_Db [12], [13] from the UCSC track (hg18) [26]. We downloaded knownToLocusLink.txt and knownToRefSeq.txt from UCSC (hg18) [26] to convert entrez gene ID to RefSeq gene ID. We took the intersection between our APA signals and polyA sites from PolyA_Db, by taking all the sites from PolyA_Db that lie up to 40 bp downstream of our signals.
For each of the 13 candidate genes, we downloaded the EST sequences (Expressed sequence tag) from UCSC (hg18, tables ‘all_mrna’ and ‘all_est’) [26] that lie within their 3′UTR region. We also downloaded their alignment to their reference mRNA sequence from UCSC [26], and the list of EST that support the considered polyA site from PolyA_Db2 [13]. We used sequence alignment to identify the allele and haplotype of each sequence, when possible. Otherwise, the APA-SNP allele was imputed, by using haplotypes from the CEU Hapmap population [11] (see Dataset). We tested the proportion of APA alleles that support the candidate APA site, versus longer transcripts, by using a 2×2 contingency table. If the 4 expected values were greater than 5: we used the 2×2 , and Fisher's exact test otherwise.
Given a 3′UTR region of a gene of interest, we took all the phased SNPs from Hapmap [11] in that region, as well as their haplotypes in the CEU population [11]. For each of those SNPs, we identified the allele in the EST sequence when possible, to identify the EST haplotype. We discarded EST haplotypes that had zero identified allele. For each remaining EST haplotype, we selected haplotypes from Hapmap that fit the identified alleles in the EST haplotype. The APA-SNP could be imputed if there was only one unique allele at that SNP in all the selected haplotypes from Hapmap.
We downloaded RNA-seq data from human primary cells from 4 individuals [15] (Short read archive accession number: SRA008367), reads in FASTQ format, length of 45 bp. We downloaded Burge lab RNA-seq [27] (Short read archive: SRA002355, and Gene expression omnibus: GSE12946): Human tissue samples (brain, liver, heart, skeletal muscle, colon, adipose, testes, lymph node, breast, MAQC, 6 Cerebellum), immortalised and cancer cell lines (BT474, HME, MCF-7, MD435, T47D, MAQC UHR), reads in FASTQ format, length of 36 bp. MAQC is a mixture of brain cell types from several donors, MAQC UHR is a mixture of several cancer cell lines, and MD435 is thought to be contaminated by the M14 melanoma cell line. Therefore those 3 cell lines were discarded from the allelic imbalance analysis.
We mapped RNA-seq reads using the RMAP software [29], with option ‘-Q’ for position weight matrix matching, based on quality score. Alignment was stored in BED files. We used the default options: 2 mismatches allowed in the 32 first nucleotides, 10 mismatches allowed in the whole read. Ambiguous reads were discarded. Paired-End reads were mapped as Single-End reads.
We mapped those reads to 3′UTR 50 bp: the reference sequence is all 3′UTR DNA sequences (from the human genome assembly HG18 [25]) from all coding genes (excluding Y chromosome because of overlap with X), including introns, extended of 50 nucleotides up- and downstream. Overlapping sequences were merged (19012 regions). We mapped reads to a second version of the reference sequence, where reference alleles of APA-SNPs were replaced by non-reference alleles.
We counted base calls based on base quality probability score and sequence alignment score: We discarded reads mapped with an alignment score , and reads that had a quality score accuracy at the SNP. Quality score probability of accuracy at a SNP was computed as follows: , where is the ASCII character of one base call in a read in FASTQ file format [30]. We computed the mapping score as , where is the alignment score given by RMAP. We counted alleles as for each allele (for all the FASTQ files of each individual). We discarded alleles that do not fit Hapmap bi-allelic SNPs. If there was only one allele left, we classified the SNP as homozygous. If there were two alleles left, with both proportions greater than 0.15, we classified the SNP as heterozygous. If there were two alleles but one had its proportion lower than 0.15, we classified the SNP as homozygous with the allele having the biggest proportion.
We mapped reads from the Burge dataset using the alignment software Bowtie [31] version 0.12.7 with default options. Bowtie generated alignments in the SAM format [32]. The transcript assembly software Cufflinks version 1.3.0 [33] was then used with the SAM files to generate a list of expressed exons for each run (default options). Those exons were then mapped back to annotated RefSeq genes. Exons that mapped to several different genes were discarded; the corresponding genes they overlapped were also discarded. For a given gene and a given run, the 3′ end of the exon that mapped the most downstream on the gene was used as an estimate of the gene's 3′ end. Finally, the distance between the estimate and the annotated transcript end was computed for each gene and each run. This distance is negative when the transcript is shorter than the annotation and had a logarithmic distribution for negative s. Few transcripts were longer than the annotated transcription end site, resulting in positive values. To handle these few positive values, we put a threshold at 30, so that all were truncated to 29. We then converted the s to the logarithm scale by using the following formula: .
Log Allelic Ratio for each heterozygous SNP is defined as , where counts of alleles are computed in a similar way as in the genotyping section (by taking base quality and alignment score into account). is positive when the transcripts with APA alleles are up-regulated compared to non-APA allele.
However, to avoid that a mapping bias towards reference alleles affects allelic ratios, we used a corrected allelic imbalance in our analyses, by combining allelic ratios computed from reads mapped to the reference genome with reference alleles, and allelic ratios computed from reads mapped to the same genome but with non-reference alleles at candidate SNPs. We defined it as the mean of the two log-ratios:
where is the allelic ratio, and are the counts of APA alleles mapped to respectively the genome with reference alleles, and the one with non-reference alleles. Similarly and are the counts of non-APA alleles.
We took all the known coding genes from the UCSC RefSeq gene database (hg18) [26]. To define the precise region of GU-analysis, for each gene, we computed the GU proportion in a 5-nucleotide long window sliding from the polyA signal downstream in a 70-nucleotide long region. Those curves represent the variation of GU proportion in the region for each gene. We then took the mean of all the curves, which showed that the increased GU region was from the window to the window (Fig. S4). We therefore defined the GU level as the mean of the GU-proportions in the 5-nucleotide windows, from the to the downstream of the polyA signal.
We downloaded human gene expression profiling of EBV-transformed lymphoblastoid cell lines from 270 unrelated Hapmap individuals [17] (Gene expression omnibus: GSE6536, data normalised across populations), and genotypes for the same individuals, from the Hapmap database release 27.
We mapped probe IDs to RefSeq genes using the BioConductor package for R [35], [36] (R version 2.10.1, AnnotationDbi package version 1.8.2 [37] and the annotation file illuminaHumanv1.db version 1.4.0). One candidate SNP could have one or several RefSeq gene IDs, which could be mapped to one or several probe IDs. Among those probe IDs, we selected the one with maximum variance across all the individuals in the dataset, and assigned it to the given SNP in the 3′UTR.
Genotype was encoded as 0, 1, and 2 for non-APA homozygotes, heterozygotes, and APA homozygotes, respectively.
We computed bootstraps of median differences: Given two groups with different sizes, we resampled with replacement in each group with their actual original size. We took the median in each resampling and computed the difference. We repeated this procedure 1000 times to create a median difference distribution, which was then used to compute the 95% confidence interval ( CI).
We mapped APA-SNPs to GWAS SNPs, using the mapping method described in Thomas et al. [22]. The mapping was based on linkage disequilibrium (LD) data from the Hapmap database (CEU population release 27). The mapping parameter was the threshold (see Thomas et al. [22]), to identify all neighbouring APA-SNP/GWAS-SNP pairs.
|
10.1371/journal.pcbi.1000045 | The Dynamics of Human Body Weight Change | An imbalance between energy intake and energy expenditure will lead to a change in body weight (mass) and body composition (fat and lean masses). A quantitative understanding of the processes involved, which currently remains lacking, will be useful in determining the etiology and treatment of obesity and other conditions resulting from prolonged energy imbalance. Here, we show that a mathematical model of the macronutrient flux balances can capture the long-term dynamics of human weight change; all previous models are special cases of this model. We show that the generic dynamic behavior of body composition for a clamped diet can be divided into two classes. In the first class, the body composition and mass are determined uniquely. In the second class, the body composition can exist at an infinite number of possible states. Surprisingly, perturbations of dietary energy intake or energy expenditure can give identical responses in both model classes, and existing data are insufficient to distinguish between these two possibilities. Nevertheless, this distinction has important implications for the efficacy of clinical interventions that alter body composition and mass.
| Understanding the dynamics of human body weight change has important consequences for conditions such as obesity, starvation, and wasting syndromes. Changes of body weight are known to result from imbalances between the energy derived from food and the energy expended to maintain life and perform physical work. However, quantifying this relationship has proved difficult, in part because the body is composed of multiple components and weight change results from alterations of body composition (i.e., fat versus lean mass). Here, we show that mathematical modeling can provide a general description of how body weight will change over time by tracking the flux balances of the macronutrients fat, protein, and carbohydrates. For a fixed food intake rate and physical activity level, the body weight and composition will approach steady state. However, the steady state can correspond to a unique body weight or a continuum of body weights that are all consistent with the same food intake and energy expenditure rates. Interestingly, existing experimental data on human body weight dynamics cannot distinguish between these two possibilities. We propose experiments that could resolve this issue and use computer simulations to demonstrate how such experiments could be performed.
| Obesity, anorexia nervosa, cachexia, and starvation are conditions that have a profound medical, social and economic impact on our lives. For example, the incidence of obesity and its co-morbidities has increased at a rapid rate over the past two decades [1],[2]. These conditions are characterized by changes in body weight (mass) that arise from an imbalance between the energy derived from food and the energy expended to maintain life and perform work. However, the underlying mechanisms of how changes in energy balance lead to changes in body mass and body composition are not well understood. In particular, it is of interest to understand how body composition is apportioned between fat and lean components when the body mass changes and if this energy partitioning can be altered. Such an understanding would be useful for optimizing weight loss treatments in obese subjects to maximize fat loss or weight gain treatments for anorexia nervosa and cachexia patients to maximize lean tissue gain.
To address these issues and improve our understanding of human body weight regulation, mathematical and computational modeling has been attempted many times over the past several decades [3]–[19]. Here we show how models of body composition and mass change can be understood and analyzed within the realm of dynamical systems theory and can be classified according to their geometric structure in the two dimensional phase plane. We begin by considering a general class of macronutrient flux balance equations and progressively introduce assumptions that constrain the model dynamics. We show that two compartment models of fat and lean masses can be categorized into two generic classes. In the first class, there is a unique body composition and mass (i.e. a stable fixed point) that is specified by the diet and energy expenditure. In the second class, there is a continuous curve of fixed points (i.e. an invariant manifold) with an infinite number of possible body compositions and masses at steady state for the same diet and energy expenditure rate. We show that almost all of the models in the literature are in the second class. Surprisingly, the existing data are insufficient to determine which of the two classes pertains to humans. For models with an invariant manifold, we show that an equivalent one dimensional equation for body composition change can be derived. We give numerical examples and discuss possible experimental approaches that may distinguish between the classes.
The human body obeys the law of energy conservation [20], which can be expressed as(1)where ΔU is the change in stored energy in the body, ΔQ is a change in energy input or intake, and ΔW is a change in energy output or expenditure. The intake is provided by the energy content of the food consumed. Combustion of dietary macronutrients yields chemical energy and Hess's law states that the energy released is the same regardless of whether the process takes place inside a bomb calorimeter or via the complex process of oxidative phosphorylation in the mitochondria. Thus, the energy released from oxidation of food in the body can be precisely measured in the laboratory. However, there is an important caveat. Not all macronutrients in food are completely absorbed by the body. Furthermore, the dietary protein that is absorbed does not undergo complete combustion in the body, but rather produces urea and ammonia. In accounting for these effects, we refer to the metabolizable energy content of dietary carbohydrate, fat, and protein, which is slightly less than the values obtained by bomb calorimetry. The energy expenditure rate includes the work to maintain basic metabolic function (resting metabolic rate), to digest, absorb and transport the nutrients in food (thermic effect of feeding), to synthesize or break down tissue, and to perform physical activity, together with the heat generated. The energy is stored in the form of fat as well as in lean body tissue such as glycogen and protein. The body need not be in equilibrium for Equation 1 to hold. While we are primarily concerned with adult weight change, Equation 1 is also valid for childhood growth.
In order to express a change of stored energy ΔU in terms of body mass M we must determine the energy content per unit body mass change, i.e. the energy density ρM. We can then set ΔU = Δ(ρMM). To model the dynamics of body mass change, we divide Equation 1 by some interval of time and take the limit of infinitesimal change to obtain a one dimensional energy flux balance equation:(2)where I = dQ/dt is the rate of metabolizable energy intake and E = dW/dt is the rate of energy expenditure. It is important to note that ρM is the energy density of body mass change, which need not be a constant but could be a function of body composition and time. Thus, in order to use Equation 2, the dynamics of ρM must also be established.
When the body changes mass, that change will be composed of water, protein, carbohydrates (in the form of glycogen), fat, bone, and trace amounts of micronutrients, all having their own energy densities. Hence, a means of determining the dynamics of ρM is to track the dynamics of the components. The extracellular water and bone mineral mass have no metabolizable energy content and change little when body mass changes in adults under normal conditions [21]. The change in intracellular water can be specified by changes in the tissue protein and glycogen. Thus the main components contributing to the dynamics of ρM are the macronutrients - protein, carbohydrates, and fat, where we distinguish body fat (e.g. free fatty acids and triglycerides) from adipose tissue, which includes water and protein in addition to triglycerides. We then represent Equation 2 in terms of macronutrient flux balance equations for body fat F, glycogen G, and protein P:(3)(4)(5)where ρF = 39.5 MJ/kg, ρG = 17.6 MJ/kg, ρP = 19.7 MJ/kg are the energy densities [3], IF,IC,IP are the intake rates, and fF, fC, 1−fF−fC are the fractions of the energy expenditure rate obtained from the combustion of fat, carbohydrates (glycogen) and protein respectively. The fractions and energy expenditure rate are functions of body composition and intake rates. They can be estimated from indirect calorimetry, which measures the oxygen consumed and carbon dioxide produced by a subject [22]. The intake rates are determined by the macronutrient composition of the consumed food, and the efficiency of the conversion of the food into a utilizable form. Transfer between compartments such as de novo lipogenesis where carbohydrates are converted to fat or gluconeogenesis where amino acids are converted into carbohydrates can be accounted for in the forms of fF and fC. The sum of Equations 3, 4, and 5 recovers the energy flux balance Equation 2, where the body mass M is the sum of the macronutrients F, G, P, with the associated intracellular water, and the inert mass that does not change such as the extracellular water, bones, and minerals, and ρM = (ρFF+ρGG+ρPP)/M.
The intake and energy expenditure rates are explicit functions of time with fast fluctuations on a time scale of hours to days [23]. However, we are interested in the long-term dynamics over weeks, months and years. Hence, to simplify the equations, we can use the method of averaging to remove the fast motion and derive a system of equations for the slow time dynamics. We do this explicitly in the Methods section and show that the form of the averaged equations to lowest order are identical to Equations 3–5 except that the three components are to be interpreted as the slowly varying part and the intake and energy expenditure rates are moving time averages over a time scale of a day.
The three-compartment flux balance model was used by Hall [3] to numerically simulate data from the classic Minnesota human starvation experiment [21]. In Hall's model, the forms of the energy expenditure and fractions were chosen for physiological considerations. For clamped food intake, the body composition approached a unique steady state. The model also showed that apart from transient changes lasting only a few days, carbohydrate balance is precisely maintained as a result of the limited storage capacity for glycogen. We will exploit this property to reduce the three dimensional system to an approximately equivalent two dimensional system where dynamical systems techniques can be employed to analyze the dynamics.
The various flux balance models can be analyzed using the methods of dynamical systems theory, which aims to understand dynamics in terms of the geometric structure of possible trajectories (time courses of the body components). If the models are smooth and continuous then the global dynamics can be inferred from the local dynamics of the model near fixed points (i.e. where the time derivatives of the variables are zero). To simplify the analysis, we consider the intake rates to be clamped to constant values or set to predetermined functions of time. We do not consider the control and variation of food intake rate that may arise due to feedback from the body composition or from exogenous influences. We focus only on what happens to the food once it is ingested, which is a problem independent of the control of intake. We also assume that the averaged energy expenditure rate does not depend on time explicitly. Hence, we do not account for the effects of development, aging or gradual changes in lifestyle, which could lead to an explicit slow time dependence of energy expenditure rate. Thus, our ensuing analysis is mainly applicable to understanding the slow dynamics of body mass and composition for clamped food intake and physical activity over a time course of months to a few years.
Dynamics in two dimensions are particularly simple to analyze and can be easily visualized geometrically [34],[35]. The one dimensional models are a subclass of two dimensional dynamics. Three dimensional dynamical systems are generally more difficult to analyze but Hall [3] found in simulations that the glycogen levels varied over a small interval and averaged to an approximate constant for time periods longer than a few days, implying that the slow dynamics could be effectively captured by a two dimensional model. Reduction to fewer dimensions is an oft-used strategy in dynamical systems theory. Hence, we focus our analysis on two dimensional dynamics.
In two dimensions, changes of body composition and mass are represented by trajectories in the L–F phase plane. For IF and IL constant, the flux balance model is a two dimensional autonomous system of ordinary differential equations and trajectories will flow to attractors. The only possible attractors are infinity, stable fixed points or stable limit cycles [34],[35]. We note that fixed points within the context of the model correspond to states of flux balance. The two compartment macronutrient partition model is completely general in that all possible autonomous dynamics in the two dimensional phase plane are realizable. Any two or one dimensional autonomous model of body composition change can be expressed in terms of the two dimensional macronutrient partition model.
Physical viability constrains L and F to be positive and finite. For differentiable f and E, the possible trajectories for fixed intake rates are completely specified by the dynamics near fixed points of the system. Geometrically, the fixed points are given by the intersections of the nullclines in the L–F plane, which are given by the solutions of IF−fE = 0 and IL = (1−f)E = 0. Example nullclines and phase plane portraits of the macronutrient model are shown in Figure 1. If the nullclines intersect once then there will be a single fixed point and if it is stable then the steady state body composition and mass are uniquely determined. Multiple intersections can yield multiple stable fixed points implying that body composition is not unique [4]. If the nullclines are collinear then there can be an attracting one dimensional invariant manifold (continuous curve of fixed points) in the L–F plane. In this case, there are an infinite number of possible body compositions for a fixed diet. As we will show, the energy partition model implicitly assumes an invariant manifold. If a single fixed point exists but is unstable then a stable limit cycle may exist around it.
The fixed point conditions of Equations 8 and 9 can be expressed in terms of the solutions of(26)(27)where I = IF+IL, and we have suppressed the functional dependence on intake rates. These fixed point conditions correspond to a state of flux balance of the lean and fat components. Equation 26 indicates a state of energy balance while Equation 27 indicates that the fraction of fat utilized must equal the fraction of fat in the diet. Stability of a fixed point is determined by the dynamics of small perturbations of body composition away from the fixed point. If the perturbed body composition returns to the original fixed point then the fixed point is deemed stable. We give the stability conditions in Methods.
The functional dependence of E and f on F and L determine the existence and stability of fixed points. As shown in Methods, an isolated stable fixed point is guaranteed if f is a monotonic increasing function of F and a monotonic decreasing function of L. If one of the fixed point conditions automatically satisfies the other, then instead of a fixed point there will be a continuous curve of fixed points or an invariant manifold. For example, if the energy balance condition 26 automatically satisfies the fat fraction condition 27, then there is an invariant manifold defined by I = E(F,L). The energy partition model has this property and thus has an invariant manifold rather than an isolated fixed point. This can be seen by observing that for f given by Equation 15, Equation 26 automatically satisfies condition 27. An attracting invariant manifold implies that the body can exist at any of the infinite number of body compositions specified by the curve I = E(F,L) for clamped intake and energy expenditure rates (see Figure 1C). Each of these infinite possible body compositions will result in a different body mass M = F+L (except for the unlikely case that E is a function of the sum F+L). The body composition is marginally stable along the direction of the invariant manifold. This means that in flux balance, the body composition will remain at rest at any point on the invariant manifold. A transient perturbation along the invariant manifold will simply cause the body composition to move to a new position on the invariant manifold. The one dimensional models have a stable fixed point if the invariant manifold is attracting. We also show in Methods that for multiple stable fixed points or a limit cycle to exist, f must be nonmonotonic in L and be finely tuned. The required fine-tuning makes these latter two possibilities much less plausible than a single fixed point or an invariant manifold.
Data suggest that E is a monotonically increasing function of F and L [36]. The dependence of f on F and L is not well established and the form of f depends on multiple interrelated factors. In general, the sensitivity of various tissues to the changing hormonal milieu will have an overall effect on both the supply of macronutrients as well as the substrate preferences of various metabolically active tissues. On the supply side, we know that free fatty acids derived from adipose tissue lipolysis increase with increasing body fat mass which thereby increase the daily fat oxidation fraction, f, as F increases [37]. Furthermore, reduction of F with weight loss has been demonstrated to decrease f [38]. Similarly, whole-body proteolysis and protein oxidation increases with lean body mass [39],[40] implying that f should be a decreasing function of L. In further support of this relationship, body builders with significantly increased L have a decreased daily fat oxidation fraction versus control subjects with similar F [41]. Thus a stable isolated fixed point is consistent with this set of data.
We have shown that all two dimensional autonomous models of body composition change generically fall into two classes - those with fixed points and those with invariant manifolds. In the case of a stable fixed point, any temporary perturbation of body weight or composition will be corrected over time (i.e., for all things equal, the body will return to its original state). An invariant manifold allows the possibility that a transient perturbation could lead to a permanent change of body composition and mass.
At first glance, these differing properties would appear to point to a simple way of distinguishing between the two classes. However, the traditional means of inducing weight change namely diet or altering energy expenditure through aerobic exercise, turn out to be incapable of revealing the distinction. For an invariant manifold, any change of intake or expenditure rate will only elicit movement along one of the prescribed F vs. L trajectories obeying Equation 12, an example being Forbes's law (14). As shown in Figure 2, a change of intake or energy expenditure rate will change the position of the invariant manifold. The body composition that is initially at one point on the invariant manifold will then flow to a new point on the perturbed invariant manifold along the trajectory prescribed by (12). If the intake rate or energy expenditure is then restored to the original value then the body composition will return along the same trajectory to the original steady state just as it would in a fixed point model (see Figure 2 solid curves). Only a perturbation that moves the body composition off of the fixed trajectory could distinguish between the two classes. In the fixed point case (Figure 2A dashed-dot curve), the body composition would go to the same steady state following the perturbation to body composition but for the invariant manifold case (Figure 2B dashed-dot curve), it would go to another steady state.
Perturbations that move the body composition off the fixed trajectory can be done by altering body composition directly or by altering the fat utilization fraction f. For example, body composition could be altered directly through liposuction and f could be altered by administering compounds such as growth hormone. Resistance exercise may cause an increase in lean muscle tissue at the expense of fat. Exogenous hormones, compounds, or infectious agents that change the propensity for fat versus carbohydrate oxidation (for example, by increasing adipocyte proliferation and acting as a sink for fat that is not available for oxidation [42]–[44]), would also perturb the body composition off of a fixed F vs. L curve by altering f. If the body composition returned to its original state after such a perturbation then there is a unique fixed point. If it does not then there could be an invariant manifold although multiple fixed points are also possible.
We found an example of one clinical study that bears on the question of whether humans have a fixed point or an invariant manifold. Biller et al. investigated changes of body composition pre- and post-growth hormone therapy in forty male subjects with growth hormone deficiency [45]. Despite significant changes of body composition induced by 18 months of growth hormone administration, the subjects returned very closely to their original body composition 18 months following the removal of therapy. However, there was a slight (2%) but significant increase in their lean body mass compared with the original value. Perhaps not enough time had elapsed for the lean mass to return to the original level. Alternatively, the increased lean mass may possibly have been the result of increased bone mineral mass and extracellular fluid expansion, both of which are known effects of growth hormone, but were assumed to be constant in the body composition models. Therefore, this clinical study provides some evidence in support of a fixed point, but it has not been repeated and the result was not conclusive. Using data from the Minnesota experiment [21] and the underlying physiology, Hall [3] proposed a form for f that predicts a fixed point. On the other hand, Hall, Bain, and Chow [10] showed that an invariant manifold model is consistent with existing data of longitudinal weight change but these experiments only altered weight through changes in caloric intake so this cannot rule out the possibility of a fixed point. Thus it appears that existing data is insufficient to decide the issue.
We now consider some numerical examples using the macronutrient partition model in the form given by Equations 18 and 19, with a p-ratio consistent with Forbes's law (13) (i.e. p = 2/(2+F), where F is in units of kg). Consider two cases of the model. If ψ = 0 then the model has an invariant manifold and body composition moves along a fixed trajectory in the L–F plane. If ψ is nonzero, then there can be an isolated fixed point. We will show an example where if the intake energy is perturbed, the approach of the body composition to the steady state will be identical for both cases but if body composition is perturbed, the body will arrive at different steady states.
For every model with an invariant manifold, a model with a fixed point can be found such that trajectories in the L–F plane resulting from energy intake perturbations will be identical. All that is required is that ψ in the fixed point model is chosen such that the solution of ψ (F,L) = 0 defines the fixed trajectory of the invariant manifold model. Using Forbes's law (14), we choose ψ = 0.05(F−0.4 exp(L/10.4))/F. We then take a plausible energy expenditure rate of E = 0.14L+0.05F+1.55, where energy rate has units of MJ/day and mass has units of kg. This expression is based on combining cross-sectional data [36] for resting energy with a contribution of physical activity of a fairly sedentary person [3]. Previous models propose similar forms for the energy expenditure [5],[7],[13],[18].
Figure 3 shows the time dependence of body mass and the F vs. L trajectories of the two model examples given a reduction in energy intake rate from 12 MJ/day to 10 MJ/day starting at the same initial condition. The time courses are identical for body composition and mass. The mass first decreases linearly in time but then saturates to a new stable fixed point. The dashed line represents the same intake rate reduction but with 10 kg of fat removed at day 100. For the invariant manifold model, the fat perturbation permanently alters the final body composition and body mass, whereas in the fixed point model it only has a transient effect. In the fixed point model, the body composition can ultimately exist only at one point given by the intersection of the nullclines (i.e., solution of I = E and ψ = 0). For the invariant manifold, the body composition can exist at any point on the I = E curve (dotted line in Figure 2D). Since a ψ can always be found so that a fixed point model and an invariant manifold model have identical time courses for body composition and mass, a perturbation in energy intake can never discriminate between the two possibilities.
The time constant to reach the new fixed point in the numerical simulations is very long. This slow approach to steady state (on the order of several years for humans) has been pointed out many times previously [3],[5],[7],[13],[18]. A long time constant will make experiments to distinguish between a fixed point and an invariant manifold difficult to conduct. Experimentally reproducing this example would be demanding but if the time variation of the intake rates and physical activity levels were small compared to the induced change then the same result should arise qualitatively. Additionally, the time constant depends on the form of the energy expenditure. There is evidence that the dependence of energy expenditure on F and L for an individual is steeper than for the population due to an effect called adaptive thermogenesis [46], thus making the time constant shorter.
In this paper we have shown that all possible two dimensional autonomous models for lean and fat mass are variants of the macronutrient partition model. The models can be divided into two general classes - models with isolated fixed points (most likely a single stable fixed point) and models with an invariant manifold. There is the possibility of more exotic behavior such as multi-stability and limit cycles but these require fine-tuning and thus are less plausible. Surprisingly, experimentally determining if the body exhibits a fixed point or an invariant manifold is nontrivial. Only perturbations of the body composition itself apart from dietary or energy expenditure interventions or alterations of the fraction of energy utilized as fat can discriminate between the two possibilities. The distinction between the classes is not merely an academic concern since this has direct clinical implications for potential permanence of transient changes of body composition via such procedures as liposuction or temporary administration of therapeutic compounds.
Our analysis considers the slow dynamics of the body mass and composition where the fast time dependent hourly or daily fluctuations are averaged out for a clamped average food intake rate. We also do not consider a slow explicit time dependence of the energy expenditure. Such time dependence could arise during development, aging or gradual changes in lifestyle where activity levels differ. Thus our analysis is best suited to modeling changes over time scales of months to a few years in adults. We do not consider any feedback of body composition on food intake, which is an extremely important topic but beyond the scope of this paper.
Previous efforts to model body weight change have predominantly used energy partition models that implicitly contain an invariant manifold and thus body composition and mass are not fully specified by the diet. If the body does have an invariant manifold then this fact puts a very strong constraint on the fat utilization fraction f. Hall [3] considered the effects of carbohydrate intake on lipolysis and other physiological factors to conjecture a form of f that does not lead to an invariant manifold. However, our analysis and numerical examples show that the body composition could have an invariant manifold but behave indistinguishably from having a fixed point. Also, the decay to the fixed point could take a very long time, possibly as long as a decade giving the appearance of an invariant manifold. Only experiments that perturb the fat or lean compartments independently can tell.
The three compartment macronutrient flux balance Equations 3–5 are a system of nonautonomous differential equations since the energy intake and expenditure are explicitly time dependent. Food is ingested over discrete time intervals and physical activity will vary greatly within a day. However, this fast time dependence can be viewed as oscillations or fluctuations on top of a slowly varying background. It is this slower time dependence that governs long-term body mass and composition changes that we are interested in. For example, if an individual had the exact same schedule with the same energy intake and expenditure each day, then averaged over a day, the body composition would be constant. If the daily averaged intake and expenditure were to gradually change on longer time scales of say weeks or months then there would be a corresponding change in the body composition and mass. Given that we are only interested in these slower changes, we remove the short time scale fluctuations by using the method of averaging to produce an autonomous system of averaged equations valid on longer time scales.
We do so by introducing a second “fast” time variable τ = t/ε, where ε is a small parameter that is associated with the slow changes in body composition and let all time dependent quantities be a function of both t and τ. For example, if t is measured in units of days and τ is measured in units of hours then ε∼1/24. Inserting into Equations 3–5 and using the chain rule yields(28)(29)(30)We then consider the three body compartments to have expansions of the form(31)(32)(33)where 〈F1〉 = 〈P1〉 = 〈G1〉 = 0 for a time average defined by and T represents an averaging time scale of a day. The fast time dependence can be either periodic or stochastic. The important thing is that the time average over the fast quantities is of order ε or higher. We then expand the energy expenditure rate and expenditure fractions to first order in ε:(34)(35)where E0(t,τ)≡E(F0,G0,P0,t,τ)+O(ε2) and i∈{F,G,P}. We assume that the expenditure fractions depend on time only through the body compartments. Substituting these expansions into Equations 28–30 and taking lowest order in ε gives(36)(37)(38)
Taking the moving time average of Equations 36–38 and requiring that 〈∂F1/∂τ〉, 〈∂G1/∂τ〉, and 〈∂P1/∂τ〉 are of order ε or higher leads to the averaged equations:(39)(40)(41)In the main text we only consider the slow time scale dynamics so we drop the superscript and bracket notation for simplicity. Hence, the system (3–5) can be thought of as representing the lowest order time averaged macronutrient flux balance equations. We note that in addition to the daily fluctuations of meals and physical activity, there can also be fluctuations in food intake from day to day [23]. Our averaging scheme can be used to average over these fluctuations as well by extending the averaging time T. A difference in the choice of T will only result in a different interpretation of the averaged quantities.
The dynamics near a fixed point (F0,L0) are determined by expanding fE and (1−f)E to linear order in δF = F−F0 and δL = L−L0 [34],[35]. Assuming solutions of the form exp(λt) yields an eigenvalue problem with two eigenvalues given by where(42)and(43)A fixed point is stable if and only if Tr J<0 and det J>0. In the case of an invariant manifold, detJ = 0, so the eigenvalues are Tr J and 0. The zero eigenvalue reflects the marginal stability along the invariant manifold, which is an attractor if Tr J<0. An attracting invariant manifold implies a stable fixed point in the corresponding one dimensional model. Unstable fixed points are either unstable nodes, saddle points or unstable spirals. In the case of unstable spirals, a possibility is a limit cycle surrounding the spiral arising from a Hopf bifurcation, where Tr J = 0 and det J>0. In this case, body composition and mass would oscillate even if the intake rates were held constant. The frequency and amplitude of the oscillations may be estimated near a supercritical Hopf bifurcation by transforming the equations to normal form. Stability of a fixed point puts constraints on the form of f. Physiological considerations and data imply that ∂E/∂L>∂E/∂F>0 [3],[36]. Thus we can set ∂E/∂F = δ∂E/∂L where δ <1 (the derivatives are evaluated at the fixed point). Then detJ>0 implies that(44)and Tr J<0 implies(45)where K = [δf+γ (1−f)](∂E/∂L)/E>0 and γ = ρF/ρL≈5.2. Hence ∂f/∂F>0 and ∂f/∂L<0 guarantees stability of a fixed point. In other words, if f increases monotonically with F and decreases monotonically with L then there will be a unique stable fixed point. For an invariant manifold, f is given by Equation 15, which immediately satisfies detJ = 0; TrJ<0 is guaranteed if E is monotonically increasing in F and L. For a Hopf bifurcation, we require ∂f/∂F = γ∂f/∂L−K and Equation 44, implying (γ−δ)∂f/∂L−K>0. Since γ>δ, f must increase with L for the possibility of a limit cycle. However, to ensure that trajectories remain bounded f must decrease with L for very small and large values of L. Hence, f must be nonmonotonic in L for a limit cycle to exist. This can also be seen from an application of Bendixson's criterion [35], which states that a limit cycle cannot exist in a given region of the L–F plane if(46)does not change sign in that region. In addition, the other parameters must be fine tuned for a limit cycle (see Figure 1D). Similarly, as seen in Figure 1C), for multi-stability to exist, nonmonotonicity and fine tuning are also required. |
10.1371/journal.pcbi.1006956 | Noise-precision tradeoff in predicting combinations of mutations and drugs | Many biological problems involve the response to multiple perturbations. Examples include response to combinations of many drugs, and the effects of combinations of many mutations. Such problems have an exponentially large space of combinations, which makes it infeasible to cover the entire space experimentally. To overcome this problem, several formulae that predict the effect of drug combinations or fitness landscape values have been proposed. These formulae use the effects of single perturbations and pairs of perturbations to predict triplets and higher order combinations. Interestingly, different formulae perform best on different datasets. Here we use Pareto optimality theory to quantitatively explain why no formula is optimal for all datasets, due to an inherent bias-variance (noise-precision) tradeoff. We calculate the Pareto front of log-linear formulae and find that the optimal formula depends on properties of the dataset: the typical interaction strength and the experimental noise. This study provides an approach to choose a suitable prediction formula for a given dataset, in order to best overcome the combinatorial explosion problem.
| Sometimes a combination of drugs works much better than each drug alone. Finding such drug cocktails is a pressing challenge in order to combat drug resistance and to improve drug effects. However, it is impossible to test all combinations of multiple drug experimentally. Therefore, researchers are looking for computational rather than experimental approaches to overcome this problem. One approach is to measure the effect of few drugs and plug it into a formula that predicts the effect of many drugs together. Existing prediction formulae typically perform best on the dataset that they were developed on, but less well on other datasets. Here we explain this observation and give a guide for the choice of an optimal prediction formula for a given dataset. The optimal formula depends on two main properties of the dataset: 1) The interaction strength between the drugs and 2) The experimental noise in the data. This study may help researchers discover effective combinations of multiple drugs and multiple perturbations in general.
| Different fields of biology ask how multiple perturbations affect a biological system. For example, to understand the function of DNA sequences such as promoters or coding regions, or to design new ones, it is important to understand how mutations combine to affect function [1–4]. Another widely studied example is how multiple drugs combine to affect cells and organisms. This question is important for developing effective combination therapy [5–9] and to reduce drug resistance [10–14].
A major challenge in these fields is the combinatorial explosion problem: the number of combinations increases exponentially with the number of perturbations. Given n different single perturbations, there are 2n possible combinations. In DNA sequences there are 4n combinations of bases so that sequences of 30bp have 1018 possible combination. Drugs present the additional dimension of doses, so that 8 drugs at 6 doses amount to 68≈106 combinations. Therefore, the number of combinations quickly outgrows experimental ability.
To overcome the combinatorial explosion problem, there are two main approaches. In the case of sequences, one can use directed evolution to find sequences with desired function [15–19]. This approach is powerful and is based on exponential expansion of the sequences with highest function. However, experimental evolution still covers only a tiny fraction of sequence space and has the potential to get stuck on local optima. In the case of drugs this approach is not feasible.
The other main approach is to use mathematical models to estimate the effects of combinations using only a small number of measurements. Machine learning studies use regression-like models to estimate drug and mutation effects based on a learning set of measurements [20–24]. For example [4] analyzed combinations of mutations on the lac promoter, and [25] analyzed a library of mutation in green fluorescent protein. One limitation of machine learning is that it requires extensive training data, which may exceed experimental ability when samples are rare and perturbations are costly, as in the case of drug combinations.
To overcome the lack of large training datasets, another line of research establishes analytical formulae to estimate combination effects based on, for example, measurements of single perturbations and pairs. Analytical formulae can include knowledge about the biology of the system and can therefore be more effective than machine learning when data is scarce. The most common baseline model, that seems to work well as a first approximation in many cases, is Bliss independence [26] in which the effect of a pair of perturbations is the product of the single perturbation effects, sij = sisj. Bliss independence is equivalent to additivity in log-effect space. Another baseline model for drugs is Loewe (dose additivity) [27], but seems to be less accurate than the Bliss approximation for high-order drug combinations [28,29].
Baseline models are generally inaccurate because they do not consider the interactions between perturbations. These interactions are called synergy and antagonism, in the case where the combination shows larger or smaller effect than the baseline model, respectively. Several studies have attempted to present formulae that take interactions into account, by including measurements for pairs. Wood et al. [30] introduced an Isserlis-like formula based on singles and pairs. For triplets, the formula is s123 = s1s23+s2s13+s3s12−2s1s2s3. This formula worked well for combinations of up to 4 antibiotics.
Zimmer et al [31] presented a model which used measurements of dose-response for single drugs and drug pairs to compute the dose-dependent effect of higher order combinations, with excellent accuracy on antibiotics and anti-cancer drugs. An additional formula, based only on pairs, performed well on small single-dose drug datasets [32].
Surveying these studies, it seems that there is no best formula that outperforms others on all datasets. Instead, each formula works well on the dataset it was developed on, but typically less well on other datasets. This situation suggests that, because datasets differ in their noise and interaction strengths, there may be a range of formulae to consider. There is therefore a need to compare formulae, to understand when formulae fail, and to develop ways to decide which formula to use when considering a given dataset.
Here, we address these questions by studying the tradeoffs inherent in formulae for combinations. We study wide classes of formulae and test them on twelve experimental datasets for drugs and sequences, as well as on synthetically generated datasets. We find that no formula outperforms the others on all datasets. Instead, each dataset has a different optimal formula. On the other hand, many formulae are suboptimal for all datasets.
We explain this result using a well-known concept from statistical learning, the bias-variance tradeoff [33–35]. Roughly speaking, good formulae should be complex or expressive enough to capture the true variability of the dataset (low bias). On the other hand, formulae should be simple enough in order to avoid overfitting the noise in the dataset (Fig 1). Hence, the optimal formula for a dataset should be dependent on the typical effect size (true variability) of the dataset as well as the experimental noise.
In order to understand this tradeoff, we use Pareto optimality [36–39]. Pareto optimality was previously used to study model selection and hyper-parameter choice in machine learning models [40,41]. We use it to define the optimal formula for each dataset, based on its noise and interaction strength. We suggest a method to choose the optimal formula for a new dataset.
For simplicity, we concentrate on the problem of predicting the effect of triplets of perturbations from data on the effects of pairs and single perturbations. We provide formula for the effects of k perturbations in the supporting information (S1 Text, [29,42]).
To establish notation and terminology, we use the term perturbation as a general term for drug, mutation or other type of change in the system. We define the effect as the measurable outcome of the perturbations on the system function, such as survival of cancer cells for anti-cancer drugs, growth rate of bacteria in case of antibiotics, and the activity of a promoter or a protein in the case of sequence mutations.
Three different perturbations will be denoted by 1,2,3. The value of the effect in the absence of perturbation (wild-type) is S∅. The effects of single perturbations are S1,S2,S3, of pairs of perturbation are S12,S13,S23. The effect of the triplet perturbation, which we wish to predict given singles and pairs effect data, is S123. For the effects normalized by the wild-type we use lower case letters sx=SxS∅
Formulae from the literature include the Bliss independence formula:
s123=s1s2s3
(1)
Machine learning approaches often use a regression formula:
s123=s12s13s23s1s2s3.
(2)
This formula results from regression where one fits the effects of singles and pairs to s = ∑iaixi+∑i,jaijxixj where xi = 0 if mutation i is absent and xi = 1 if it is present.
A third formula uses only information from pairs [32]:
s123=s12s13s23.
(3)
These formulae belong to the class of log-linear formulae, and hence we focus on this class. The most general formula in this class, taking into account the symmetry in perturbation indices (re-naming drugs 1, 2 and 3 should not affect the prediction for S123) is:
S123=S∅α(S1S2S3)β(S12S13S23)γ
To make the calculation linear, we use the logarithm of the un-normalized effects Lx = log(SX), resulting in
L123=αL∅+β(L1+L2+L3)+γ(L12+L13+L23)
The log-linear formulae thus have three parameters, α,β and γ. They include the previous formula discussed above: Bliss independence is when α = −2,β = 1,γ = 0 and regression is α = γ = 1,β = −1.
We now evaluate the precision of each formula. As an operational definition of precision, we use a Taylor-series approach. We assume that the log effect is a smooth function f of multiple inner variables of the system. Each perturbation is represented by a change in one of these inner variables.
Without loss of generality, we can assume that without perturbations, L∅ = f(0,0,0). Then L1, the log effect of perturbation 1, is L1 = f(x,0,0), for some value of x. Similarly, the other two single perturbations are L2 = f(0,y,0) and L3 = f(0,0,z). The pair log effects are L12 = f(x,y,0),L13 = f(x,0,z),L23 = f(0,y,z). To predict the triplet, we need to estimate L123 = f(x,y,z). Mathematically, this is equivalent to the question of estimating a function on one vertex of a 3D box given its values on the other 7 vertices [42].
Even though in reality perturbations are sometimes not small, we will next assume that they are in order to give an operationalized and analytically solvable way to discuss precision. When the values of x,y and z are such that they represent small perturbations, one can use a Taylor expansion and ask which of the formulae are precise to which order of expansion (no matter what the exact form of f).
Here we will derive conditions for a formula to be precise to the 0th, 1st and 2nd orders in Taylor series. But first we explain intuitively what these precisions orders mean. Formulae precise to 0th order have the property that if all effects are equal, L∅ = Li = Lij = C the prediction for the triplet is equal to that effect: L123 = C. Formulae accurate to first order have the property that if all pairs are Bliss independent in the sense that sij = sisj, then the predicted triplet is also Bliss independent s123 = s1s2s3.
We now derive the conditions for precision to different orders. The Taylor expansion of L123 is, to first order:
L123=f(x,y,z)=f(0,0,0)+∂f∂x(0,0,0)x+∂f∂y(0,0,0)y+∂f∂z(0,0,0)z+o(|x|,|y|,|z|)
We equate this to the Taylor expansion of the log-linear formula:
αL∅+β(L1+L2+L3)+γ(L12+L13+L23)==αf(0,0,0)+β[f(x,0,0)+f(0,y,0)+f(0,0,z)]+γ[f(x,y,0)+f(x,0,z)+f(0,y,z)]==(α+3β+3γ)f(0,0,0)+(β+2γ)[∂f∂x(0,0,0)x+∂f∂y(0,0,0)y+∂f∂z(0,0,0)z]+o(|x|,|y|,|z|)
We therefore obtain the condition for a formula to be precise to 0th order:
1=α+3β+3γ
From now on, we restrict ourselves to the class of models that are precise to 0th order. We next ask which models are precise to 1st order. The condition for 1st order precision is:
β+2γ=1
All the formulae on this line in beta-gamma space give exact approximation to the first order. For example, the Bliss (β = 1,γ = 0), the regression (β = −1,γ = 1) and pairs formula (β=0,γ=12) fall on this line of first order precision.
We can define the deviation from 1st order precision as follows:
P1st(α,β,γ)=(1−β−2γ)2
We next ask which formulae are precise to the second order. We find that there is only one log-linear formula which is precise to the second order–the regression formula of Eq 2 (S1 Text) (α = γ = 1,β = −1). The deviation of other formula from second-order precision can be represented by the sum of the coefficients of the second order error (S1 Text):
P2nd(α,β,γ)=(12−β2−γ)2+(1−γ)2
The precision findings are summarized in (Fig 2A and 2B). The figures plot contours of accuracy to different orders as a function of β and γ. In the plots, α is evaluated by the zero-order precision demand α = 1−3β−3γ. The plots are therefore restricted to 0th order precise formulae. It is seen that optimal first-order accuracy occurs on a line in model space which includes the Bliss and regression models, and that second-order precision has elliptical contours maximal at the regression model.
If precision was the only factor at play, one would expect the regression model to outperform others. However in most real datasets this model does poorly [31]. The reason is that it is sensitive to experimental noise. To estimate the robustness to noise of different models, we model experimental noise in the measured effects, Li = Li+χi and Lij = Lij+χij, where χ are independent Gaussian noise with equal STD σ for all measurements (similar conclusions apply to the case of non-independent noise, S1 Text). This corresponds to log-normal multiplicative noise for the effect measurements. Such log-normal noise is typical for experiments on drug and mutation effects [31,32].
Here we derive an expression for the noise in the predicted triplet effect. We must separate between two cases. Case I occurs when there is experimental noise in L∅ (the wild-type), as is the typical case for sequence (mutation) data, so that L∅ = L∅+χ∅. Case II is when L∅ is noiseless, as often happens for drug combinations when the effect is cell survival and L∅ = 0 by definition.
To compute the variation in the prediction of a triplet s123 given the noise in the pair and single inputs, we assume independent noise for each variable. The noise (std) for case I depends on the three parameters of the model α,β and γ:
PN,WTnoise(α,β,γ)=σα2+3β2+3γ2
And in case II (noiseless L∅) only on the parameters β and γ:
PN,WT=1(α,β,γ)=σ3β2+3γ2
Note that noise is minimal when α = β = γ = 0, a formula that always predicts 0. This model is not precise even to 0th order. Considering only models precise to 0th order, we obtain the minima of the noise performance function in case of noisy wild type (S1 Text):
argmin(PN,WTnoise(α,β,γ))=(17,17,17)
Which simply averages the inputs L∅,Li and Lij, and in case of noiseless wild-type simply taking the wild-type value S∅ as the prediction
argmin(PN,WT=1(α,β,γ))=(1,0,0)
Contours of this function in the cases of presence and absence of noise in the wild-type appear in (Fig 2C and 2D). In both cases noise grows with distance from the single minimum.
In order to compare models according to the two tasks, precision and noise, we use the Pareto front approach. The Pareto front is defined as the set of formula for which there is no other formula that is better at both tasks. Given the two performance functions of noise and precision, we compute the Pareto front as the set of points of external tangency of the performance contours [43,44]. The resulting front is a one-dimensional curve in the space of formulae (beta-gamma space). In the case of first-order precision and noise robustness, the front is a straight line.
In the absence of wild-type noise the Pareto front is defined by (see S1 Text and Fig 3A):
γ=2β
Or in the presence of wild-type noise (see S1 Text and Fig 3B):
5β+2γ=1
If noise and first-order precision are the only tasks faced by formulae, it is expected that all optimal formulae will fall on this line.
We next computed the Pareto front where the two tasks are noise robustness and second order precision. In the case of noiseless wild-type, this give the conic defined by the equation (see S1 Text and Fig 3C):
−2γ2+7βγ+2β2+γ−6β=0
In the case of noisy wild-type we find (see S1 Text and Fig 3D):
5β2+28βγ−22β+16γ2−20γ+5=0
It is now possible to compute the entire Pareto front which consists of optimizing the three performances together. The boundary of the Pareto front is defined by the Pareto fronts of the pairs of tasks. The entire Pareto front in the cases of noiseless and noisy wild-type is composed of two thin triangle-like shapes that meet at a vertex, as shown in Fig 4A and 4B (black region).
We note in passing that the typical solution for a Pareto front with three tasks resembles a single triangle with the optima for the three tasks at the three vertices[43,44]; the elongated two-shape pattern found here results from the fact that the optima for one task, first order precision, falls on a line and not a single point.
The present approach can be applied to any class of formulae. To illustrate this we compute the Pareto front for a class of generalized mean formulae in S1 Text.
In order to test the relevance of the Pareto front to real data, we compiled a set of thirteen published experimental datasets for drugs and mutations (Table 1). This includes data on the effects of drugs (antibiotics, cancer drugs) on cells, and the effect of mutations on proteins and organisms. The datasets include the effects of singles, pairs and triplets of perturbations. For each dataset, we scanned formulae (scanning β and γ with α = 1−3β−3γ to provide 0th order precision) and found the formula that gives the smallest root-mean-square error for triplet predictions. This formula, a point in the β,γ plane, is the optimal formula for that dataset. In order to control for outliers and variation in the data, we repeated this for each dataset on 30 bootstrapped datasets, in which we built a new dataset sampled from the original data with replacements. Thus, each dataset yields 30 additional optimal formula points.
We find that the optimal formula for all datasets lie close to the Pareto front (Fig 4A). The large datasets fall neatly on the Pareto front (E. coli antibiotics 1, A549 and others), whereas smaller datasets tend to deviate more due to their larger bootstrapping variance (Dihydrofolate reductase, H1299, E. coli antibiotics 2). Note also that the main direction of variability of the bootstrapping distribution is parallel to the Pareto front [45].
In the presence of noise in the measured wild-type effect (case II above), the datasets also fall on the Pareto front (Fig 4B). In this case the datasets are larger, hence they have less variability in the bootstrapping. Here, we used an expansion trick to increase the amount of usable data from small fully-factorial datasets. In the expansion trick, we consider treatment with a single perturbation Li as wild-type L∅. We then consider treatments with an additional second perturbation Lij as a single perturbation on the wild-type background, Li, treatments with three perturbations Lijk as the pairs Ljk, in order to predict the triplet Ljkm given by the quadruplet Lijkm in the original data (Fig 5). We also used pairs and higher order combinations as wildtype to the extent allowed by the dataset. This increases the number of triplets in the fully factorial dataset of order k from (k2) to at most (k2)2k in its most extended form (Table 1 shows both original and expanded triplet number).
We further tested 61 datasets from the UniProbe database [46] on protein-DNA binding interactions in fully factorial datasets of 8 mutations. We use the expansion trick using 1000 randomly chosen starting point sequences as a wild-type from each fully factorial dataset. We find that the optimal formulae for these datasets all fall close to the Pareto front (Fig 4C). The results are near the noise-robustness archetype, suggesting that noise is a dominant source of variation in these protein-binding microarray experiments.
We see that optimal formulae for different datasets are close to the Pareto front. We next asked how the properties of the dataset affect which formula is optimal for that dataset. To do so, we generated synthetic datasets with different parameters, so that we could control the level of noise and the level of interaction strength (deviation from the Bliss formula, see S1 Text), the two factors that influence the performance of the formula.
To generate simulated data we used a third order polynomial f(x,y,z) with random coefficients, sampled at different random points, with Gaussian noise added (which varies between datasets). The goal is to predict triplets from pairs and singles, that is to predict f(x,y,z) from the projections on axes and planes (S1 Text) e.g. f(x,0,0), f(x,y,0) etc. The noise amplitude of each dataset is the standard deviation of the Gaussian noise added to log effect. The interaction strength (that is synergy/antagonism) of each dataset is given by its mean deviation from the Bliss approximation I=|sij−sisj||sisj|. To control I, we sampled the function at various distance from the origin (S1 Text), where the larger x y and z, the larger the nonlinearity and hence the interaction.
For each such synthetic dataset, we computed its optimal formula among the log-linear family and found that for datasets with small interaction strengths, the optimal formula falls close to the curve defining the Pareto front (Fig 6A). Interestingly, when interaction strength become larger, points go a bit beyond the second order precision archetype (Fig 6A, solid arrow), and when interaction strength was increased even further, points start to go back to the (0,0) point deviating from the Pareto front (Fig 6A, dashed arrow).
To see the trends described above we plotted the optimal values of γ and β as function of noise and interaction strength (Fig 6B and 6C). We start by considering the region above the solid arrow (small interaction strength), we see that in this region γ increases and β decreases with interaction strength. This is the expected result since larger γ and smaller β means getting closer to the second-order-precision archetype. Second-order precision becomes more important relative to noise as interaction strength gets larger, noise robustness becomes more important than second order precision when the noise in dataset is larger. These results summarize the prediction of Pareto optimality theory.
Interestingly, there are simulated datasets for which the optimal formulae are beyond the second order archetype (Fig 6A, right side). It was found that formulae of second order tend to approximate higher order interactions better than expected [47]. The points beyond the Pareto front are example of formulae of second order which are especially better in predicting the value of the third order function.
In the region of large interaction strength (Fig 6B and 6C over the dashed arrow and, Fig 6A dashed arrow), we see the opposite trend of decreasing γ and increasing β with interaction strength. The explanation of this surprising result is that formulae in this region no longer satisfy the assumption of precision to 0th order. The interaction strengths in this case are so large, such that the Taylor approximation approach no longer gives the optimal formulae. Fig 6D shows that indeed the formulae found for higher interaction strength no longer gives predictions which are accurate to the 0th order.
These results indicate that one can predict the optimal formula for a dataset if one can estimate its noise and interaction strengths.
One general question is to what extent high order interactions exist in biological systems that can’t be explained by pairs. High order interactions in this context are defined as the deviation of the measurement form a null model that includes the effects of single and pair perturbations. Thus, choice of null model can affect the results.
For example, a standard definition of pairwise interaction is:
ϵ12=s12−s1s2
This formula is based on a Bliss independence null model for the combined single effects: s12 = s1s2.
Different studies of triplets use different null models [48,49]. For example, a recently study measured the effects of about 150,000 triple gene deletions in yeast, and compared them to single and pair deletions [49]. Third-order interactions were estimated using an Isserlis null model (s123 = s1s23+s2s13+s3s12−2s1s2s3) yielding
ϵ123=s123−s1s2s3−ϵ12s3−ϵ13s2−ϵ23s1
Evaluating the triplet interaction using the absolute value of ϵ123 as defined above gives a mean absolute triplet interaction of 0.044. Significant triplet interactions were estimated to be about 100 times more common than significant pair interactions.
We used the present approach to predict the optimal model using the noise and effect size in the pair measurements in this study. The best null model is similar to the pairs model (Eq 3), which is less noise-prone than the Isserlis model. With this null model, the mean absolute triplet effect is 23% lower.
In this study, we find that the problem of predicting the combined effects of perturbations does not have a unique optimal solution. Instead, different solutions and formulae are optimal for different datasets. We analyze the Pareto front of models that trade-off noise robustness and precision. This Pareto front of optimal formulae matches observations on the best formula for a range of real and synthetic datasets.
The present study offers a way to predict the best formula based on the noise and effect size of pairs data. By measuring interaction strength based on pairs, and experimental noise using repeats, one can judge where on the Pareto front the optimal formula might lie for a given dataset.
One important use of these formula is to estimate high-order effects between genes. For example, a third-order effect ϵ123 is defined by the measured effect of three perturbations minus a null model based on single and pair perturbations. The better the null model, the more accurate the estimation of the high-order effect. We find that the present approach can improve the null model used for estimating the effects of triple yeast gene deletions in a recent large scale study (Kuzmin,2018). The improved estimation lowers the number of apparent three-gene interactions that can’t be explained by pairs. This is relevant for the design of systematic gene perturbation experiments, because it indicates that pairs may be enough to capture most of the interactions. Pair scans are much more feasible than triple-perturbation scans, suggesting an optimistic outlook for understanding complex gene interactions.
This study used a Taylor expansion to define precision. Taylor series strictly apply only to small perturbations. Despite this limitation the method seems to work well for mutation and drug combination dataset. One reason for this is that higher-order effects in biological systems seem to be smaller than low order ones [29], which is equivalent to the underlying assumption of the Taylor approximation. It would be fascinating to use the present approach to analyze additional classes of formula, and to understand the effects of multiple perturbations on additional biological systems.
Computations of the maxima of the different performance functions (Fig 2), and the Pareto fronts of multiple performance functions (Figs 3 and 4) were performed analytically, and are detailed in the result section and S1 Text.
All simulations and Figs 1–4 and 6 were produced using MATLAB 2017.
Evaluation metric for formulae performance was RMSE. Therefore, the coefficients of the optimal formula were computed using linear regression on a dataset (Figs 4 and 6).
In Fig 6, simulated data was generated using random symmetric polynomials of degree 3 according to the formula:
f(x,y,z)=a0+a1(x+y+z)+a2(x2+y2+z2)+a3(x3+y3+z3)+a4xyz+a5(xy+xz+yz)+a6(x2y+y2x+x2z+z2x+y2z+z2y)
Where ai were sampled randomly and uniformly between 0 and 1. To get the simulated dataset such random formulae were evaluated at random points in the box [0,ϵ]×[0,ϵ]×[0,ϵ]. The approximation distance ϵ varied logarithmically between [0.06,0.06∙29]. To the synthetic dataset we added random log-normal noise N(0,σ), where σ varied logarithmically between [0.01,0.01∙29]. Each point in Fig 6 is based on average of 10 different simulated dataset generated from 10 different random functions, each simulated dataset consists of 300 points.
|
10.1371/journal.pcbi.1002373 | Joint Analysis of Multiple Metagenomic Samples | The availability of metagenomic sequencing data, generated by sequencing DNA pooled from multiple microbes living jointly, has increased sharply in the last few years with developments in sequencing technology. Characterizing the contents of metagenomic samples is a challenging task, which has been extensively attempted by both supervised and unsupervised techniques, each with its own limitations. Common to practically all the methods is the processing of single samples only; when multiple samples are sequenced, each is analyzed separately and the results are combined. In this paper we propose to perform a combined analysis of a set of samples in order to obtain a better characterization of each of the samples, and provide two applications of this principle. First, we use an unsupervised probabilistic mixture model to infer hidden components shared across metagenomic samples. We incorporate the model in a novel framework for studying association of microbial sequence elements with phenotypes, analogous to the genome-wide association studies performed on human genomes: We demonstrate that stratification may result in false discoveries of such associations, and that the components inferred by the model can be used to correct for this stratification. Second, we propose a novel read clustering (also termed “binning”) algorithm which operates on multiple samples simultaneously, leveraging on the assumption that the different samples contain the same microbial species, possibly in different proportions. We show that integrating information across multiple samples yields more precise binning on each of the samples. Moreover, for both applications we demonstrate that given a fixed depth of coverage, the average per-sample performance generally increases with the number of sequenced samples as long as the per-sample coverage is high enough.
| Microorganisms are extremely abundant and diverse, and occupy almost every habitat on earth. Most of these habitats contain a complex mixture of many different microorganisms, and the characterization of these metagenomic mixtures, in terms of both taxonomy and function, is of great interest to science and medicine. Current sequencing technologies produce large numbers of short DNA reads copied from the genomes of a metagenomic sample, which can be used to obtain a high resolution characterization of such samples. However, the analysis of such data is complicated by the fact that one cannot tell which sequencing reads originated from the same genome. We show that the joint analysis of multiple metagenomic samples, which takes advantage of the fact that the samples share common microbial types, achieves better single-sample characterization compared to the current analysis methods that operate on single samples only. We demonstrate how this approach can be used to infer microbial components without the use of external sequence data, and to cluster sequencing reads according to their species of origin. In both cases we show that the joint analysis enhances the average single-sample performance, thus providing better sample characterization.
| Metagenomic samples are pooled samples of the genomes of multiple microorganisms living in the same environment. They can be taken either from the outer environment or from microbial populations colonizing other living organisms. Metagenomic studies focus on the taxonomic and functional characterization of the microbial populations contained in such samples. These studies have been boosted by advances in Next Generation Sequencing (NGS) technologies. Particularly, Whole Genome Shotgun (WGS) sequencing provides reads sampled randomly along the genomes, and enables simultaneous phylogenetic and functional analysis of the samples. Although WGS datasets contain plenty of information, they are hard to decipher, as we will further explain below. In a nutshell, the natural way to explore their composition is by aligning the sequencing reads against known databases of whole genomes or of marker genes, however these databases are seriously limited and biased. In addition, one cannot a-priori tell which reads originated from the same genome, and therefore many methods attempt to cluster the reads according to species of origin as a preliminary stage; unsupervised binning methods face an especially hard challenge, and are currently practiced mostly on extremely simple or simulated datasets.
Along with the increasing availability of single-metagenome WGS datasets, datasets consisting of multiple metagenomic samples are also becoming abundant. These datasets typically include samples taken from similar environments, such as ocean water sampled from different locations or depths [1], or microbiomic samples taken from a group of human individuals [2]. To date, the primary analysis of the resulting sequences is performed separately for each sample. Our principal observation is that combining information from multiple samples improves the characterization of each of the samples. We give two demonstrations of this principle: First, we present a method for the unsupervised characterization and quantification of shared hidden components across samples. Second, we present a binning method that operates on multiple samples simultaneously in order to achieve high per-sample precision.
We consider an unsupervised learning approach, in which we aim at learning the shared components of the different samples in an attempt to answer the prominent question of metagenomics, “what's in the mix”, without relying on any prior knowledge. While the use of stored sequences of whole genomes [3] or of marker genes, such as the 16S rRNA subunit [4], is currently the most effective way of analyzing large-scale metagenomic samples, it is considerably hindered by the incompleteness of existing databases: In addition to including only a small fraction of the species expected to be found in the samples, the set of species which these databases do include is highly biased, and this bias in turn causes a bias in the analysis results. Supervised analyses also often assume that the properties of the samples which are of biological or medical interest correspond to known taxonomic or functional annotations, although this is not necessarily the case. An intriguing counter-example is the recently discovered enterotype classes [5], which are three robust classes to which human gut metagenomic samples can be classified. Although generated using a supervised technique, these classes are characterized by a complex combination of the abundance of many bacterial species, which do not correspond to specific taxonomic units.
Aiming to avoid these disadvantages, we developed a method for the inference of hidden components within the data, which leverages on the fact that these unknown components are shared by the different samples. Each of the components is characterized by its sequence composition pattern, specifically the frequency of different short -mer words in the sequence, which is known to characterize bacteria at different phylogenetic scales [6], [7]. Due to the unsupervised nature of the method, we do not expect the components to represent any easily-interpretable biological entity, but instead to provide a composite characterization of the samples. Unlike the enterotypes clustering procedure, our method does not require an alignment stage and does not classify the samples to distinct classes. Instead, we search for the best components that explain the data, and each sample is assigned a distribution over these components; this is done by utilizing Probabilistic Latent Semantic Analysis (PLSA) [8], a technique applied to fields such as information retrieval and natural language processing. Despite these differences, there are some correlations between the inferred components and the enterotypes, which we mention in the Discussion.
Unsupervised component estimation can be used for multiple purposes, and we choose to demonstrate its applicability to a new paradigm for studying statistical association between metagenomic content and phenotypes, which we now introduce. We look for DNA words - long -mers - whose abundance in the sequencing reads of the different samples correlate with the phenotype at question. For large enough , differences in the abundance of certain -mers would capture differences in the abundance of specific species, genes or functional domains which cause the phenotype or are affected by it.
The proposed framework is analogous to the widely used paradigm of Genome Wide Association Studies (GWAS), which is used to test for associations between genetic variants in the human genome and phenotypes. In a typical GWAS the frequency of millions of variants, spanning the entire genome, is compared between a group of cases and a group of controls, and variants whose frequencies differ significantly between the two are considered to be statistically associated with the condition. In the context of metagenomic association, the -mers are analogous to the genetic variants studied in GWAS, and in both applications the goal is to find statistically significant associations between the measured variants and the condition. However, while GWAS searches for specific mutations which are associated with increased risk for the condition, we aim to capture modifications in the bacterial composition - functional or taxonomic - which are associated with the disease. As in the case of GWAS, the advantages of our approach are its computational efficiency, statistical rigor, cross-study comparability, and the fact that it does not require a supervised stage or comparison to existing databases.
Interestingly, when testing this approach on a publicly available dataset [2] containing 124 deeply sequenced samples of human gut microbiomes collected as part of the MetaHIT (Metagenomics of the Human Intestinal Tract) project, we found that the abundance of a large fraction of the -mers vary with some of the phenotypes, even for as small as 3. In the GWAS context this is known as a case of stratification: the null hypothesis of equal distribution between phenotype groups does not hold for the typical variant. For example, when the case and control groups have different ethnic composition, the minor allele frequency of an exceptionally large number of Single Nucleotide Polymorphisms (SNPs) may appear correlated with the disease, but these correlations reflect the fundamental genetic difference between the groups, instead of being relevant to the disease.
In order to correct for the stratification and conduct a proper association analysis, we integrate into the association test the estimates provided by the probabilistic model, specifically the estimated proportion of each component within each sample. We chose to characterize the components according to the short -mers distribution in the samples. Recently, Meinicke et al. [9] propose to model the -mers distribution of a single metagenomic sample as a mixture over the distributions of already-sequenced genomes; however, the use of multiple samples in our method allows our method to remain unsupervised.
As a second demonstration of the joint analysis approach we consider the task of binning sequence reads into an unknown set of species. Binning is an important preliminary step for further metagenomic analysis, and has been heavily investigated in the past few years, including the development of multiple unsupervised methods [10]–[16]; however, all existing methods operate on single samples only. We suggest an unsupervised coverage-based approach, and demonstrate that when the samples share a common species core, information can be integrated between them to improve binning precision. In other words, if one wishes to bin a given sample, then the simultaneous binning of other samples would yield better precision for the original sample. Moreover, we show that for a fixed depth of coverage, dividing the sequencing reads between additional related samples improves precision on the sample of interest.
Over the last few years, there have been many reports of associations between the content of metagenomic samples, especially human microbiomic samples, and phenotypes. Different studies report associations with different properties of the samples, such as the abundance of certain taxonomic units, mostly phyla and species, the overall taxonomic and functional diversity of the samples, and the abundance of certain genes or groups of genes, such as those participating in specific metabolic pathways (see Turnbaugh et al. [17] for a comprehensive study of obesity which tested most of these properties). In addition, dimensionality reduction techniques such as PCA [18] are often used on top of the raw data. While examining many properties of the samples allows to capture a wide range of associations, it is not always possible to accumulate results over different studies, in order to perform meta-analysis. In addition, it is hard to perform a rigorous statistical analysis, and especially to control for multiple hypotheses, when many different types of tests are carried out.
As a more rigorous approach, we propose to test the abundance of all possible DNA words of a fixed length for association with the phenotype. This test examines a limited but well-defined group of variants, and hence while it is not expected to capture the entire spectrum of possible associations, its results are statistically robust and easy to compare across studies and accumulate for future meta-analysis studies. It does not require alignment or comparison against any existing database, and therefore it can capture associations with unannotated sequences; due to the latter, the test is also fast and easy to implement.
Formally, for a given value of , the number of occurrences of all -mers across each of the samples are normalized to obtain sample-specific relative abundances. The counts of complementary -mers are summed together as they are indistinguishable in the sequencing data. We denote by the relative abundance of -mer in sample , and by the phenotype of sample . We test the association between the -mer and the phenotype by fitting a regression model of the form(1)For a given phenotype we solve the model for each -mer by generating the appropriate vector , where is the number of samples. We use simple regression and logistic regression for continuous and dichotomous phenotypes respectively.
In the Results section we report that some phenotypes are correlated with a large fraction of the -mers. These correlations reflect large-scale differences in the genetic composition of the samples between the phenotype groups; specifically, a plausible assumption is that there exists a group of common microbial components, and that each sample is a mixture of these components, in unique proportions. The components might be, for example, different bacterial phyla, and a certain phenotype might correlate with a higher proportion of a certain phylum; since there are differences in sequence composition between the phyla, this would cause phenotype-correlated differences in the distributions of many -mers. However, we are interested not in the large-scale variation, but in the -mers which remain correlated with the phenotype after taking this variation into account. Assuming there are components and denoting by the proportion of component in sample , , the estimation of the matrix would allow us to construct a corrected model:(2)
This equation is similar to equation 1 but includes the additional confounding components as covariates. For a given phenotype , we again solve the model for each k-mer while keeping the covariate expressions fixed. Due to this addition, the association of a -mer whose association with the phenotype is explained by the covariates will not be statistically significant, as desired. Note that is not included in the equation because of the linear dependency .
In order to estimate we use the following probabilistic model. We assume that the sequencing reads for metagenomic samples are given, and that the DNA content of the samples is composed of a common set of components; each read has originated from one of the components, and each component is characterized by a typical distribution over the group of all possible -mers in the sequence, for some fixed small value of (e.g., ). The model is parametrized by two row-stochastic matrices, and : the th row of , denoted , defines a sample-specific multinomial distribution over all components, and the th row of , denoted , defines a component-specific multinomial distribution over , the group of all -mers. When we sample a short -mer from a random position on the reads of sample , we first sample a component according to the distribution , and then sample the -mer according to the distribution . Being defined as general multinomial distribution, some entries in and may have a zero value; in particular, some components might not be represented in some of the samples.
We note that while the previous subsection discussed long -mers (e.g., ), which are each tested for association with the phenotypes, in this section we use short -mers as characteristics of the components we attempt to learn. Specifically, we chose to use = 4 since it has been shown that 4-mer distributions are characteristic of phylogenetic units [6], [7], and since the 4-mer distribution captures both the codon distribution and possible codon biases.
We now turn to calculate the likelihood of the metagenomic data. Since the model explains the -mer distribution in the reads, we extract the first -mer from each read, and denote by the number of times -mer was extracted from sample . The likelihood of the counts data R is(3)where is the number of samples, is the number of components, and is the group of all possible -mers, as the counts for complementary -mers are joined. Our goal is to estimate the distributions in the matrix , which are the distributions over the components for each sample.
We note that there is a simple relation between and the above notation, given by . Furthermore, under the model assumptions, we have that . One can verify that if there is a solution such that , then this solution maximizes the likelihood in Equation 3. Thus, we can view the maximization of the likelihood function as an approximation of the factorization of the matrix, which is row-stochastic, into two other row-stochastic matrices:This factorization corresponds to a set of linear equations, to which each additional sample adds variables but also a much larger number of equations - ; this could serve as an intuition for the advantage conveyed by sample multiplicity. In addition, this factorization is a variant of the non-negative matrix factorization (NMF) technique, with the added stochasticity constraints (so that the sum of each row in and is 1). NMF is used to unveil hidden structures within data, and its major advantage over methods such as the widely-used PCA is the high interpretability of the inferred components [19]. A recent paper [20] used NMF in a metagenomic context, however the factorized data matrix was generated using alignment to sequenced genomes, in contrast to our method which does not rely on prior knowledge. The stochasticity constraints turn our model to an exact instantiation of PLSA [8], [21], a generative model from the statistical literature. PLSA was originally applied to the field of text analysis for the discovery of topics in a corpus of documents [22]. Due to its great flexibility, it was successfully applied to multiple problems in the field of text learning [23]–[25] as well as to image content analysis tasks [26].While strong similarities exist between PLSA and NMF, the fact that PLSA is based on a probabilistic model allows us to refine the model to better match the properties of the sequencing data, as we do below.
In the above procedure we extract only the first -mer from each read because the model assumes that the -mers are sampled independently according to , conditioned on the read's component. Extracting multiple -mers would result in a deviation from the model due to the dependencies between neighboring -mers on the same read, as well as dependencies between -mers extracted from multiple reads covering the same genomic region.
Interestingly, the simulations presented in the Results section demonstrate that extracting multiple -mers from the same read improves performance, despite the dependencies. The reason is that under reasonable coverage and when is not too large, the relative abundances approach a constant value, and the exact sampling strategy has no effect on the final counts data. It is therefore of benefit to extract multiple -mers from each read when processing the sequencing reads, however it turns out that the best strategy is to choose not all -mers present on the read but only a subset, while using a slightly different model, as we explain next.
We use Expectation-Maximization algorithms in order to approximate the maximum likelihood solutions of both the original model (defined in Equation 3) and the refined model (Equation 4), beginning with the refined. The observed variables are groups of extracted -mers, one group for each read, and the latent variables are the assignments of a component to each of the reads. Let be the unknown assignments, and let be the number of occurrences of -mer in the multiset of -mers extracted from read , denoted . The algorithm can now be written as
E-step:where
M-step:
The running time of each iteration of the above algorithm is , being the number of -mers extracted from each read. For realistic values of this is time consuming. In addition, for large datasets the entire data cannot fit in memory. Consider, for example, the case where the number of individuals is , the number of reads per individual is , and all non-overlapping -mers from reads of length bp are used. In this case, the amount of memory required is at least GB, even if every nucleotide letter is stored in two bits. We note that by changing the order of the summation, one can use a considerably smaller amount of memory, however in each iteration of the EM the entire dataset will have to be read again to memory. Therefore, when analyzing large datasets we recommend to use the simpler model, described by Equation 3, which ignores the relation between -mers on the same read. The input counts are extracted in a single pass through the data. The latent variables are the assignments of each pair (sample, -mer) to a component, and the EM algorithm is as follows:
E-step:where
The M-step is similar to the one described for the refined model. Note that the running time of each iteration of the EM algorithm is now , and it is therefore very efficient as long as is fixed.
The model we presented infers common components in the samples but does not assign the reads to these components; it provides for each read a probability distribution over the components that could be used for an assignment, but in general is not optimized for this goal. Such assignment, or binning, is an important preliminary step in the analysis of metagenomic samples, especially binning according to species of origin. We therefore devised an unsupervised algorithm which performs binning over multiple samples simultaneously, again leveraging on sample similarity, this time assuming a common species core. Most previous unsupervised binning methods are based on sequence composition [10]–[15]. For example, CompostBin [11] computes for each read its 6-mer distribution, similarly to the process performed by our component inference algorithm, and clusters these distributions using spectral methods. The main limitation of composition-based approaches is that they require relatively long reads (1000 bp in the case of CompostBin) due to the variance in sequence properties along the genome. Recently, a coverage-based method, AbundanceBin, was developed [16] with the advantage of being able to bin even very short reads (as small as 75 bp). Since it relies on abundance differences for binning, AbundanceBin is only able to discern between species whose abundance levels are considerably different (they report that a ratio of 2∶1 is required). Our algorithm is also coverage-based, but because it operates on multiple samples it can use abundance difference in any of the samples to tell between such species.
Assume we are given metagenomic sequencing samples, consisting of a total of bacterial species. We wish to divide the reads in all samples into bins that correspond to the species from which they were sequenced. The binning algorithm, which we term MultiBin, proceeds as follows:
In the last stage, the distance between every two vectors , was computed as . -medoids clustering was performed using a local search procedure, in which we start from a random choice of centers and attempt to improve the solution by swapping at least one of the centers with another vertex, until no further improvement can be made.
The initial stage of distance computation takes , and each iteration of the clustering algorithm takes ; in our experiments convergence was reached within three iterations or less. The running time of the alignment stage, as well as the size of , depends on the composition of the mixtures. We note that clustering is performed only on the tag reads, whose number is approximately bounded by the sum of the genome sizes in the samples divided by the read length.
The above procedure assumes that the number of species in the samples in known. In the Results section we describe a procedure for determining the number of species based on examining the clustering results for different numbers of bins.
We evaluated our methods using both real data and simulated data. We used the MetaHIT dataset (downloaded from EBI, accession ERA000116), which includes over 0.5 terabases of sequence generated from the gut microbiomes of 124 European individuals using the Illumina Genome Analyzer technology. The average amount of sequence per individual is 4.5 gigabases, and the paired-end read length is 44 or 74, depending on the sample. We used the publicly available raw reads, which were obtained after filtering human and Illumina adapter contaminant reads and low quality reads. The sampled individuals vary on the following variables: country of origin (Denmark/Spain), age, BMI (Body Mass Index), gender, and status for infectious bowl diseases (Ulcerative Colitis/Crohn's disease/disease free). In the context of this paper all the variables will be referred to as phenotypes of the human host, although country and age are in fact determinants of the metagenomic content, instead of being affected by it.
In the initial stage we simply compared, for each -mer, its relative abundance in different phenotype groups using a two-sample t-test. Surprisingly, the relative frequencies of many of the -mers are significantly correlated with many of the phenotypes. This is true even for as small as 3: for example, the frequency of 69% of the -mers and of 61% of the -mers differs between the Spanish and the Danish samples at the 0.05 level. To the best of our knowledge, such dramatic differences in sequence composition between samples from different countries have not been observed previously. This effect might be partially due to differences in sample preparation and DNA extraction procedures, which are known to exist between the MetaHIT samples from the two countries; however, we also observed significant differences across phenotypes within each country: The frequency of 40% of the -mers differs between the Crohn and the healthy Spanish samples, and the frequency of 18% of the -mers differs between the 10 highest- and lowest-BMI Danish groups. Permuting the phenotype labels times yielded p-values of 0.0027, 0.0521 respectively for these fractions of rejected nulls.
A possible concern regarding the counts statistics are possible biases in the GC content distribution of the reads. We note that unlike different single-genome samples, different metagenomic samples are not expected to contain the same sequence composition characteristics, and therefore normalizing for such biases is a challenging task. We note that in the context of an association study between a phenotype and the metagenome, it is possible to avoid this problem using a permutation test, at the expense of power reduction.
The majority of the correlations between -mers and phenotypes are false positives resulting from a hidden stratification which confounds the -mer distributions. In order to reveal the components and to quantify them, we solved the probabilistic model described in the Methods section. Because the MetaHIT dataset is large, we used the more efficient version of the algorithm (which solves Equation 3). The input to the algorithm is a counts matrix of size [124×136], detailing for each of the samples how many occurrences of each possible -mer it includes (there are only possible -mers, instead of , because complementary strings are indistinguishable in the sequencing data). Since extracting multiple -mers from each read did not seem to considerably change the results in this particular case (this does not hold in general, as illustrated later), we used only the first, highest-quality -mer. The model was solved for three components ( = 3) by running EM multiple times from random starting points and choosing from the maximum-likelihood run. The solution provides, for each sample , the components proportions , and , such that .
We first tested each of the components for correlation with each of the phenotypes in the following regression model:(5)We solved this model for each phenotype and for each cluster , attempting to discover biologically meaningful components. Simple regression and logistic regression are used for continuous and dichotomous phenotypes, respectively. As can be seen in Table 1, highly significant p-values were obtained for predicting country among healthy individuals and for predicting BMI among the Danish. These results remain consistent also after correcting for the other measured phenotypes by entering them as covariates into the regression models. Results significant at the level were obtained also for predicting colitis and Crohn status among the Spanish. In the case of Crohn's disease, the power of the regression model was limited due to the small number of cases (only ), but permutations yielded a p-value of . Figure 1 visually demonstrates the separation of Crohn cases and controls on the plane defined by the components proportions.
We note that we also solved the model for . In these cases we found that the smallest per-component p-values, as well as the proportions of explained phenotypic variance captured by all components together, were similar to those obtained for . We therefore report the estimates for three components throughout the paper.
The results from the last section, showing that the components proportions are correlated with some of the phenotypes, suggest they indeed may be used to correct the association between these phenotypes and long -mers; we therefore attempted to perform this correction on -mers.
Figure 2 demonstrates such a successful correction. Two quantile-quantile curves compare the uniform distribution to the distribution of the p-values obtained by testing association between all possible -mers and BMI within the Danish samples. The black curve shows the uncorrected p-values, and its shape reflects the fact that they are highly deflated. The red curve shows the p-values obtained by adding the components proportions to the regression equation; this curve approaches the identity, indicating that these statistics capture the variance in the phenotype explained by most of the -mers.
We compared the precision of the two models, the original and the refined, utilizing different strategies for -mer extraction. Performance was tested on simulated data, each simulation consisting of mixtures of the following bacterial species: Listeria monocytogenes (phylum Firmicutes), Bacteroides vulgatus (phylum Bacteroidetes), Bifidobacterium longum (phylum Actinobacteria), Pseudomonas stutzeri (phylum Proteobacteria). The components distributions matrix was randomly drawn from the uniform distribution and normalized to row-stochastic, and the -mer distributions matrix was computed from the actual genomes. For each sample 1,000 sequencing reads of length 100 bp were simulated by sampling a bacterium according to P, and then a random position in the genome of that bacterium as the starting point of the read. We chose a read length of 100 bp since it is currently a length that is obtained by most high-throughput sequencing technologies.
Figure 3 compares the effect of extracting different subgroups of -mers along each read and using them either in the original or in the refined model. Under the original model, extracting multiple -mers improves estimation precision of compared with extracting only the first -mer; this is the case even when these -mers overlap, and are therefore highly correlated. Shifting to the refined model greatly improves the estimation precision when choosing all non-overlapping -mers along the read; however, even better results are obtained when using only nine sparsely dispersed -mers along the read, as using a small number of distant -mers decreases the dependencies of -mer sequence between and within reads.
PLSA is a dimensionality reduction method, and we therefore compared its performance to the widely used principal component analysis (PCA). PCA has been extensively used in metagenomic studies [2], [5] for sample visualization and classification. We used the simplified setting in which the counts matrix is generated from mixtures of multinomials: Setting we generated and by drawing their entries from the uniform distribution followed by normalizing to row-stochastic, and then generated the counts matrix by sampling each row according to the multinomial distribution specified in , with varying numbers of counts per sample.
Both PLSA and PCA were tested in the task of estimating the matrix . Since PCA operates with no stochasticity constraints, we estimates precision as the average squared correlation coefficient () between the true vectors and either the three strongest principal components (for PCA) or the vectors (for PLSA). For both methods we chose the ordering of the components that yielded the highest score.
Figure 4 shows that PLSA's estimates of are considerably more accurate than those obtained by PCA. This result confirms that PLSA is indeed a more appropriate method for characterizing mixture components in our context.
Nine datasets, each consisting of five metagenomic samples, were generated using MetaSim [27]. All samples in a given dataset contained the same set of bacterial species in different, randomly drawn proportions. The datasets differed in the number of species they contain, ranging from 2 species to 10 species in each sample of the most complex dataset. The species distribution of a sample containing species was generated incrementally by adding a random number sampled uniformly from [0,1] to the species proportions of an existing sample and normalizing to 1. In the first experiment we generated reads of length bp from each sample, and in the second experiment reads of length bp each. Overlaps between reads were determined by running BLAT [28] and requiring an exact match at the edges of the reads; the BLAT parameters we used restricted the results to matches of length bp and above. Precision was computed as the fraction of reads assigned to the correct species, averaged over all species.
We compared the precision obtained by MultiBin to the performance of AbundanceBin, a program implementing the equivalent coverage-based approach which was shown to perform precise binning of species exhibiting different abundance levels using reads of lengths 400 and 75 bp. AbundanceBin operates on single samples only, and therefore was run separately on each sample, while MultiBin was run on all five samples in each dataset simultaneously. We note that the separate execution of AbundanceBin conveys no information about the correspondence between the bins across samples, and so we chose the best matching between the bins and the species in each sample so as to maximize the total precision.
As can be seen in Figure 5, MultiBin performs better than AbundanceBin over both read lengths and over all dataset complexities. For the 400 bp reads MultiBin maintains a precision of over 0.8 even for mixtures of five species. MultiBin is also able to bin the 75 bp reads, although with lesser success; we note that the ability to bin short reads is unique to coverage-based approaches, and that in principle there is no advantage in having longer reads, assuming the coverage is high enough. As for AbundanceBin, its performance on the simple mixtures exhibits high variation between samples because of its reliance on large abundance differences within each sample. AbundanceBin's performance also deteriorates more rapidly as the number of species per sample increases compared with MultiBin.
We went on to evaluate MultiBin under the realistic scenario in which the reads have sequencing errors. To do so, we adjusted the alignment stage of MultiBin, currently performed by BLAT, to be more permissive. We tested this modification by again generating for the above datasets reads of length 400 bp, this time introducing base substitutions into the reads. When the substitution rate was increased to and then to , the precision for mixtures of two species decreased from 1.0000 to 0.9975 and then to 0.9966. Overall we conclude that the effect of these errors is not dramatic, and that for realistic error rates they could be largely moderated by adjusting the alignment procedure.
We note that integrating information across samples enables MultiBin to perform precise binning even when the variance in species distribution across the samples is relatively small. For example, when simulating 400 bp reads from five nearly-balanced mixtures of two species - the relative abundance of the more abundant species were 0.57, 0.55, 0.70, 0.53 and 0.56 - AbundanceBin still obtained a precision of 0.93. These results also demonstrate that MultiBin achieves precise binning on nearly-balanced samples; in contrast, a coverage-based method which did not integrate information across samples would produce extremely poor results on each of these samples alone.
Determining the number of clusters is an issue widely explored in the literature, and particularly, several approaches exist and have been tested for similar problems [29]. We found that running the algorithm multiple times for different values of , measuring the Hartigan index [29], [30] for each value and choosing the value at which the index decreases sharply and reaches a plateau gave accurate results, as long as the binning itself was accurate enough (precision of and above).
Our results so far demonstrate that joint modeling of multiple metagenomic samples can be helpful in the analysis stage. A further question has to do with the design stage: Given a fixed coverage depth and a potential pool of related metagenomic samples, how many of the samples should be sequenced in order to achieve optimal characterization of the underlying microbial composition? There seems to be a tradeoff between sequencing with high coverage a small number of samples and the superficial sequencing of many samples. We tested this tradeoff in both the components estimation problem and the binning task.
For components estimation, our task is to best characterize the components, or in other words to estimate the matrix; the matrix varies with the number of samples and therefore cannot be compared here. We simulated instances of the proposed probabilistic model by uniformly drawing the P and F matrices followed by normalization to row-stochastic, and then drawing observations from the corresponding multinomial distributions. We used components, each defined by a multinomial distribution over possible results. Initially, was defined for 1,000 samples, and for each sample 1,000 counts were drawn. The model was then solved for a decreasing number of samples by joining samples together to obtain 1000, 500, 200, 100, 50, 20, 10, 5, 2 and 1 samples. As can be seen in Figure 6(a), the optimal estimation of was achieved for 50 samples, each including 20,000 counts: Increasing the number of samples further past this point does not allow enough data to be gathered from each sample, resulting in a decrease in performance.
For the binning task, we ran our algorithm on mixtures of 15 species and reads of length 400 bp. Due to efficiency considerations we did not produce actual reads using MetaSim, but instead drew read start positions randomly along the genomes, and determined an overlap for reads which physically overlapped by more than 100 bp at the edges. Figure 6(b) shows the resulting precision when binning is performed over a fixed number of 8,192,000 reads allocated to an increasing number of samples, in which the species proportions are again uniformly drawn. The highest precision is obtained for 32 samples.
Both plots demonstrate that sample multiplicity is an advantage given a fixed coverage: as long as the per-sample coverage is reasonable, allocating the sequencing reads to as many samples as possible improves components characterization and binning precision.
We demonstrated the advantage in joint modeling of multiple metagenomic samples, by showing that it allows the unsupervised inference of hidden genetic component, and increases the precision of coverage-based binning. This advantage holds for both the analysis and the design stage; as for the latter, the results suggest that when wishing to characterize a given metagenomic sample, it is useful to divide the coverage between additional samples from similar environments. It might also be possible to apply a biological or chemical treatment to some of the samples, which would further accentuate the differences between them; when the samples are analyzed jointly, these differences are expected to further enhance performance. Similarly, sequencing data available from previous experiments can be used to improve the analysis of new samples. A similar tradeoff between the number of samples and per-sample coverage has been observed for testing the power of rare variant discovery in sequencing data [31], and sample multiplicity is likely to become a key issue in the future design of both standard and metagenomic sequencing studies.
We note that the components estimates obtained by solving the probabilistic model are interesting by themselves, outside of the context of association correction; particularly, they can allow for the characterization of variability patterns in metagenomic samples and for sample classification. Different estimates will be obtained by setting the parameters and to different values; our choice of and was meant to capture a high-level division, possibly taxonomic, of the microbial population, as it is known that bacterial phyla have characteristic sequence composition. A recent paper [5] identified metagenomic variability components using a supervised approach and divided the Danish samples accordingly to discrete classes termed enterotypes; interestingly, there are some correlations between these enterotypes and the components we obtain. For example, samples belonging to the first enterotype have a low proportion of the fourth component (p-value = ) when solving the model for and , and those belonging to the second enterotype have a low proportion of the first component () when solving the model for and . However, as explained in the Methods, our algorithm is fundamentally different from the PCA used to identify the enterotypes, and is expected to yield components of different nature, on top of it being unsupervised.
As for the proposed binning algorithm, unlike most other algorithms it can be used on datasets containing short reads, since the reads need only be long enough so as to determine unique sequence overlaps between them. In addition, the algorithm can be further improved to use not only coverage information but also other features, such as sequence composition, by adding them to the vectors on which clustering is performed.
Lastly, the implementation of the proposed association test for the MetaHIT dataset was limited by the sequencing quality, which forced us to extract only the first -mer from each read and therefore to examine only relatively short -mers (), otherwise the counts data would become too sparse. We believe that this problem could be addressed by the integration of sequencing uncertainties into the counts data. With the expected improvements in high-throughput sequencing technology in terms of read length and read accuracy, these issues may be of lesser importance in the future.
|
10.1371/journal.pbio.1001306 | Hormonal Signal Amplification Mediates Environmental Conditions during Development and Controls an Irreversible Commitment to Adulthood | Many animals can choose between different developmental fates to maximize fitness. Despite the complexity of environmental cues and life history, different developmental fates are executed in a robust fashion. The nematode Caenorhabditis elegans serves as a powerful model to examine this phenomenon because it can adopt one of two developmental fates (adulthood or diapause) depending on environmental conditions. The steroid hormone dafachronic acid (DA) directs development to adulthood by regulating the transcriptional activity of the nuclear hormone receptor DAF-12. The known role of DA suggests that it may be the molecular mediator of environmental condition effects on the developmental fate decision, although the mechanism is yet unknown. We used a combination of physiological and molecular biology techniques to demonstrate that commitment to reproductive adult development occurs when DA levels, produced in the neuroendocrine XXX cells, exceed a threshold. Furthermore, imaging and cell ablation experiments demonstrate that the XXX cells act as a source of DA, which, upon commitment to adult development, is amplified and propagated in the epidermis in a DAF-12 dependent manner. This positive feedback loop increases DA levels and drives adult programs in the gonad and epidermis, thus conferring the irreversibility of the decision. We show that the positive feedback loop canalizes development by ensuring that sufficient amounts of DA are dispersed throughout the body and serves as a robust fate-locking mechanism to enforce an organism-wide binary decision, despite noisy and complex environmental cues. These mechanisms are not only relevant to C. elegans but may be extended to other hormonal-based decision-making mechanisms in insects and mammals.
| During development, many animals choose between mutually exclusive fates, such as workers, soldiers, or queens in bees or ants. The choice between states is uniform throughout the animal since mixtures of these fates are not observed in the wild. The nematode Caenorhabditis elegans larvae integrate environmental conditions and have two choices: mature into reproductive adults or arrest development as dauer larvae—a latent form that can survive harsh conditions. The decision between both fates is governed by the hormone dafachronic acid (DA), however its regulation during development in response to environmental conditions has been unclear. In this study we show how two mechanisms are responsible for the integration of environmental conditions and the coordination of the decision between many tissues. We first show that a threshold mechanism integrates population density with the internal amount of DA made in the head. A normal population density has a low threshold of DA needed for worms to become adults, whereas a high population density increases this threshold and leads worms to develop into dauer larvae. We then show that the low levels of DA released from the head are amplified in the hypodermis (the main body syncytial epithelium) via a positive feedback loop, coordinating the decision over the animal. Disruption of this positive feedback yields abnormal adults. We propose that the positive feedback serves as a fate-locking mechanism enforcing an organismal binary decision—either adult or dauer—despite noisy and uncertain environmental conditions.
| During development, organisms often face unpredictable and unfavorable environmental conditions that may decrease their fitness. In some cases, organisms of the same genotype develop into alternate phenotypes, each better adapted to a particular environment. Alternative phenotypes entail changes in metabolism, developmental programs, behavior, or morphology [1]. To predict the capacity of the environment to provide for reproductive development, animals integrate external environmental conditions, the internal state of nutrient supplies, and other variables [2]. In many cases, the integration culminates in a decision between two mutually exclusive alternative phenotypes. Therefore, robust developmental mechanisms have evolved to ensure that the animals coordinate development exclusively into only a unified phenotype, as uncoordinated development will be detrimental [3].
Organism-wide binary decisions are common throughout the animal kingdom and include examples such as sex determination, changes in coloration as a function of season, and caste differentiation in insects [3]. However, our knowledge of the mechanisms that regulate the decision between alternative phenotypes and coordinate the outcome across a multi-cellular organism is fragmentary, drawing on principles derived from studies of different model organisms [4]. Insects coordinate the fate of an alternative phenotype by altering hormone amounts above or below a threshold, during a hormone sensitive period, prior to metamorphosis [2]. A threshold distinguishes the two alternatives, yet it remains unclear how the thresholding mechanism is regulated and how the amounts of hormone are maintained throughout the body post-decision.
Transcriptional amplification mechanisms such as positive feedback loops have been shown to lock in binary fate decisions in phage [5], bacteria [6], and yeast [7]. A hallmark of such positive feedback mechanisms is that signals from a noisy environment can be forced into a bi-stable response by a threshold. Signal levels above the threshold will be amplified and maintained at a high abundance, acting as a memory module for a decision, thus enforcing a cell-specific fate. We sought to understand if such principles can be extended to hormonal regulation in multi-cellular organisms, specifically as a means to threshold, coordinate, and maintain the alternative phenotype in the selective environment.
To understand the interaction between genetic and hormonal regulatory mechanisms that integrate environmental conditions and coordinate a discrete developmental fate, we looked at the dauer decision in the free-living nematode C. elegans. During this decision, C. elegans integrates environmental conditions and chooses between the mutually exclusive fates, dauer or reproductive development. In favorable environments—plentiful food, moderate temperatures, and low population density—C. elegans develops rapidly through four larval stages (L1–L4) separated by molts, into a sexually reproductive adult. In unfavorable environments—high population density (indicated by high levels of a constitutively secreted dauer pheromone) limiting food or high temperature—animals can decide to develop into an alternative third larval stage, the dauer diapause, a developmentally arrested, long-lived form geared towards survival [8]. Dauer larvae do not feed and can endure harsh conditions, including starvation, desiccation, heat, and oxidative stress [8]. Accordingly, dauer larvae have profound morphological changes including an assault-resistant cuticle, pharyngeal constriction, and sealing of body cavities. Whereas adult nematodes live for about 3 wk, dauer larvae can survive several months. When returned to favorable conditions, dauer larvae resume development, molting into L4 larvae and adults [9]–[11]. In either reproductive or dauer mode, it is essential that the execution of the decision will be robust and that no mosaic phenotypes arise, as this will compromise survival. Moreover, understanding the decision-making process of diapause entry in C. elegans can illuminate analogous processes in parasitic nematodes whose infective stages are like dauer larvae and are regulated by some of the same signaling pathways [12]. A deeper understanding of the decision to become an infective juvenile as well as exit from this stage will facilitate the design of therapeutics that inhibit parasitic infection.
Although the major signaling pathways regulating dauer formation have been identified, the cellular and molecular basis of this binary decision is not clear. Environmental cues are detected by multiple sensory neurons that integrate inputs into hormonal outputs by unknown means [13]–[16]. Molecular analysis has revealed at least four signaling pathways. Components of neurosensory structure and guanylyl cyclase signaling are involved in sensing temperature, nutrients, and dauer pheromone [17], which regulate secretion of insulin/insulin-like growth factor and TGFβ peptides. Insulin and TGFβ signaling converge on a steroid hormone pathway, which metabolizes dietary cholesterol into several bile acid-like steroids, called the dafachronic acids (DAs) [18]–[23]. DAs serve as hormonal ligands for the nuclear hormone receptor transcription factor DAF-12, which regulates the life cycle fate decision. Liganded DAF-12 promotes reproductive development, whereas unliganded DAF-12, together with the co-repressor DIN-1S, directs the dauer fate. Thus, DAF-12 serves as a DA-responsive switch that determines whether an animal will undergo reproductive or dauer development [20]–[27]. The regulation of DAs during development as a function of environmental conditions remains largely unknown.
Here we identify the times of integration and commitment to this life cycle fate choice and show that environmental conditions affect the threshold at which discrete levels of DA bypass the dauer fate. Higher amounts of DA are necessary to implement and coordinate the reproductive decision throughout the whole animal. We show that a positive feedback loop, which amplifies the amounts of DA in the hypodermis (hyp7; WBbt:0005734), is required under some circumstances to produce the higher amounts of DA, while a negative feedback loop keeps hormones with normal bounds. Finally, we demonstrate that the amplification of DA in the hypodermis is responsible for the irreversibility of the decision and the proper execution of reproductive programs. We propose that hypodermal amplification of a hormonal signal acts as a commitment mechanism that enforces the binary decision.
The decision between dauer arrest and reproductive growth is made at two points during early larval development. During late L1, worms develop into the L2 stage in favorable environments or into the dauer-capable pre-dauer (L2d) stage in unfavorable environments. During the mid-L2d stage, worms commit to either the dauer or resume reproductive development as L3 larvae (Figure 1A) [10],[28]. Golden and Riddle (1984) [28] showed that worms must be exposed to pheromone before the L1 molt in order to develop into the L2d stage and commit to the dauer fate before the mid-L2d stage. We re-visited these experiments and modified them to liquid culture to increase scale, homogeneity, and throughput. We measured the frequency of dauer formation in response to dauer pheromone while grown in the presence of sufficient food for adult development. Mean frequencies of life stages in favorable and unfavorable growth conditions were tightly distributed and highly reproducible, indicating the homogeneity of the liquid culture conditions (Figure S1A–B).
To identify when worms commit to reproductive development as L2 larvae, we performed a “shift-to-unfavorable” experiment by adding a high concentration of pheromone to synchronously hatched worms at progressive times. Worms stopped responding to pheromone at 18–20 hours post-hatch (hph; Figure 1B; 16.6%±21.4% dauer formation), which coincides with the beginning of the L2 stage. After this time, animals initiated reproductive development despite exposure to unfavorable conditions. To identify when worms commit to the dauer fate, we performed a “shift-to-favorable” experiment by growing synchronously hatched worms in unfavorable conditions (high concentration of pheromone) and washing away pheromone at progressive times. L2d worms committed to dauer during mid-L2d at 33 hph (Figure 1C; 29.1%±25.1% dauer formation), 18 h after the L1/L2d molt. Shifting worms to favorable conditions after this time did not affect their propensity to become dauers. We next identified when L2d worms commit to L3. We reasoned that L2d worms exposed to favorable conditions for longer times would have a higher propensity to develop into adults. The “shift-to-favorable” experiment was modified by growing worms under unfavorable conditions to obtain L2d animals, followed by a shift to favorable conditions at 24 hph. Worms were then returned to unfavorable conditions after varying amounts of time. We found that a 3 h pulse into favorable conditions was sufficient to commit L2d animals to reproductive development (Figure 1D; 0.02% dauer formation). Pulses at different start times during L2d had similar responses (Figures 1E, S1D). The L2d stage is thus divided into two periods: integration (from the beginning to the middle of the L2d stage, 16–33 hph) and commitment and implementation of the dauer program (33–48 hph).
We sought to find cellular and molecular candidates that could account for the dauer and L3 commitments. The two species of the DAs, Δ4-DA and Δ7-DA, are good candidates because synthetic DAs can fully rescue the Daf-c phenotypes of the null allele daf-9(dh6) (WBGene00000905) as well as daf-7/TGFβ (WBGene00000903) and daf-2/InsR (WBGene00000898) mutants [21]. Partial reduction of daf-9 function results in animals that bypass the dauer stage yet exhibit abnormal gonadal morphogenesis and migration (Mig; WBPhenotype:0000594) and occasionally aberrant cuticle shedding (Cut; WBPhenotype:0000077) defects (Figure 2A) [18],[19]. Exogenous DA can also rescue these phenotypes [20],[21],[23],[24]. We thus hypothesized that a low amount of DA is required to bypass dauer and commit to L3, whereas a high amount is required for normal development.
To understand the physiological response to DA dose, dauer-constitutive daf-9 loss-of-function mutants were treated with increasing amounts of Δ7-DA and measured for dauer and reproductive adult fates. Most daf-9(dh6) null animals developed into abnormal adults when supplemented with a minimum of 10 nM DA (Figure 2B, 74%±42% non-dauers), suggesting that a threshold of DA has to be crossed before committing to adult fate (dauer bypass DA threshold). Increasing the levels to 25 nM DA decreased the frequency of dauers to 99%±1% with about 66% of animals developing normally (Figure 2B). Further increase of DA to 50 nM increased the frequency of normal adults (Figure 2B). For a distribution of Mig and Cut phenotypes, see . Similar results were observed with animals homozygous for daf-9(e1406) or daf-9(m540) (Figure S2; worms were not synchronously hatched), both of which are strong loss-of-function alleles. By contrast, the weak loss-of-function allele daf-9(rh50) does not result in Daf-c phenotypes, but in highly penetrant Mig defects (95%±3%) [18]. In these animals, only 10 nM of DA was required to rescue over 90% of the Mig phenotypes (Figure 2C), revealing a 5-fold decrease in the amount of exogenous DA required to promote normal development compared to the stronger daf-9 mutants (dh6, e1406, and m540; Figure 2S). Thus, daf-9(rh50) animals produce sufficient amounts of DA to bypass dauer development but require additional DA to develop into normal adults, consistent with our finding that different levels of DA are required for the two processes.
Many Daf-c mutants have a slower developmental rate, but the basis of this is not well understood (A. A., unpublished observations). To test the effects of DA on developmental rate, daf-9(dh6) worms were synchronously hatched in different concentrations of DA and scored for developmental stage at 48 hph (the time at which wild-type worms grown in favorable conditions are young adults (YAs) and worms grown in unfavorable conditions are dauers; Figure 1A) and for egg production the following day. At 25–50 nM DA, worms developed into L4, whereas worms supplemented with 75–175 nM DA already developed into YAs. Worms that were in the L4 or YA stages at 48 hph were gravid the next day. Exogenous addition of DA had no effect on wild-type growth rates (Figure 2D). These trends indicate that increases of DA levels can accelerate growth rate until it matches that of wild-type worms (Figure 2D).
DA and dauer pheromone have opposite effects on dauer formation, with DA preventing and dauer pheromone promoting the dauer stage. We investigated the dose-response relationship when administered together, with respect to bypass of the dauer diapause and normal reproductive development. We hypothesized that the same dose of DA would be required to overcome dauer induced by pheromone as that seen in daf-9(dh6) null mutants.
Synchronized populations of daf-9(dh6) worms were supplemented with a combination of DA and pheromone at different concentrations and scored for dauer, abnormal, and normal adult development at 48 hph (Figure 3A–C). Unexpectedly, addition of pheromone at 1%, 3%, or 6% (which induce 47%±4%, 92%±2%, and 95%±2% dauer in wild-type worms, respectively; Figure S1D) increased the concentration of DA necessary to exceed the dauer bypass DA threshold to 30, 45, and 58 nM, respectively (Figure 3F). Moreover, 90% of the population developed into normal adults if worms were supplemented with 30 nM of DA more than the amount required to bypass the dauer bypass DA threshold (Figure 3D–F), similar to the concentration of DA needed to bypass the dauer bypass DA threshold in daf-9 worms without pheromone. These experiments demonstrate that dauer pheromone increases the amount of DA required to bypass the dauer bypass DA threshold and normal reproductive development.
To understand the time of action of DAs and their role in life cycle fate decisions, we sought to identify three key points in the response to DA: (i) the time at which daf-9(dh6) animals start responding to DA to bypass dauer, (ii) the end of response to DA for the dauer decision, and (iii) the requirements of exposure to DA for normal development to maturity. Synchronously hatched daf-9(dh6) worms were shifted from media containing DA dissolved in EtOH to media containing EtOH alone (downshift) or vice versa (upshift). Analysis of downshift experiments revealed that worms started responding to DA after 15 hph, the same time that wild-type worms commit to L2 mediated reproductive development (Figure 4A). When DA was washed away before 15 hph, worms developed into dauers, despite previous exposures of DA indicating that previous exposures to DA had no effect on commitment to reproductive development. Removal of DA at time points after 15 hph prevented dauer formation to increasing extents, which could be divided into two phases: a minimum of 3 h on 100 nM DA during the responsive period was sufficient to prevent 61.7%±19.6% of the population from becoming dauers, but these animals developed as abnormal adults (Figure 4A, 15 to 18 hph), whereas an additional 12 h were necessary to drive 100% of the population to normal adult development (Figure 4A, 18 to 30 hph). To determine when daf-9(dh6) worms became refractory to DA, upshift experiments were performed during the L2d stage. Worms responded to DA until 33 hph, precisely at the same time that wild-type worms became refractory to pheromone (Figure 4B; correlation coefficient = 0.996). Next, we asked whether the total time exposed to DA or the specific time (stage) of exposure to DA were regulating the fate decision and development of normal adults. Pulse experiments revealed that worms committed to bypass dauer when exposed to DA at 15 hph or 24 hph for as little as 3 h (Figure 4C,D). When DA was supplemented for 3 h at 24 hph worms developed into normal adults (Figure 4D). These data correspond with addition of DA at 15 hph for 12 h, indicating that development into normal adults is a function of stage (27 hph, mid-L3) and a persistent exposure to DA (Figure 4D). Similar results were seen with the daf-9(e1406) allele (Figure S4). In sum, DA can affect the decision during a specific temporal window (15 to 33 hph) during the L2d stage, the same time that wild-type L2d worms integrate pheromone. Worms become committed to bypass dauer with a minimal exposure of 3 h in DA, but additional persistent exposure to DA over 12 h is necessary for normal adult development.
We wanted to understand how the spatiotemporal and tissue-specific regulation of daf-9 is related to hormonal activity and stage commitments. The two bilaterally symmetric XXX cells (WBbt:0007855) express daf-9 throughout all stages, suggesting that they may produce steady levels of DA [18],[19]. Hypodermal daf-9 expression is more complex: hypodermal daf-9 is weakly expressed in L3 larvae growing in favorable, low-stress conditions, strongly expressed in L3 larvae growing in mild stress conditions, and not expressed under high stress conditions that trigger dauer formation [20],[23].
First, we investigated the expression of daf-9 mRNA during L2 and L2d in favorable and unfavorable conditions by whole animal qPCR in wild-type worms. Second, we examined the expression of DAF-9 protein levels and distribution with a translational DAF-9::GFP fusion by fluorescent microscopy (strain AA277; lin-15(n765), dhIs64[daf-9::GFP, lin-15(+)]; Gerisch et al., 2001 [18]; strain AA277 grows slower than N2 and therefore commitment to dauer occurs at 36 hph; Figure S5). We found that daf-9 is regulated differently in favorable and unfavorable environmental conditions.
In favorable conditions that promote reproductive development, total daf-9 transcripts were upregulated 7±1.1-fold at 16 hph and peaked at 30 hph, with 10-fold upregulation (Figure 5A). All observed daf-9 upregulation was due to the daf-9a isoform as we were unable to detect the daf-9b isoform (see Materials and Methods). Eighteen percent of worms started expressing hypodermal DAF-9::GFP at 21 hph, mid-L2 stage, reaching a maximum of 75%±12% at 30 hph, mid L3 (Figure 5B; p<0.0001). Presumably the delay between daf-9 upregulation detected by qPCR to that observed by GFP is due to the translation of mRNA to protein and slower developmental rate of the AA277 strain.
Previous genetic experiments demonstrate that hypodermal daf-9 expression is regulated by DAF-12 and hormone biosynthetic genes [20],[23]. To monitor daf-9 transcriptional regulation in specific tissues as a function of DA, we performed experiments using a pdaf-9::gfp transcriptional promoter construct in the daf-9(dh6) background. This promoter construct largely recapitulates the behavior of the translational fusion, suggesting that the majority of regulation occurs at the level of transcription. At low DA concentration (0–0.5 nM), expression was seen only in the XXX cells, and all animals developed as dauers (Figure 5C–D). As DA concentration was increased to 0.75–7.5 nM, expression in the hypodermis dramatically increased by mid-L2, suggesting positive amplification (Figure 5C–D). Notably, within the range of 1–5 nM DA, animals bypassed dauer but exhibited the abnormal development phenotypes (Figure 5C,D). Hypodermal expression was decreased at 10 nM or shut off (50–100 nM), suggesting suppression of daf-9 expression. At these higher concentrations (>10 nM) all animals developed into normal adults. Exogenous DA and hypodermal daf-9 upregulation have an inverse relationship; intermediate and high levels of DA promote intermediate, and low levels of hypodermal daf-9, which correspond to states of abnormal and normal development, respectively.
In unfavorable conditions, total daf-9 transcripts were not significantly upregulated in L2d animals committed to dauer (Figure 5E,J; p = 0.14). All observed daf-9 upregulation was due to the daf-9a isoform as we were unable to detect the daf-9b isoform (see Materials and Methods). Nearly all worms grown in unfavorable conditions failed to show hypodermal DAF-9::GFP expression during L2d or dauer (Figure 5F; 92%–100%, p = 0.18). Shift-to-favorable experiments revealed a 40-fold upregulation of daf-9 transcripts in wild-type worms committed to reproductive development (Figure 5H). When L2d worms were pulsed into favorable conditions at 24 hph for a 6-h window, 76%±12% showed hypodermal DAF-9::GFP expression with onset as early as 27 hph (Figure 5G). Hypodermal daf-9 expression was retained even when worms were shifted back to unfavorable conditions. Conversely, 93%–99% of worms shifted to favorable conditions for 1 h did not express hypodermal DAF-9 (Figure 5I). These results correlate temporally with the minimum time that wild-type worms require a pulse in favorable conditions to bypass dauer and suggest that hypodermal daf-9 expression could be a cause or consequence of a decision to develop into L3.
Worms committed to reproductive development had transcriptional upregulation of daf-9 in the hypodermis, likely resulting in the production of the high levels of DA. We asked whether the XXX cells play a role in hypodermal daf-9 upregulation. Notably, we observed that after a shift from unfavorable to favorable conditions, hypodermal DAF-9::GFP expression was observed in a spatiotemporal manner along the anterior posterior axis (Figure 6), first and most strongly in the head region before expression spread to more posterior regions. This led us to hypothesize that under these conditions, XXX cells (located at the anterior) might act as a source of DA, releasing a small amount that is amplified and propagated in the hypodermis from anterior to posterior.
To test this hypothesis, we removed the XXX cells with a laser microbeam. We ablated XXX cells in worms expressing a translational DAF-9::GFP fusion developing in high pheromone concentration at 24 hph (mid L2d, pre-commitment), and worms were allowed to recover in favorable conditions. Nearly all (30/31) XXX-ablated worms lacked hypodermal DAF-9::GFP expression and developed as dauers, while 29/31 control mock-ablated L2d animals developed into adults (Figure 6B; p<1×10−10, Figure S6A). Therefore, intact XXX cells are necessary for L2d larvae to respond to favorable conditions, committing to reproductive development and initiating hypodermal daf-9 expression.
We next tested whether DA could rescue the dauer arrest caused by ablation of XXX cells. Worms were grown in high pheromone concentration and XXX cells were ablated at 24 hph (L2d before commitment), and shifted to growth in favorable conditions supplemented with 0, 1, 5, or 10 nM DA. An increasing frequency of both hypodermal DAF-9::GFP and normal adult development was observed as higher concentrations of DA were supplemented. Rescue with 1 nM DA yielded 22% adults and 78% dauers (N = 18), rescue with 5 nM DA yielded 56% adults and 44% dauers (N = 30), and rescue with 10 nM DA yielded 92% adults and 8% dauers (N = 39; Figures 6C and S6B). All XXX-ablated worms supplemented with exogenous DA developed either as normal adults or as dauers, with none of the Mig or Cut phenotypes seen in daf-9(dh6) worms at these concentrations of exogenous DA (Figure 2B). These results suggest that in the absence of the XXX cells, hypodermal daf-9 upregulation can be induced with as little as 1 nM DA, resulting in normal adult development. By contrast, in the daf-9 null background hypodermal daf-9 amplification is not possible, leading to abnormal development at low DA levels.
To test whether XXX cells act as a source of DA later in development, we ablated XXX cells after commitment to L3. Ablation at this time had no effect and resulted in worms that expressed hypodermal DAF-9::GFP and matured to adulthood (Figure 6D; p = 2×10−9, Figure S6C). Therefore, XXX cells act as a source of DA during the dauer decision and become dispensable later.
Here we characterized a molecular mechanism connecting environmental signals to hormonal regulation during the commitment to reproductive development in C. elegans. We identified specific time windows during which worms integrate their environmental conditions and make a decision between reproductive development and diapause. We demonstrate how environmental conditions regulate the threshold of hormone required to commit to a certain fate. Finally, we find that environmental information is funneled to the neuroendocrine XXX cells, which upon a decision to commit to adulthood initiate activation of a positive feedback loop of hormonal production. This mechanism, consisting of a proposed gene regulatory network, regulates robust hormonal amplification, thus enforcing the fate decision throughout the organism.
We analyzed the commitment of developing C. elegans larvae to reproduction (L3 larvae) or delayed reproduction (dauer larvae) in liquid culture, enabling a large and highly synchronized brood, amenable to facile and reproducible changes in environmental conditions. We showed that larvae exposed to favorable conditions for the first 12–18 h no longer entered dauer arrest when subjected to later pheromone exposure (Figure 7A). Conversely, if grown in unfavorable conditions between 12 and 18 hph, worms were induced to the pre-dauer stage, L2d. Worms integrate environmental conditions during the L2d stage and become irreversibly committed to the dauer fate by 33 hph. However, L2d worms can commit to the reproductive fate if pulsed with favorable conditions for 3 h before the 33 hph commitment point is crossed (Figure 7A). Our identification of the times at which these life cycle fate decisions occur allowed us to couple changes in the environment to the known molecular and cellular components involved in this decision.
Timing of the decision period is congruent with the requirement for the hormone DA and the expression of the DA biosynthetic gene daf-9. In this view, favorable conditions equate with the presence of DA, while unfavorable conditions equate with its absence. Starting from 15 hph, pulses of DA 3 h or longer will bypass dauer with no memory to previous exposures to DA. Similarly, daf-9 mutants stop responding to DA at 33 hph, mid-L2d stage. Thus, these periods of DA sensitivity overlap with the response to changes of population density in the environment.
Our studies suggest that two thresholds of DA must be crossed to ensure proper reproductive development: DA levels above the dauer DA bypass threshold will specify reproductive development, and DA levels below the normal adult threshold will specify dauers. If worms produce DA levels between the two thresholds, they will develop into abnormal adults. Importantly, the two thresholds only become apparent in daf-9 mutants, which uncouple the dauer bypass threshold from the normal adult threshold. Addition of DA to daf-9 mutants indicates that 10 nM DA is sufficient to overcome the dauer DA bypass threshold in liquid culture and 1–5 nM are sufficient on plates, therefore committing worms to L3 development. These animals require an additional 30 nM DA to promote normal gonadogenesis and cuticle formation, thus developing into normal adults. Higher levels of DA increase developmental rate. In addition, dauer pheromone can raise both DA thresholds, thus increasing the fraction of dauers in a population. Therefore, worms that produce DA levels above the dauer DA bypass threshold, or lower the threshold itself, will develop into adults. The difference of 30 nM between both thresholds is constant regardless of the amount of pheromone to which worms are exposed. These observations suggest that pheromone has additional targets downstream or parallel to DA production that antagonize reproductive development (Figure 7B).
What might be the molecular and cellular correlates of these two thresholds? The cytochrome P450 DAF-9 is limiting for DA production. daf-9 is expressed in the XXX cells from hatch and throughout development, and in the hypodermis starting from mid-L2 until L4. The timing requirements for DA described above suggest that the commitment to adult development through the L2 stage is made early in L2 between 15 and 18 hph, a time that precedes visible hypodermal daf-9 expression.
At high population density, the XXX cell appears to be source of DA required for the dauer decision, and the hypodermis amplifies DA production leading to normal development. When worms are shifted from unfavorable to favorable conditions, the dauer bypass DA threshold is lowered for a sufficient amount of time and the XXX cells presumably make a sufficient amount of DA to pass that threshold. Once the XXX cells release a small amount of DA, the hypodermis amplifies this signal leading to normal adult development. This amplification is visible as anterior to posterior propagation of hypodermal daf-9 expression originating in proximity of the XXX cells. If the XXX cells are ablated, there is no source of DA to trigger hypodermal daf-9 transcription and animals develop as dauer larvae. Hypodermal daf-9 amplification is triggered if XXX-ablated animals are supplemented with as little as 1 nM DA. Lastly, the onset of hypodermal daf-9 upregulation renders worms insensitive to removal of XXX, thus conferring the irreversibility of the decision and committing worms to the reproductive fate (Figure 7A).
In favorable conditions, the XXX cells and the hypodermis may share responsibilities. Under these conditions daf-9 expression in the XXX cells appears steady and hypodermal expression low. The XXX cells are sufficient but not necessary for committing to reproductive fate: rescue of the daf-9(dh6) putative null by a XXX cell-specific DAF-9 construct leads to adult development. Ablation of the XXX cells during the L1 stage, in worms grown in favorable conditions, results in 30% of animals developing as dauers [29]. However, the hypodermis can overcome this deficiency of XXX signaling by daf-9 upregulation [18],. Hypodermal expression of daf-9 works as a homeostatic regulator since low amounts of DA increase transcription of daf-9 in a daf-12-dependent manner, whereas sufficient production of DA by the XXX cells is not followed by hypodermal upregulation of daf-9 during the L2 and L3 stages.
From the daf-9 expression pattern, we infer that in favorable conditions DA is released in low levels over a long period of time, whereas worms developing in unfavorable conditions to adulthood release a burst of DA over a short period of time. This also implies that worms have a mechanism of counting and integrating hormone levels to reach the threshold of the dauer decision (Figure 7B). We speculate that this could be achieved by various levels of DA swapping DAF-12/DIN-1 or other co-repressor complexes for DAF-12/co-activator complexes, a known mechanism in nuclear receptor signal transduction [30].
Consistent with the importance of the XXX cells to the dauer decision, many components of the dauer regulatory pathways are expressed in these cells, including ncr-1, the Niemann-Pick C1 homolog, hsd-1 encoding a 3β-hydroxysteroid dehydrogenase, and sdf-9 and eak-6, which encode tyrosine phosphatases, eak-3, eak-4, and eak-7, novel proteins localized to the plasma membrane [18],[19],[27],[29],[31],[32]. These components as well as others could regulate enzymatic activities, availability, or hormone transport to and from the XXX. Additional activities in the dauer pathways could regulate the amount of DA produced in the XXX cells and the adult DA threshold, in endocrine or target tissues.
The spatiotemporal and homeostatic regulation of daf-9 provides a molecular mechanism that can explain the phases in the decision between dauer and reproductive development: integration of environmental and internal information, signaling of the decision, and implementation of a coordinated and irreversible decision. We propose that integration occurs by a process of information reduction and that signaling and implementation occur by a process of amplification and propagation upon reception of a discrete signal (Figure 7B). Complex information from the environment is measured by at least six neuron pairs (ASI, ADF, ASG, ASJ, ASE, and ASK [13],[15],[16]) and reduced in complexity by the Insulin/IGF, TGFβ [33], guanylyl cyclase [17], and steroid hormone pathways [18]–[21],[23],[24]. We propose that the XXX cells then integrate information from these signal transduction pathways and commit to reproductive development by releasing DA (Figure 7B). The latter phases, signaling and implementation of the decision, involve amplification and distribution of the discrete signal via a positive feedback loop in the hypodermis, thus ensuring a coordinated response over the whole animal. Amplification of the signal leads to independence from the integration apparatus. Should environmental conditions change and XXX cells cease to release DA, the amplification and diffusion over the whole body guarantees that all tissues will receive the necessary amounts of DAs required for normal adult development, thereby preventing inappropriate development in any cell lineage. In the future, it will be important to dissect further components of these life cycle commitments as well as the upstream mechanisms that weigh the decision within the XXX cells.
Although there are clear differences across taxa, hormonal regulation in C. elegans bears many similarities to insect and mammalian hormonal regulatory mechanisms. First, external cues such as nutrients and photoperiod as well as internal cues such as body size or organ development affect developmental progression. Second, the hormone sensitive period overlaps with the environmental sensitive period and acts as a cue integrator also observed in the insects lepidoptera, hymenoptera, and diptera [34]. Third, developmental progression and coordination of development are relayed via the insulin/IIS and TGFβ/activin signal transduction pathways. In particular, the insulin/IIS pathway positively regulates reproductive growth in C. elegans [35],[36], insects [37], and mammals [38],[39], and may do so by converging on steroidogenic pathways [20],[23],[37],[40]. Conversely, a reduction of Insulin/IGF signaling increases the propensity for diapause in nematodes [4], and insects [37], as well as torpor in mammals [41],[42]. TGFβ/activin signaling also controls steroidogenic enzymes and influences mammalian reproductive development [43]. Fourth, insect TGFβ/activin signaling regulates metamorphosis by controlling expression of a subset of steroidogenic enzymes through Insulin/IGF signaling and the prothoracicotropic hormone PTTH. PTTH is an insect peptide hormone that regulates developmental timing and body size at metamorphosis [44]. Its release from prothoracicotropic neurons and binding to its cognate receptor Torso in the prothoracic gland results in stimulation of steroidogenic gene expression and the production of Ecdysone [45]. In mammals, pulses of gonadotropin-releasing hormone (GnRH) may work analogously to PTTH to signal commitment to reproduction [2],[38]. Although C. elegans lacks PTTH or GnRH hormones, conceivably other neuropeptides could take on this role. Fifth, the discrete spatial regulation of primary and secondary sources of steroid metabolism resembles somewhat those in insects, in which the precursor Ecdysone is produced in the PG, but the final product 20-hydroxyecdysone is converted in the peripheral tissues including the epidermis, midgut copper cells, Malpighian tubes and the fat body [44]. Last, hymenopterans and coleopterans have been shown to regulate alternative phenotypes by modulating the hormonal threshold via secreted pheromones [2]. The ant Pheidole bicarinata regulates the threshold of Juvenile Hormone (JH), causing the differentiation between worker and soldier ants. Worker ants are determined by a sub-threshold dose of JH, while soldier ants are determined by above threshold amounts of JH. Ants that have committed to the soldier caste will secrete a soldier inhibiting pheromone, which raises the JH threshold in pre-committed ants [46].
We have demonstrated that a simple network architecture of positive feedback can both lock in a fate decision and convey irreversibility of a decision. Because the hormonal regulatory mechanisms found in the worm are similar to insect and mammalian systems, the relative simplicity of the C. elegans may prove beneficial in elucidating the environmental, cellular, and molecular mechanisms of decision-making involved in reproductive commitments in multi-cellular organisms. It will be particularly interesting to determine if the commitment role of hormonal amplification and feedback plays an analogous role in other animals as observed in C. elegans.
All worms were handled using standard growth and cultivation techniques using the bacterial strains HB101 and OP50 as food sources [47]. Unless otherwise stated all liquid cultures were grown in glass flasks at ∼1 worm per µl at 20°C in S complete medium supplemented with 7.5 mg/ml HB101 as described in [47] in an Innova 4230 incubator at 180 RPM. The wild-type strain used was N2 (Bristol).
Worms were hatched synchronously essentially as described by [48]; changes are described in the SOM.
Crude pheromone was prepared as described in [28]. Each pheromone extract was tested on N2 worms (1 worm per µl) and diluted so that 3% (v/v) would yield 90%±2% dauer arrest in a culture supplemented with 7.5 mg/ml of HB101.
Synchronous broods were grown as described above to the L2d stage by supplementing media with 3% (v/v) pheromone, partitioned into multiple parallel cultures, and grown in glass tubes. Shift to favorable: at specified times, broods were washed 3 times in S basal to remove pheromone. Cultures were re-suspended in S complete medium containing HB101 and calibrated for density. Shift to unfavorable: broods were supplemented with 3% (v/v) pheromone and grown in glass tubes. At specified time points (L2d), worms were partitioned into a control sample and experimental samples, which were washed 3 times with S basal. Worms were suspended in S complete medium and allowed to grow for specific time periods until 3% pheromone (v/v) was added.
Liquid culture: Δ7-DA was solubilized in 100% EtOH to necessary concentrations. Liquid culture assays were performed by adding EtOH-solubilized Δ7-DA in S basal medium. NG agar plate assays were performed by resuspending EtOH-solubilized Δ7-DA in S basal with OP50 and spreading on plates. Worms were picked onto Petri plates not more than 1 d after Δ7-DA was added to those plates. For the pdaf-9::gfp experiment Δ7-DA was added on 3 cm NG agar plates, seeded with OP50.
daf-9(dh6), daf-9(e1406), and daf-9(m540) worms were grown in liquid culture with different concentrations of Δ7-DA as described above. Worms were washed once with S basal medium to remove HB101 and mixed with S basal medium containing 1 mM sodium azide (to limit worm movement), spotted onto a 24-well plate. Worms were scored for gonad migration and cuticle shedding. Phenotype frequencies were calculated as the means of three biological replicates ± standard deviation.
Plots of daf-9(dh6) supplemented with 0%, 1%, 3%, and 6% pheromone as a function of DA were fit to a sigmoidal curve of the form f(x) = 1/(1+xn), where x is a log transformed concentration of DA, and n is the hill coefficient of the slope. Each pheromone concentration hill coefficient and EC50 were used to solve the EC90; the concentration of DA to bypass 90% dauer formation or 90% normal adult formation (Figure 3F, yellow and blue curves) according to the equation: EC90 = (90/(100-90))1/n * EC 50.
Frequencies were calculated within each biological replicate and means of frequencies ± standard deviation were calculated between biological replicates. We determined the point of commitment at the measurement times with highest standard deviation as it represents the tipping point of a transition between non-committed to committed worms. We calculated a q-statistic based on a Tukey type multiple comparison test for differences among variances (Zar 2009 [49]; Table S1).
Stage distributions were compared between three biological replicates in favorable and unfavorable conditions. A Bartlett's test [49] was used to determine if variances were significantly different between all stages of development.
Analysis was performed by a one-way ANOVA (Figure 5B,F). Significance of hypodermal upregulation in favorable versus unfavorable conditions after different time windows in favorable conditions was analyzed using a two-tailed t test between worms scored 30 hph (Figure 5G,I). Significance of transcriptional upregulation was analyzed by one-way ANOVA across all time points (Figure 5A) and paired t tests between L2d uncommitted to L2d committed to dauer and to L2d committed to L3 (Figure 5H,J).
AA277 worms (lin-15(n765), dhIs64[daf-9::GFP, lin-15(+)]) were grown to L2d stage in pheromone as described above. We found it necessary to use fluorescently labeled XXX cells as they migrate from the nose tip to the posterior region of the anterior bulb [50]. Worms were placed on glass slides with 5% agarose and 1 mM sodium azide in S basal, and laser microbeam ablations of the XXX cells were performed as described [51]. Worms were allowed to recover for 2 h before re-mounting on slides and verifying successful ablation by determination that no fluorescence signal was seen from either XXX cells. Worms were then transferred to either NG agar plates or NG agar plates supplemented with 1, 5, or 10 nM Δ7-DA. All ablations were coupled with mock-ablation controls. Statistical significance of observed differences between ablations and controls was determined using Fisher's exact test [49].
Strain AA277 was grown in liquid culture as described above. At specific times, worms were washed once in S basal medium and plated on glass slides with 5% agarose and 1 mM sodium azide in S basal. Worms were scored for hypodermal DAF-9::GFP under 40× magnification using a Zeiss Axiovert 200 microscope with a 200W mercury bulb.
Anterior posterior DAF-9::GFP expression: Each worm was imaged using both Nomarski and fluorescence using a CoolSnap HQ camera (Photometrics, Tucson, Arizona, USA) run through Metamorph software (MDS Analytical Technologies, Toronto, Ontario, Canada). Four to six worms were imaged per time period at 5 ms per Nomarski image and 400 ms per fluorescence image. Worms were straightened computationally, normalized to length, and mean grey value was quantified using custom software written in Matlab (see Text S1 for details).
Different concentrations of DA were added to NG agar plates, seeded with OP50. One day later, 10 reproductive daf-9(dh6) pdaf-9::gfp adults (grown in the presence of 250 nM DA) were placed on each plate for egg laying. F1 progeny were scored for hypodermal daf-9 expression levels and dauer, molting, and gonadal cell migration phenotypes. Experiments were performed at 20°C and repeated at least twice. The GFP fluorescence was imaged through a Zeiss Axio Imager Z1 and photographed with an AxioCam MRm camera. Pixel intensity over a fixed area was measured with AxioVision 4.7 software.
Synchronous populations of worms were grown at 20°C either in favorable (2 worms per µl, 15 mg/ml HB101) or in unfavorable conditions (3% pheromone v/v, 2 worms per µl, 15 mg/ml HB101). At each time point, 104 worms were washed 3 times in S basal medium without cholesterol (pH = 6) to decrease bacterial load and to wash off excess pheromone. Samples were concentrated in 100 µl volume and suspended with 1 ml TRIzol reagent (Invitrogen, USA) and mixed with 0.6 µl/ml Linear Poly acrylamide, used as a carrier [52], flash frozen using liquid nitrogen, and stored at −80°C until processed.
RNA purification using TRIzol was adapted from the manufacturer's protocol and is described in the SOM. RNA was subjected to quality control by Nanodrop spectrophotometry (A260/280 ratio) and Agilent Bioanalyser (S28 to S18 ratio). Samples were processed if the A260/280 ratio was above 1.9 and the S28 to S18 ratio was above 1.8. RNA was digested with RNase-free DNase (Ambion, Austin, Texas) according to the manufacturer's instructions. Total RNA was made into cDNA by reverse transcription reaction using Superscript III (Invitrogen, San Diego, California). mRNA was selected for reverse transcription by using oligo dT primers. Reactions containing no reverse transcriptase were carried out in parallel. cDNA was purified on silica columns (Qiagen, Venlo, Netherlands) and diluted to 16 ng/µl for subsequent qPCR analysis.
daf-9 transcripts were analyzed with three pairs of primers spanning different exons according to the WS190 gene model (http://ws190.wormbase.org). Each of the three amplicons was between 115 and 183 bp in length and included sequence from two exons (Table S2). All qPCR reactions were prepared using Roche SYBR Green I Master (Roche Diagnostics) and carried out in a Roche Lightcycler LC480. Data analysis was performed according to the ΔΔCt method [53]. Efficiency values of each primer set were empirically determined by performing a dilution series on pooled cDNA. Transcripts were analyzed if they crossed the Ct threshold before 34 cycles. Control genes were determined empirically by measuring gene expression that did not change significantly (Pearson correlation >0.995) during larval development (L1 through L4) and dauer fate. daf-9 relative abundance was determined as follows: for mRNA processed from worms grown in favorable conditions, daf-9 was normalized to the geometric mean of control genes pmp-3 and- Y45F10D.4 [54]. mRNA processed from worms grown in unfavorable conditions was normalized to relative abundance levels of ver-2, a gene expressed only in the ADL neurons [55]. daf-9 fold change was determined by normalizing all time points to relative abundance in the L1 stage. Error bars represent mean fold change ± standard deviation across two technical replicates originating from three biological replicates (six data points).
Accession numbers from http://www.wormbase.org: Genes: daf-2: WBGene00000898, daf-7: WBGene00000903, daf-9: WBGene00000905, daf-12: WBGene00000908, daf-16: WBGene00000912, din-1: WBGene00008549, ncr-1: WBGene00003561, hsd-1: WBGene00012394, sdf-9: WBGene00004748, eak-3: WBGene000022356, eak-4: WBGene00009955, eak-6: WBGene00008663, eak-7: WBGene00010671.
Phenotypes: Mig, WBPhenotype:0000594, Cut, WBPhenotype:0000077.
Cells: XXX: WBbt:0007855, hyp7: WBbt:0005734.
|
10.1371/journal.pntd.0002888 | Strengths and Weaknesses of Global Positioning System (GPS) Data-Loggers and Semi-structured Interviews for Capturing Fine-scale Human Mobility: Findings from Iquitos, Peru | Quantifying human mobility has significant consequences for studying physical activity, exposure to pathogens, and generating more realistic infectious disease models. Location-aware technologies such as Global Positioning System (GPS)-enabled devices are used increasingly as a gold standard for mobility research. The main goal of this observational study was to compare and contrast the information obtained through GPS and semi-structured interviews (SSI) to assess issues affecting data quality and, ultimately, our ability to measure fine-scale human mobility. A total of 160 individuals, ages 7 to 74, from Iquitos, Peru, were tracked using GPS data-loggers for 14 days and later interviewed using the SSI about places they visited while tracked. A total of 2,047 and 886 places were reported in the SSI and identified by GPS, respectively. Differences in the concordance between methods occurred by location type, distance threshold (within a given radius to be considered a match) selected, GPS data collection frequency (i.e., 30, 90 or 150 seconds) and number of GPS points near the SSI place considered to define a match. Both methods had perfect concordance identifying each participant's house, followed by 80–100% concordance for identifying schools and lodgings, and 50–80% concordance for residences and commercial and religious locations. As the distance threshold selected increased, the concordance between SSI and raw GPS data increased (beyond 20 meters most locations reached their maximum concordance). Processing raw GPS data using a signal-clustering algorithm decreased overall concordance to 14.3%. The most common causes of discordance as described by a sub-sample (n = 101) with whom we followed-up were GPS units being accidentally off (30%), forgetting or purposely not taking the units when leaving home (24.8%), possible barriers to the signal (4.7%) and leaving units home to recharge (4.6%). We provide a quantitative assessment of the strengths and weaknesses of both methods for capturing fine-scale human mobility.
| Being able to quantify human movement is important for studying activity patterns, exposure to pathogens and developing realistic infectious disease models. We compared fine-scale human mobility data obtained by Global Positioning System (GPS)-enabled devices and semi-structured interviews (SSI) from 160 individuals in Iquitos, Peru, in order to assess the quality of data using these two different approaches and our ability to measure fine-scale human mobility patterns in a resource-poor urban environment. Using various methods to process the GPS data, we found the SSI identified more locations a person had visited than GPS. Though the GPS gave more precise data, there were behavioral, technical, and analytical barriers. The SSI provided richer context and was easier to process, but also had more false positives. SSI was the only option for identifying locations retrospectively.
| Knowledge of daily and routine individual human mobility patterns within urban settings are important for urban planning [1]–[3], developing transportation models [3], promoting healthy lifestyles [4], and understanding infectious disease dynamics [5]–[13]. Measuring mobility at fine spatial and temporal scales through classic data collection methods (e.g., interviews, diaries, direct observations) presents significant challenges, such as marked heterogeneities in the ability of individuals to recall the locations they visit, changes in people's lives that affect their daily mobility (e.g., new partners, change of jobs, school vacation) as well as privacy issues [11], [14]. These challenges can be exacerbated in resource-poor settings [6], [7], [10], [15], [16], such as our study site in Iquitos, Peru, due to the lack of complete and updated address maps (affecting geo-coding of self-reported addresses) and limitations in spatial literacy of interviewed individuals [11]. There is an urgent need to develop and validate easily deployable and culturally-sensitive tools that characterize a person's routine mobility in order to link such information to health outcomes [6], [10], [13], [17], [18]. This is of particular relevance for understanding infectious disease dynamics, given the dominant role mobility has in driving infectious contacts and thus pathogen transmission, emergence, persistence and propagation [5], [6], [8]–[13], [18]–[22].
The wide availability of emerging location-aware technologies such as Global Positioning System (GPS)-phones or data-loggers provides new opportunities to quantify human mobility at fine spatial and temporal scales. Their use in research projects is feasible: they have decreased in cost and size, the technology has improved (i.e., GPS chipsets are more efficient in acquiring and fixing a signal as well as in power consumption) and the units are widely accepted by study populations [6], [10], . Over the past ten years, GPS tracking (often coupled with other sensors) has taken a prominent role in physical activity and exposure research [17], [26], [27]. Their implementation, however, in infectious disease research has been limited in part due to the challenges in linking the positional data generated by such sensors with temporally and spatially discrete locations (i.e., a person's home) where pathogen exposure occurred, and more importantly, the complexities associated with the analysis of the vast amount of data that these sensors can generate. A recent systematic review [17] shows that most studies using GPS to track physical activity involve few participants (<20), track individuals over short time periods (<12 days) and are focused on specific age groups (children vs. adults) or environmental correlates of activity (e.g., park vs. school movement) [17], [27]. GPS-based tracking presents enormous opportunities for improving our understanding of individual space-time activities and how they influence health outcomes, which has been done in various studies [6], [7], [10], [11], [15], [16], [28].
GPS technology, however, also has limitations that need to be addressed before considering it a “gold standard” for mobility research [26]. Rates of GPS data loss can reach 92% due to signal drop-outs, dead batteries, participants not wearing the units, signal loss during the initialization period or misuse of the device [17]. In Stothard et al.'s study in Uganda, the authors found that the track logs of the small, wearable GPS units (i-gotU) were accurate compared to a more sophisticated and costly unit (Garmin Oregon 550t) – discordance of <7 m for the 15 households tested – but there was GPS malfunction in units that was believed to be related to “insufficiently robust hardware for field conditions” possibly due to humidity or quality of the software [10].
As part of a larger study investigating risk for dengue (a human disease caused by a mosquito transmitted virus) in Iquitos, Peru, we simultaneously implemented two methods to capture fine-scale human mobility patterns: GPS data-loggers and semi-structured interviews (SSI). Dengue is a mosquito-transmitted viral disease of humans in tropical and subtropical regions of the world that is a rapidly growing public health problem [29], [30]. The main goal of this observational study was to compare and contrast the information obtained through these two methods to assess the issues affecting data quality, and identify strengths and weaknesses of each approach. We used two methods to analyze GPS data, and compared GPS results obtained via both methods with the results from the SSI.
Our study took place in Iquitos, a large and geographically isolated city in the Amazon Basin of northeastern Peru that is accessible only by boat or plane [31], between September 2008 and August 2010. The city of Iquitos has a high population density (∼390,000 inhabitants), and a very informal and dynamic economic structure (33.4% of those economically active are either unemployed or informally employed) [31]. As observed in other resource-poor cities, Iquitos lacks a unified and updated address system. Car access and public transportation are limited and residents rely on personal motorcycles, ∼20,000 motorized rickshaws [“moto-taxis”], and a few bus lines to move throughout the city. The major industries in the area are small commercial enterprises, fishing, oil, lumber, tourism, and agriculture [7]. Iquitos is the home-base of an extensive, ongoing, long-term project since 1999 led by the University of California at Davis/U.S. Naval Medical Research Unit 6-Iquitos group [5]–[7], [15], [16], [32] studying the environmental, entomologic, epidemiologic and behavioral determinants of dengue virus transmission.
Two methods for obtaining fine-scale human mobility data were simultaneously implemented: (1) GPS data-loggers (“i-gotU GT120”, Mobile Action Technology Inc.) and (2) semi-structured interviews (SSIs).
Descriptions of GPS features, spatial accuracy, acceptance by participants and device deployment associated to this study were reported previously [11], [12]. The main attributes of selected units were: (1) data storage capacity and battery life capable of recording at least 3 days of data; (2) high spatial accuracy (∼4–10 m); (3) durable, water resistant and tamper-proof; (4) light weight (<50 g); (5) carrying mechanism (lanyard around neck) widely accepted by participants of different ages/sex; (6) little to no maintenance required by study participants; (7) low cost ($49); and (8) password protection and a special socket for data download (to protect participant's confidentiality). The units are easily worn on a neck strap or in a pocket, and have been used to track routine movement patterns of Iquitos residents over the past three years with a high level of acceptance (98%) [16].
Based on the known limitations of classic interview instruments to capture overt behaviors in space and time [2], [33]–[35], and guided by findings from focus group discussions performed in Iquitos [16], we designed a SSI for capturing positional and temporal information of routine human mobility. Key findings from the focus groups that guided the survey development included [16]: (1) people could clearly identify many of the routine locations they visited, although they sometimes needed certain “triggers” for recall (and these were identified), (2) there were marked differences in reported mobility routines by gender and age groups; and (3) there were clear “common activity spaces” (markets, recreational spots, etc). The developed SSI contained one section listing commonly visited locations, such as markets, health facilities, and schools, and a section that used field-tested triggers to help people recall “individual” locations visited (such as relatives' houses) in the last 14 days. Participants also gave estimates of time spent in each location per week. High resolution satellite (Quickbird, Digitalglobe, CO) and digitized street maps were used during the interview to prompt recall and to mark the position of the places mentioned.
Participant recruitment was not random: we used purposive sampling and focused on two Iquitos neighborhoods participating in an ongoing longitudinal study on dengue epidemiology [32], [36], seeking a balanced number of males and females representing age ranges between 7 and 74 (see Table 1). We only excluded those who planned to spend more than a day outside of Iquitos during the following 14 days. Recruitment was performed by trained local technicians who provided a description of the study together with a pamphlet with specific information about the GPS units and the study in general [16]. In the first phase, conducted between September 2008 and March 2009, 59 participants were asked to use the GPS units at all times for a period of 14 days and respond to the SSI on day 15 asking for all the places they visited while GPS-tracked during those 14 previous days. The research team scheduled an exchange of the GPS units every three days to download data, verify function, and recharge batteries. At the time of GPS unit exchange, participants were asked about their experiences with the GPS, whether they had used it, if it had been forgotten and, if so, on what days. GPS units were programmed to track a person's position (latitude, longitude and time stamp) every 150 seconds. The second phase was conducted in July and August of 2010 with 101 participants, who were asked to follow the same procedures as before; use the GPS unit for 14 days and respond to the SSI on day 15. One component was added in this phase: within 3 days of data collection, survey data was entered into a database and GPS-collected data was processed so that information on the locations identified as visited by each method were overlaid in a Geographic Information System (ArcGIS 10, ESRI). With a series of maps noting the position of each place visited by either method, field technicians returned to the participants within 4–5 days to ask them about any discordant information (i.e., locations on the survey, but not registered on the GPS or vice versa). For Phase 2, the GPS collection frequency was increased to every 15 seconds (45 participants) and 90 seconds (56 participants) to assess the impact of data collection frequencies on GPS-SSI concordance. Whereas with 150 second programming, we could collect and recharge GPS units every 3 days, individuals wearing GPS units programmed at 15 and 90 seconds were provided with a charger and asked to charge the units daily because of the reduction in battery life. Our sample size was sufficient for a descriptive analysis and was limited due to intense participant follow-up for ∼20 days; i.e., recruiting and consenting, distributing GPS units, exchanging charged GPS units and collecting ones losing power, interviewing participants with SSI at day 14, geocoding locations immediately, inputting all data from GPS and SSI to overlay in a GIS, returning to participants for follow up interview. Considering these complexities, participant recruitment was limited to what was logistically feasible for our field teams.
All locations reported on the SSI were identified in the Iquitos GIS and received a unique location code with geographic coordinates that link directly to a SQL database containing participant information. If the location was not already in our system or if there were doubts about the specific location, a research team member went to the described place to assign a geo-code. Based on geo-referenced city-block maps (courtesy of the Peruvian Navy) and field sketch maps, geo-referenced aerial photographs and high resolution satellite imagery (Quickbird, Digitalglobe, CO), a total of 48,365 Iquitos lots were digitized prior to initiation of this study. Given the lack of a formal and consistent address system, we assigned a unique code to each lot. A local GIS specialist on our research team updates the maps on a regular basis, making the Iquitos GIS one of the most complete and up to date geo-spatial databases generated for a resource-poor city of its size.
To obtain locations recorded by GPS units, the raw data was processed using an agglomerative algorithm (i-Cluster [15]). In simple terms, when GPS raw data was plotted over a satellite image of the city, we observed “clouds” over specific locations that were frequented by an individual [15]. These “clouds” mark locations that are the product of the frequency of going to that place and the time spent there. This data reduction algorithm works by aggregating consecutive GPS readings that are within a spatial (d) and temporal (t) window, and estimating the total time a participant spent within such a spatio-temporal buffer [15]. The algorithm also allows for identification of locations intermittently visited by applying a threshold time (tintv) in between visits. Based on the inherent spatial error of GPS data (e.g., 5–10 m) we determined the following configuration: d = 20 m, t = 15 min and tintv = 30 min, for tracking Iquitos participants. The resulting place derived from the i-Cluster algorithm was then manually assigned the nearest location ID in the Iquitos GIS.
For the analysis, we directly compared the raw GPS data to the SSI data. Because we know the exact GPS coordinates of every location reported in the SSI data, we could test to see how frequently the GPS unit reported that the individual was in the vicinity of each location. Specifically, for every participant and every location they visited, we calculated the distance from every GPS point registered for that participant to that location. For many locations, we have not just the location, but the footprint of the structure as a polygon within the Iquitos GIS. As such, we could calculate the distance from each GPS point to the boundary of each location (taking GPS points that were within the polygon to have distance zero from the structure). For both locations that we have the footprint of the structure and those that we just have a single GPS location representing the centroid of the building, we consider the location “visited” if there are a sufficient number of raw GPS points within a certain threshold distance of the location. We then vary the number of raw GPS points deemed sufficient (here we used 1, 5, and 10 points), as well as the distance threshold selected (defined as the distance allowed for what constitutes a “match” between locations recorded in the SSI compared to a nearby GPS point, in this study, ranging from 0 to 100 meters), to investigate the sensitivity of visitation.
We quantified the concordance between SSI and GPS in identifying places visited by participants by comparing the interview locations with (1) i-Cluster-derived locations and (2) raw GPS positions. To compare the interview with the i-Cluster inferred locations we mapped the locations identified by each method in a GIS (ArcMap 9.3; ESRI). Locations identified both by the GPS and the SSI were considered “concordant” and did not require follow up. All locations that were captured by either GPS or the SSI, but not both, were considered “discordant” and a research assistant was sent back to the participant's home to ask them about the potential causes of discordance. Before interviewing each participant, the research assistants checked the original SSI to determine how the respondent had described the location (e.g., “aunt's house” or “internet cabin”) or the GIS maps to locate a nearby reference point that might help the participant identify each discordant location (e.g., 2 blocks from market). Research assistants (nurses and biologists) were native Iquitos residents who received specific training on all steps of the interview process to ensure they were aware of sensitive issues they might encounter both when gathering initial SSI information, as well as while following up with discordant locations.
Participants were given a 24–48 hour period to decide whether to participate or not in the study. For children, verbal assent of the minor and written consent of the parent or caretaker were required, whereas for adults, a written consent was required. After GPS data collection, a strict protocol for storage (in a secure MySQL database) and management was followed. The procedures for enrollment of participants and GPS data management were approved by the Institutional Review Boards (IRB) of the University of California at Davis (2007.15244), Emory University (IRB9162) and Tulane University through an inter-institutional IRB agreement with the United States Naval Medical Research Center Unit No. 6 (NAMRU-6). The NAMRU-6 IRB, located in Peru, also reviewed and approved the study (NMRCD 2007.0007). This IRB functions as a Peruvian IRB and is registered with the Peruvian Regulatory Agency for Clinical Trials with the number RCEI-78.
More than half of the 160 enrolled participants were females (58.5%) (Table 1). The lower number of males was due to the difficulty in finding them at home during regular interviewing hours. Recruitment was stratified by age; the age range sampled was 7 to 74 years. Recruitment varied across age groups (range of 25–46 per age group), with 7–18 year olds accounting for 28% of the tracked individuals (Table 1). Although not perfectly balanced among sexes and age groups, the recruited population represents a large and diverse demographic sample of the local population.
Of the 2,566 locations identified by SSI and/or i-Cluster algorithm, 14.3% were concurrently identified by both (i.e., concordant). SSI identified 2.3 times more locations than the i-Cluster algorithm, with residential (42.5%), commercial (26.4%) and educational (10.8%) spaces accounting for the highest degree of concordance between methods (Table 2). A total of 2,047 places were reported in the SSI as visited by all participants over the 14-day tracking period (of these 2047 places mentioned, 1057 were unique places, see Table 2). Most (96.7%) places were located within the urban and peri-urban areas of Iquitos (Figure 1A). Participants reported visiting a median (Q1–Q3) of 12 (9–16) places over the 14-day period, with the number of places not differing significantly between sexes (Wilcoxon rank sum test with continuity correction, W = 3140.5, P = 0.89). The most commonly reported location types on SSI that were not visualized using the i-Cluster algorithm (considering 1609 locations with land-use information) were commercial locations (34.2%) followed by residential (22.1%) and recreational (17.0%) locations (Table 2). The i-Cluster algorithm identified a total of 886 places as visited by participants while tracked (716 unique locations); 98.7% of which were found within the urban and peri-urban areas of Iquitos (Figure 1B). A significantly lower median (Q1–Q3) number of places per participant was registered by the i-Cluster algorithm in comparison to the SSI (7, 4.0–10.0; W = 11990, P<0.001). Residential spaces represented 58.6% of the 454 i-Cluster-identified locations with land-use information that were not reported on the SSI, followed by commercial (11.4%), educational (4.2%), and recreational (3.5%) locations (Table 2). Locations with highest percentage of concordance (i.e., per type of location, the number of concordant sites divided by the total number of sites obtained for that type of location through SSI and/or GPS) were educational settings (24%), followed by residential (19%), other (18%), and religious or market spaces (both at 13%).
When the SSI-reported locations were compared to the raw GPS data (Figure 1), differences in the concordance between methods were observed based on the location type, distance threshold selected, GPS collection frequency and number of GPS points considered to define a visit (Figure 2). Both GPS and SSI had perfect concordance in identifying each participant's home (see Figure 2) at either combination of collection frequency, distance or number of points. There was more concordance for residential sites than non-residential sites at 15 and 90 seconds collection frequency; this difference was minimal at 150 seconds (Figure 2). Not depicted due to the small numbers in each category, there was much variation in concordance when examining by type of location. For example, when examining specific categories such as schools, “other” (ports, storage buildings, empty lots) and lodging places (i.e., rustic “hostels” for visitors from outside Iquitos, or couples might go for a few hours) there was a concordance of 80–100% between methods, whereas other residential places (i.e., friends' or relatives' homes), commercial locations (i.e., shops, markets) and religious buildings (i.e., churches) showed a concordance of 50–80%.
As distance from the SSI reported location increased, the concordance between SSI and raw GPS data increased, independently of the type of location (Figure 2). When at least one raw GPS point was considered (solid lines in Figure 2), concordance between methods was highest at up to 20 meters from each location. Beyond that distance, no dramatic increases in concordance were observed. There was less concordance when we restricted our analysis to 5 GPS points (broken lines) or 10 points (finely broken lines), but the pattern was similar to the line created when 1 point was considered a match. Interestingly, increasing frequency of GPS data collection from 150 to 15 seconds was not associated with a proportional increase in concordance between SSI and GPS (Figure 2). Battery power loss observed at 15 second collection frequency may help explain such results: of the 508 GPS exchanges performed, 56 (11%) of GPS units programmed to collect data every 15 seconds had issues due to battery loss at the time of data download in comparison to 2% (9/379) for GPS units programmed to collect data every 150 sec.
At 20 meters from each SSI location, and when 1 GPS point was considered to define a match, overall concordance averaged 72.6% (SD: 20.7%) for 15 seconds, 65.8% (30.8%) for 90 seconds and 70.3% (23.3%) for 150 seconds (Figure 3). When ten points were required to define a match, concordance was reduced to 59.1% (31.6%), 54.3% (31.0%), and 55.7% (30.7%) respectively (Figure 3). Cemeteries, public buildings, recreational areas and health centers were the location types that consistently showed the lowest concordance values (Figure 3). Increasing the data collection frequency from 15 to 150 seconds did not translate into significant variation in the concordance between SSI and GPS across all location types (average [min-max] variation across locations, 2.3% [0.7%–9%]) (Figure 3).
In comparison to using the raw GPS points (Figure 2), the i-Cluster algorithm evidenced much higher discordance rates for all location types (Table 2). However, this method allowed identifying a total of 519 locations not mentioned in the SSI and not able to be inferred when the raw GPS positions were visualized (Table 2).
In Phase 2, with the subset of 101 participants, we further explored the possible causes of discordance between GPS and SSI. Specifically, within 2–3 days of administering the SSI, we used GIS to develop maps identifying “discordant” SSI and i-Cluster locations (Figure 4) (i.e., locations that were only mentioned in the SSI or only visualized using the GPS data). These maps were used when probing participants about possible causes of discordance. In this phase, regarding locations identified on the SSI, but not detected by GPS (total of 656 locations, Table 3), the most common response to questions about the discordance was an affirmation that these locations had been visited (35.8%) – they could not explain the discordance. The second most common response was that units had “seemed to be turned off” (30%). Indeed, GPS units initially deployed could accidentally be turned off, so respondents who noticed the lack of a flashing blue light inferred correctly. Once this problem was reported, we programmed GPS units to not allow them to be turned off manually, reducing this problem half-way through this study. Other explanations for the discordance included those who admitted forgetting to take units to some locations (12.5%; i.e., rushing out and simply forgetting), not wearing the GPS units to locations that were near their house (3.4%) or to locations where they might get stolen (3.5%), and leaving units home to recharge (4.6%). A small percentage (4.7%) affirmed having the GPS unit in some locations, but questioned whether the placement of the GPS unit in their purse might have impeded the signal.
Regarding locations identified on the GPS unit but not mentioned in the SSI (204 locations, Table 3), the most common response was that they simply forgot to mention it in the SSI (38.2%), and a few made the additional observation that they had forgotten this location because it was not part of the regular routine (15.2%). Some locations were not mentioned (until probed directly about them) because they were either transient or en route to another location (22.1%; i.e., a path always taken, a bus stop) or because they were outdoors (13.2%; i.e., outdoor food kiosk). After further examination of the reasons for discordance between SSI and GPS, we identified 75 locations as being affected by technical failures in generating the maps (the locations were not properly mapped or marked the location next door, 57.3% and 42.7% respectively, and hence were incorrectly considered discordant at the time of interview).
GPS technology is increasingly used in behavioral research. Its use has moved beyond feasibility tests [15], [35], [37], [38] to the actual use of GPS-enabled devices (often coupled with other sensors such as accelerometers, air pollution sensors or cameras) in studies quantifying various aspects of human mobility and spatial behavior [7], [10], [11]. As the technology continues to be embraced by researchers across disciplines, it is easy to assume that due to the wealth and resolution of the data it provides, some might consider GPS data to be a “gold standard” for mobility research and a replacement of classic survey instruments [35]. By performing a field validation study tracking 160 individuals, we assessed both the limitations and possibilities of GPS technology for mobility research, and provided evidence of multiple sources of error/uncertainty that can affect quality of data in comparison to survey methods. It is important to mention here that based on our experience, we would expect different results with different GPS units, different SSI and other methods of data analysis.
Under perfect conditions of satellite geometry and signal strength, GPS provides very accurate information about the position (latitude, longitude, elevation, time of day) of any stationary object on earth. Wearable GPS devices provide all the essential pieces of information to reconstruct and quantify human movement: positions associated to places visited, time stamp for each potential visit, and routes followed to connect visits. Given technical (e.g., signal noise, multipath errors, signal obstruction inside buildings, battery life) and human behavioral limitations (e.g., compliance of use, individuals forgetting to take or charge units), GPS signals are prone to error and estimates of mobility parameters that they generate are considered uncertain. Signal processing algorithms have been developed to reduce such errors and improve interpretation of complex data [39]–[44]. In our study, the application of a signal clustering algorithm (i-Cluster) allowed identifying locations where individuals spent their time, but also added significant uncertainty by flagging locations transiently visited (e.g., a bus stop; 35.3%). Such errors were the main contributor to the 85.7% discordance between methods observed when i-Cluster inferred locations were considered. Because most research describing automated algorithms rely on single (or few) days of data or low sample sizes [39]–[44], the errors found by our study are a likely outcome of the type of error those algorithms may encounter if applied within the same context. Our results can be used as a guide for the development of improved and more accurate methods for GPS location extraction and human movement quantification.
An interesting finding was that higher GPS collection frequencies (e.g., 15 seconds) were not associated with a proportional and significant increase in concordance between methods. Issues of battery life, not securing the “off” option at the start of the study (remedied quickly), and compliance of participants in charging the units compromised the quality of data collected. Similar issues were observed across multiple studies quantifying physical activity [17], [35]. Implementing GPS-enabled smart-phones could have reduced the issue of battery loss, because there is more motivation for individuals to charge the phones overnight and to use them during the day. Because Iquitos is slowly making its transition into smart-phone technologies, different issues were pointed out by a subset of 10 participants when asked about the possibility of using GPS-phones instead of data-loggers: (a) older individuals were intimidated by the technology and by the possibility of having the units stolen (the latter was a concern shared by individuals across all age groups), and (b) school age children mentioned they are not allowed to take phones to elementary or high school or locations where their phones could be taken by older children [7]. When cell-phones can be properly deployed they can provide valuable information. For instance, in Canada a study comparing GPS data collected by cell-phones and self-reported surveys reported (using rudimentary indices of concordance such as convex hulls and kernel density estimations) that 75% of questionnaire-reported activity locations were located within 400 meters of an activity location recorded on the GPS track [26]. In weighting the possibility of adopting novel technologies, consideration of cultural and local concerns will be key for both GPS and SSI instruments [13], [16], [18].
Turning the large amounts of raw GPS positional data into meaningful locations individuals visited is another challenge. Unprocessed raw GPS data can be used to either describe zones or areas in which individuals spend their time or to assess the accuracy of the GPS in identifying precise locations against information provided by another method (i.e., locations identified by SSI). In our study we implemented a simple algorithm based on an agglomerative clustering method (i-Cluster) to identify locations visited by individuals carrying a GPS unit. Our analysis shows that the algorithm presents low levels of sensitivity and specificity in identifying places reported as visited by participants. This poor performance could be due to: (a) the algorithm's limited ability to account for changes in accuracy of the GPS signal or to the occurrence of intermittent positions as a consequence of GPS signal loss and (b) the fact that not all reported locations were actually visited by participants while tracked. More complex methods of location extraction that account for signal errors, such as hierarchical dynamic Bayesian network models [39], [41], [44], are being currently developed and are viewed as a promising means of reducing the uncertainty associated with the identification of locations visited by participants [39], [44]. Once those methods are validated, their integration into health research applications will increase our ability to accurately infer the location of potential infectious disease exposure areas.
Classic methods (surveys, diaries) have long been considered too limited to quantify behavior due to marked heterogeneities in the ability of individuals to recall the locations they visit, interviewer error, behavior changes and issues associated to privacy [35]. By working with the local community, addressing potential cultural barriers and concerns and adapting the language of interviews, we developed a culturally-sensitive SSI to quantify movement (and potential exposure to dengue). Our comparative analysis shows that, for a 14-day recall period, interviews provide accurate estimates of the locations visited by people (of a total of 892 locations for which we investigated causes of discordance, only 109 [12.2%] were visited and not reported). The SSI not only identified places, but also characterized the context of visits (i.e., grandmother's house), information impossible to obtain directly from GPS. SSI data entry and processing are much more straightforward and faster than of GPS: (a) maps with marked locations were digitized in the Iquitos GIS and each premise reported as visited was assigned a location code and (b) the location code was then linked to the database containing all the SSI information. We concluded that a validated survey instrument that can be adapted to different contexts can be used to understand the role of human mobility in infectious disease dynamics.
We encountered several limitations in our study design. Although our sample size was relatively large, the low numbers of participants assigned to each age group precluded statistical tests to look at different causes of discordance. Given that we needed to obtain results quickly to ask participants about possible causes of discordance, we relied on a single GPS data reduction algorithm (i-Cluster). As observed on the survey (Table 3), most of the discordant records occurred due to this algorithm providing false positive or negative results. Since the time this study was performed, new and more sophisticated methods to process GPS data have been developed [39]–[44], as well as more accurate and less error-prone GPS units have likely become available. Future research will involve performing comparative studies to quantify sensitivity/specificity as well as applicability to specific study questions. Also, we considered that our concordance estimates could be, in part, dependent on size and placement of houses in Iquitos. An average household in this city measures 5 m in width, which is within the mean error of a GPS (5–10 m). This could explain the high percent (∼60%) of residences identified by GPS that were not reported on the SSI. Thus, accuracy in identifying locations is not only dependent on the factors explained above, but also on key attributes of the urban landscape (e.g., household size, prevailing building material, density of high-rise buildings, vegetation cover). We did not test for differences in the SSI results of participants with more contact with our research team (i.e., those with more frequent GPS exchanges due to differing data collection times) compared to those with minimal contact. We do not expect differences, however, because contact was focused on the GPS exchange and SSI questions about their movement and activities were only asked at the end of the 14 day period. None of the participants, therefore, had an advantage over others regarding the types of questions they would be asked. We also did not estimate nor compare the cost and technical expertise to apply and process by these methods. Both the GPS and SSI capture very complex data. GPS data is in digital form, but needs to be processed. SSI data needs to be verified (i.e., in our study, someone might go to a location described to geocode the location), entered and mapped. There were costs associated to purchasing GPS units (∼$49/unit), training personnel to set and distribute units, downloading and analyzing the GPS data. Similarly, there were costs associated to developing, refining and improving the SSI, training personnel to apply it, and entering the data in a GIS system. Ultimately, decisions regarding using an SSI or GPS units in a study depend strongly on the study question and the urban context, because both SSI and GPS can provide different but equally valuable information that need to be carefully weighted at the planning stage.
For infectious diseases in general, and vector-borne diseases in particular, the need to tie potential exposure to specific locales requires the retrospective investigation of multiple routes of pathogen transmission. Survey instruments like the one we developed in this study not only provide accurate information of places visited, but can also be used to retrospectively infer the likely location where infection occurred [5]. This need to tie exposure to a specific place(s) has limited the use of GPS technology in infectious disease research, but GPS technology could be used in prospective movement studies or in studies obtaining information provided by phone companies. As observed in our study, once locations are identified, the raw GPS positions can be analyzed to quantify temporal patterns of mobility (days and times a person visits such locations, regularity of visits, overlap with other tracked individuals) and to accurately quantify routines and movement of a large segment of a population. This way, key information about mobility and behavior can be inferred and used to parameterize mathematical models that allow better forecasting of disease transmission or design policies targeting activities or segments of the population at greatest risk.
No gold standard exists for obtaining and analyzing human mobility data, instead different errors may occur with different methods. Despite the continually improving accuracy available with GPS, barriers persist, including: behavioral aspects (i.e., people remembering to use the unit), technical aspects (i.e., accuracy of 5–10 meters in a location with houses averaging 5 meters width), and analytical aspects (i.e., differences in concordance based on method of analyzing complex data as reported in this article). The SSI is not a gold standard either. Even with the possible drawback of more locations reported than true (i.e., false positives), compared to GPS units, the SSI provided more true locations, more context about locations, and data were easier to process and analyze. For our study, in which we needed to identify locations retrospectively for possible exposure to dengue virus, the SSI was the only choice because of the logistical and financial difficulty of fitting GPS units on a large sample and, even if that had been possible, being able to quickly identify locations recently visited within a short enough time frame to initiate our possible exposure investigations. For now, SSI remains the most comprehensive method to identify such locations.
|
10.1371/journal.pntd.0000507 | Importance of Coverage and Endemicity on the Return of Infectious Trachoma after a Single Mass Antibiotic Distribution | As part of the SAFE strategy, mass antibiotic treatments are useful in controlling the ocular strains of chlamydia that cause trachoma. The World Health Organization recommends treating at least 80% of individuals per community. However, the role of antibiotic coverage for trachoma control has been poorly characterized.
In a collection of cluster-randomized clinical trials, mass oral azithromycin was administered to 40 villages in Ethiopia. The village prevalence of ocular chlamydia was determined before treatment, and at two and six months post-treatment. The mean prevalence of ocular chlamydia was 48.9% (95% CI 42.8 to 55.0%) before mass treatments, decreased to 5.4% (95% CI 3.9 to 7.0%) at two months after treatments (p<0.0001), and returned to 7.9% (95% CI 5.4 to 10.4%) by six months after treatment (p = 0.03). Antibiotic coverage ranged from 73.9% to 100%, with a mean of 90.6%. In multivariate regression models, chlamydial prevalence two months after treatment was associated with baseline infection (p<0.0001) and antibiotic coverage (p = 0.007). However, by six months after treatment, chlamydial prevalence was associated only with baseline infection (p<0.0001), but not coverage (p = 0.31).
In post-hoc analyses of a large clinical trial, the amount of endemic chlamydial infection was a strong predictor of chlamydial infection after mass antibiotic treatments. Antibiotic coverage was an important short-term predictor of chlamydial infection, but no longer predicted infection by six months after mass antibiotic treatments. A wider range of antibiotic coverage than found in this study might allow an assessment of a more subtle association.
| Trachoma, caused by ocular chlamydia infection, is the most common infectious cause of blindness in the world. The World Health Organization (WHO) recommends the SAFE strategy (eyelid surgery, antibiotics, facial hygiene, environmental improvements) for trachoma control. Oral antibiotics reduce the transmission of ocular chlamydia, but re-infection of treated individuals is common. Therefore, the WHO recommends annual mass antibiotic treatments to the entire village. The success of treatment is likely based on many factors, including the antibiotic coverage, or percentage of villagers who receive antibiotics. However, no studies have analyzed the importance of antibiotic coverage for the reduction of ocular chlamydia. Here, we performed multivariate regression analyses on data from a clinical trial of mass oral antibiotics for trachoma in a severely affected area of Ethiopia. At the relatively high levels of antibiotic coverage in our study, coverage was associated with post-treatment infection at two months, but not at six months. The amount of infection at baseline was strongly correlated with post-treatment infection at both two and six months. These results suggest that in areas with severe trachoma treated with relatively high antibiotic coverage, increasing coverage even further may have only a short-term benefit.
| The World Health Organization (WHO) recommends the SAFE strategy (eyelid surgery, mass antibiotics, facial hygiene promotion, and environmental improvement) for the control of trachoma, the world's leading infectious cause of blindness[1]. Mass antibiotic treatments target the ocular strains of chlamydia that cause trachoma, and are a crucial component of the SAFE strategy. A single dose of oral azithromycin is clearly effective in eliminating infection from individual cases[2],[3]. A mass distribution of azithromycin to an entire community has been shown to dramatically reduce the prevalence of infection. Unfortunately, infection returns in areas with hyper-endemic trachoma[4]–[6]. Theoretically, repeated treatments can progressively reduce the prevalence of, and even eliminate, infection[7]. However, models suggest that in severely affected areas, treatment would have to be given frequently and to a large portion of the population[8],[9].
There have been few studies examining the role of antibiotic coverage for trachoma control, aside from the observation that mass antibiotic treatments with high coverage have resulted in a considerable reduction in ocular chlamydia prevalence, and even elimination of infection[10],[11]. Many think that low antibiotic coverage may play a crucial role in persistent ocular chlamydial infection[12]. Currently, the WHO recommends a goal of 80% antibiotic coverage for trachoma programs[1]. However, the relationship between antibiotic coverage and treatment efficacy at the community level has not been well characterized. Mathematical models have suggested that at higher antibiotic coverages, less frequent mass treatments will be required, and elimination will occur in a shorter period of time[7],[9]. It is not clear what level of coverage will be necessary, or whether different guidelines will be necessary for more severely affected areas[5],[8]. Here we assess how the prevalence of infection two and six months post-treatment is dependant on the antibiotic coverage and the amount of endemic infection at baseline.
The Committee on Human Research at the University of California, San Francisco approved this post-hoc analysis of existing data, and approved the use of verbal informed consent, which was obtained by local Amharic-speaking health workers from all study participants at the time of each procedure. Verbal consent was performed due to the high amount of illiteracy in the region.
As part of a larger, multiple arm, group-randomized trial, 40 villages were enrolled in the Gurage Zone of southern Ethiopia[6],[7],[13],[14]. Although the 40 villages were distributed between five study arms, they received identical treatment and monitoring for the initial six months, the results of which are reported here. An initial census of the study area was performed by trained local health workers; the names of all permanent residents in each village were recorded. The population of villages ranged from 122 to 976, with a median of 368 persons (interquartile range 243 to 502). Those aged 1 year and older were offered a single dose of directly observed, oral azithromycin (1g in adults or 20 mg/kg in children). Pregnant women and those allergic to macrolides were offered a six-week course of topical 1% tetracycline ointment (applied twice daily to both eyes and not directly observed). Antibiotic coverage was defined as the proportion of permanent residents eligible for treatment in the village (i.e., those ≥1 year of age) who accepted directly observed treatment with oral azithromycin or not-directly observed topical tetracycline, as determined from the baseline census by the health workers who distributed the antibiotics. Individuals known to have either moved permanently or died between the census and the scheduled treatment were not included in the denominator. Children aged 1–5 years were monitored at baseline, two months, and six months post-treatment, as described below. After the six month monitoring, some communities received biannual treatment, some annual treatment, and some no further treatment unless infection surpassed a pre-assigned level. The results of these trials are published elsewhere[6],[7],[13],[14].
All children aged 1–5 years in treated villages were assessed for the presence of ocular chlamydia infection at baseline (pre-treatment), two months post-treatment, and six months post-treatment. A dacron swab was passed firmly across the right upper tarsal conjunctiva three times, rotating between each pass. Examiners wore new gloves for each study subject. All samples were kept at 4°C in the field and frozen at −20°C within six hours. The swabs were shipped at 4°C to San Francisco where they were stored at −70°C until processed. The AMPLICOR PCR test (Roche Diagnostics, Branchburg, NJ) was used to detect chlamydial DNA. Pre-treatment samples were tested individually, and post-treatment samples were analyzed as pooled samples. PCR pooling is a well established, cost-effective technique for diagnosis of genital and ocular chlamydia[15],[16]. Post-treatment samples from the same village were randomized and pooled into groups of 5, with a possible remainder pool of 1–4 samples. Each pool was then tested according to the AMPLICOR protocol. If two-thirds or more of the pools were positive, the individual samples were re-pooled randomly into groups of two and re-processed to allow more accurate estimation[16]. If PCR of any pool was equivocal, then all samples from the pool were individually re-tested. As per the AMPLICOR protocol, an internal control was performed for each pool to rule out the presence of PCR inhibitors. Any inhibitory pools were re-tested, and if still inhibitory, the samples were tested individually. While samples necessarily were diluted in the pooling process, this is not thought to significantly impact the sensitivity of the test[17]. The prevalence of ocular chlamydia infection in each village was obtained by maximum likelihood estimation[7]. The number of positive individual samples most likely to have resulted in the observed pooled PCR results was chosen as the estimate for that village (Mathematica 5.0, Wolfram Research Inc., Champaign, IL).
All statistical analyses were conducted at the village level using village prevalence; no individual-level data was used in this study. This is appropriate, since trachoma interventions occur at the village level, and treatment success is measured at the village level, not at the individual level. The distribution of the prevalence of antibiotic coverage was depicted with a kernel density plot, using the Epanechnikov kernel function, the Sheather-Jones plug-in bandwidth estimate, and upper boundary correction using the renormalization method. The Wilcoxon signed rank test was used to compare the prevalence of infection at baseline with two months, and at two months with six months. The Spearman rank order correlation coefficient and 95% confidence interval was calculated for pairwise combinations of baseline infection, antibiotic coverage, infection at two months after treatment, and infection at six months after treatment. Multivariate regression was performed to assess the relationship between the prevalence of infection post-treatment with antibiotic coverage and endemic (baseline) infection. Linear regression models were constructed using the prevalence of chlamydial infection at either two or six months as the response variable, and baseline prevalence of infection and antibiotic coverage as explanatory variables, using the robust variance calculation based on the HC3 heteroskedasticity consistent covariance matrix estimator, due to concerns about heteroskedastic residuals[18]. Infection prevalence at all time points was square-root transformed to minimize heteroskedasticity and maximize normality of the residuals from the linear regression analysis (analyzed by plotting the residuals vs. the fitted values and residuals vs. the predictors; in addition, no heteroskedasticity was demonstrated with the Cook-Weisberg test, and no departure from normality was observed with the Shapiro-Francia test, using a significance level of 0.05). Linearity of the predictors in the model was adequate, as assessed with component-plus-residual plots comparing the linear fit of the predictor to the LOWESS curve. Multivariate regression models were constructed including the multiplicative interaction term for antibiotic coverage and baseline infection, but interaction terms were not significant, and therefore not included in the final model. All statistical analyses were performed with STATA 10.0 (Statacorp, College Station, TX).
The mean number of children ages 1–5 examined in each village at baseline was 54.2 (95% CI 45.7 to 62.8). No villages were lost to follow up. The mean pre-treatment prevalence of infection in 1–5 year old children among the 40 study villages was 48.9% (95% CI 42.8 to 55.0%). Antibiotic coverage data was present for 38 of the study villages, and ranged from 73.9% to 100%, with a mean of 90.6% (95% CI 88.7 to 92.4%). As is evident in a density plot, the majority of villages had an antibiotic coverage between 80–100% (Figure 1). Two months after treatment, infection decreased significantly from baseline, to a mean of 5.4% (95% CI 3.9 to 7.0%), p<0.0001. Between two and six months after treatment, the village prevalence of infection increased, to a mean of 7.9% (95% CI 5.4 to 10.4%), p = 0.03, compared to two months).
Using Spearman's test of correlation, prevalence of infection in 1–5 year old children at two months was strongly correlated with baseline infection (rs = 0.62, 95% CI 0.39 to 0.78), and moderately correlated with antibiotic coverage (rs = −0.31, 95% CI −0.58 to 0.01). The prevalence of infection in 1–5 year old children at six months was strongly correlated with infection at baseline (rs = 0.55, 95%CI 0.28 to 0.73) and at two months (rs = 0.73, 95% CI 0.54 to 0.85), but only weakly correlated with antibiotic coverage (rs = −0.16, 95% CI −0.46 to 0.17).
Multivariate regression models demonstrated that at two months after treatment, chlamydial infection in 1–5 year old children was predicted by both baseline chlamydial infection and antibiotic coverage (R2 = 0.53, Table 1). By six months after treatment, baseline chlamydial infection remained a significant predictor of chlamydial infection, but antibiotic coverage did not (R2 = 0.35, Table 1). Because the square root transformation of the response variable made the regression coefficients difficult to interpret, we used the models to calculate the role of antibiotic coverage in predicting post-treatment chlamydia in a hypothetical community, holding the baseline prevalence of infection constant at 48.9% (the mean baseline infection in this study). As depicted in Figure 2, antibiotic coverage had a greater effect in predicting chlamydial prevalence at two months compared to six months, evident from the steeper curve and narrower 95% confidence intervals.
Mathematical models and clinical trials have demonstrated the importance of vaccine coverage for conveying immunity on a population[19],[20]. Analogously, antibiotic coverage has been touted as an important determinant in the long-term success of the WHO's mass antibiotic treatments for trachoma[5],[7],[9],[21],[22]. Some have suggested that a single mass antibiotic treatment may prevent infection from returning, if given to a sufficiently large proportion of the community[23],[24]. In this study, coverage was important at two months after treatment, but we were unable to demonstrate its importance at six months.
In our study, baseline infection was a significant predictor of chlamydial infection at both two months and six months, whereas antibiotic coverage predicted chlamydial infection only at two months. Thus, in this severely affected area, the amount of endemic infection appears to be a stronger determinant of chlamydial infection than antibiotic coverage. This may be the case for at least three reasons. First, communities with more initial infection will tend to have more residual infection after an incomplete mass treatment, as demonstrated by mathematical models[8]. Secondly, re-infection after mass treatments likely occurs more rapidly in areas with severe trachoma, due to underlying transmission characteristics in these areas, such as poor hygiene and sanitation, travel to untreated communities, genetic variation among chlamydial strains and a myriad of other risk factors[25]–[27]. This suggests that in the long run, the forces that result in the return of ocular chlamydia into severely affected communities may overwhelm any temporary advantage conferred by high antibiotic coverage[26],[28],[29]. Third, children under 1 year of age were not treated with oral antibiotics in this study, and may therefore have served as a reservoir for infection. It is possible that the level of endemic infection is an indicator of infection in these untreated children. If so, then our finding of a strong relationship between endemic infection and post-treatment infection may indicate that chlamydial transmission from this young age group was more important than antibiotic coverage in predicting the prevalence of chlamydial infection after treatment.
If chlamydial infection does depend more on endemic infection than antibiotic coverage, this would not support devoting more resources to increasing the target antibiotic coverage from WHO guidelines, which currently recommend 80% antibiotic coverage. It is important to note that this regression model should not be extrapolated outside the range of our data; therefore, this conclusion may be generalizable only to severely affected areas with relatively high coverage. It is possible that antibiotic coverage may carry more importance in areas with milder trachoma and slower return of chlamydia after mass treatments.
This analysis supports the theory that treatment frequency and duration could be tailored to areas based on pre-treatment prevalence. Currently, the WHO recommends three annual mass treatments to trachoma-endemic areas, with re-evaluation after the third treatment[1]. This strategy has proven very successful in nearly eliminating infection in an area with a modest-moderate amount of trachoma[30]. However, fewer treatments may be sufficient in communities with hypoendemic trachoma and low pre-treatment chlamydial prevalence[24]. In contrast, communities with hyperendemic trachoma and high pre-treatment chlamydial prevalence may require a greater number of treatments, or more frequent treatments, as suggested by mathematical models[7]–[9],[13] and clinical trials[13],[14],[31]. In this study, chlamydial prevalence after mass antibiotic treatments was strongly predicted by the pre-treatment prevalence of infection, which supports the idea that villages could be stratified by pre-treatment chlamydial prevalence and offered a tailored mass antibiotic treatment regimen. Further research is needed to determine whether this could be a feasible strategy.
Few previous studies have addressed the question of antibiotic coverage for ocular chlamydia, aside from noting the success of mass antibiotic efforts with high coverage[10],[11],[32]. However, our findings are consistent with a study of mass azithromycin in multiple villages in the Gambia[29]. In the Gambian study, re-emergence of chlamydial infection in a subset of villages could not be explained by antibiotic coverage, which averaged 83% among the villages. These findings are consistent with the possibility that at the 80% antibiotic coverage target currently recommended by the WHO, other predictors become more important than differences in coverage.
There are several limitations of this study. We focused on one component of the SAFE strategy: mass antibiotic distributions, and are unable to comment on the role of antibiotic coverage in the setting of other trachoma interventions. We analyzed predictors of chlamydial infection at the community level, and therefore cannot make any conclusions regarding individuals in the community. However, individual-level data are not particularly helpful for trachoma programs, which make treatment decisions based on community indicators of trachoma. There was a relatively short surveillance time of six months, at which point some of the communities had scheduled re-treatments. Migration in the study area was not studied. The range of antibiotic coverage was relatively narrow, which may have decreased the likelihood of finding a significant effect of antibiotic coverage in the regression models, and may decrease the generalizability of the study. Finally, since communities in this study were not randomized to a pre-specified antibiotic coverage, unmeasured confounders may have affected the regression analyses. We would expect, however, that these confounders would have biased toward an association between antibiotic coverage and infection, since those communities where it is difficult to attain high antibiotic coverage may also have conditions that enhance the transmission of ocular chlamydia.
In conclusion, in post-hoc analyses of a large clinical trial of mass azithromycin for trachoma, we found that antibiotic coverage predicted chlamydial infection at two months after treatment, but not at six months after treatment. Far more important was baseline chlamydial infection, which was a strong predictor of infection at both two months and six months. This suggests that the WHO's recommendation of 80% antibiotic coverage is reasonable, and that trying to further increase antibiotic coverage may be less successful than targeting more intensive treatments to highly prevalent communities. Clinical trials in which communities are randomized to different antibiotic coverage levels will be important to more fully characterize the relationship between antibiotic coverage and chlamydial infection after mass antibiotic treatments.
|
10.1371/journal.pbio.0050256 | Interlocking Transcriptional Feedback Loops Control White-Opaque Switching in Candida albicans | The human pathogen Candida albicans can assume either of two distinct cell types, designated “white” and “opaque.” Each cell type is maintained for many generations; switching between them is rare and stochastic, and occurs without any known changes in the nucleotide sequence of the genome. The two cell types differ dramatically in cell shape, colony appearance, mating competence, and virulence properties. In this work, we investigate the transcriptional circuitry that specifies the two cell types and controls the switching between them. First, we identify two new transcriptional regulators of white-opaque switching, Czf1 and white-opaque regulator 2 (Wor2). Analysis of a large set of double mutants and ectopic expression strains revealed genetic relationships between CZF1, WOR2, and two previously identified regulators of white-opaque switching, WOR1 and EFG1. Using chromatin immunoprecipitation, we show that Wor1 binds the intergenic regions upstream of the genes encoding three additional transcriptional regulators of white-opaque switching (CZF1, EFG1, and WOR2), and also occupies the promoters of numerous white- and opaque-enriched genes. Based on these interactions, we have placed these four genes in a circuit controlling white-opaque switching whose topology is a network of positive feedback loops, with the master regulator gene WOR1 occupying a central position. Our observations indicate that a key role of the interlocking feedback loop network is to stably maintain each epigenetic state through many cell divisions.
| The opportunistic fungal pathogen Candida albicans can switch between two heritable states—the “white” and “opaque” states. These two cell types differ in many characteristics, including cell structure, mating competence, and virulence. Recent studies of the molecular mechanism of regulating the white-opaque switch identified a master transcriptional regulator, Wor1. In this study, we identified two transcriptional regulators, Czf1 and Wor2, as new regulators of white-opaque switching. By constructing a series of single and double mutants and by examining where the master regulator Wor1 binds throughout the genome, we generated a molecular model of the bistable switch that regulates white-opaque switching. The regulatory model consists of interlocking positive feedback loops, which mutually reinforce one another and stabilize the opaque state. These results show how an organism can exist in two distinctive, heritable states without changes in the nucleotide sequence of its genome.
| Transcriptional circuits are central to the regulation of many biological processes. Often the logic of the circuit, rather than the nature of its components, makes up its most critical feature. In this paper we describe an interlocking network of positive feedback loops that underlies white-opaque switching in the human fungal pathogen Candida albicans. White-opaque switching is an epigenetic phenomenon, where genetically identical cells can exist in two distinctive cell types, white and opaque [1]. Each cell type is stably inherited for many generations, and switching between the two types of cells occurs stochastically and rarely—roughly one switch in 104 cell divisions [2]. The white form is the default cell type, and we propose that the main purpose of the network of interlocking feedback loops is, once excited, to stably maintain the opaque cell type through many cell divisions. Thus, we propose that the feedback loop network driving opaque formation is activated infrequently, but once activated, it is stably propagated through many cell generations.
Despite possessing the same genome, white and opaque cells have many phenotypic differences. Approximately 400 genes are differentially expressed between the two cell types, and the cells differ in their appearance under the microscope and in the color and shape of the colonies they produce on solid media [1,3,4]. They also differ in their behavior toward other C. albicans cells: opaque cells, but not white cells, are highly competent for mating; they respond to mating pheromone with polarized growth, and they can subsequently undergo cell and nuclear fusion with opaque cells of the opposite mating type [5–8]. Finally, the two types of cells appear to interact differently with their mammalian host, with opaque cells appearing more suited for skin infections, and white cells appearing to fare better in blood stream infections [9,10].
Several transcriptional regulators have been identified that play key roles in maintaining the white and opaque cell types, and in controlling the switching between them. Cells of mating type a/α are blocked for white-opaque switching, with all cells remaining locked in the white phase [5]. This block occurs through the action of two homeodomain proteins (a1 and α2), encoded at the mating type-like a (MTLa) and MTLα loci, respectively. The a1 and α2 proteins likely act together as a heterodimer to repress transcription of WOR1, the product of which is a positive regulator of the opaque state [11–13]. Wor1 is required for establishment and maintenance of the opaque state, and ectopically expressed WOR1 drives cells into the opaque form. In a and α cells (both of which are permissive for switching), WOR1 is expressed at low levels in white cells, but in opaque cells Wor1 activates its own synthesis, and WOR1 expression levels rise dramatically. High levels of Wor1, produced by this positive feedback loop, are necessary to maintain the opaque state. Finally Efg1, which has been studied extensively for its role as regulator of the filamentous growth and pathogenesis in a/α (i.e., non-switching) strains of C. albicans, also plays a part in white-opaque switching: in a and α cells (but not a/α cells), cells deleted for EFG1 exist almost exclusively in the opaque state [14,15] (this work). Thus EFG1 is needed to stably maintain the white state.
In this paper we identify two additional transcription regulators of white-opaque switching, CZF1 and white-opaque regulator 2 (WOR2). The former has been previously studied as an important regulator of filamentous growth in a/α (non-switching) cells [16], but a role in white-opaque switching has not been previously described. WOR2 has not been previously described in any context, and we named the gene WOR2 based on its key role white-opaque switching, as described in this paper.
In order to understand the genetic relationships among WOR1, EFG1, CZF1, and WOR2, we constructed a large set of single and double mutants and analyzed them for white-opaque switching. We also ectopically expressed these regulators in mutant strains in various combinations and monitored their effects on switching and maintenance of the white and opaque states. Finally, we carried out chromatin immunoprecipitation (ChIP) experiments to establish direct regulatory connections between the central regulator, Wor1, and the other targets. We found that Wor1 binds upstream of: (1) all four transcriptional regulators investigated in this paper (WOR1, CZF1, WOR2, EFG1); (2) genes whose transcription is regulated by the white-opaque switch; and (3) a large number of genes that are not differentially transcribed during the white-opaque switch, suggesting an additional, previously unrecognized component of white-opaque switching, one that may require additional environmental inputs to fully reveal. Based on the combined results of these experiments, we have placed MTLa1, MTLα2, WOR1, CZF1, WOR2, and EFG1 into a single genetic circuit regulating the white-opaque switch. This circuit is formed from a network of interlocking positive feedback loops, and we believe that this network can account for the stability of the white and opaque states through many cell generations.
One of the most striking characteristics of the white-opaque switch is the large number of genes that are differentially regulated between the two types of cells. Approximately 400 genes have altered transcription; roughly half are up-regulated in the white phase, with the remaining half up-regulated in the opaque phase [3,4]. Several of these regulated genes encode transcriptional regulators, as predicted by the presence of a sequence-specific DNA-binding motif encoded in the gene. We tested a set of opaque-enriched transcription factors for possible roles in regulating the white-opaque switch.
Opaque-enriched genes were previously identified through microarray analyses that compared the gene expression profiles of isogenic white and opaque cells [3,4]. Genes up-regulated in opaque cells were searched for homology to transcriptional regulatory proteins using BLAST searches (http://www.ncbi.nlm.nih.gov/BLAST/) and Pfam motif searches (http://www.sanger.ac.uk/Software/Pfam/). From a set of 237 opaque-enriched genes, we chose to study six genes encoding putative transcriptional regulators: CZF1, WOR2, HAP3, orf19.4972, CSR1, and PHO23. Both CZF1 and WOR2 are predicted to contain a Zn(2)-Cys(6) zinc cluster motif, a known DNA-binding domain in fungal transcriptional regulators [17]; indeed WOR2 had been provisionally named ZCF33 to indicate it was the 33rd protein annotated with this zinc cluster motif. HAP3 is predicted to contain a motif similar to the CCAAT-binding factor. The genes CSR1 and orf19.4972 are predicted to each contain multiple C2H2 zinc fingers, a well-characterized DNA-binding domain. Because chromatin structure has been proposed to play a role in regulating the white-opaque switch [18], we also chose to investigate the opaque-enriched transcript PHO23, which encodes a protein containing a PHD domain and is predicted to be a part of the RPD3 histone deacetylase complex.
For each of these candidate genes, we attempted to make homozygous deletion mutants in a white strain that is mating type a, and is thus permissive for switching to the opaque cell type (C. albicans is diploid, and it is therefore necessary to knock out two copies of each gene). Multiple independent deletion mutants of each target gene were made from independent heterozygous mutants. Despite numerous attempts, we were unable to create a homozygous knockout mutant of the CSR1 gene and did not study CSR1 further. White-opaque switching can occur in strains that are mating type a or α but not a/α. Most of the work presented in this paper was performed in a strains, but we know that Wor1 is also required for opaque formation in α strains (unpublished data and [13]), suggesting that white-opaque switching is controlled the same way in a and α strains. Consistent with this idea, a large set of microarray data indicates that the set of genes regulated by white-opaque switching is virtually identical in a and α cells [3].
For the five remaining candidates, we performed quantitative white-to-opaque switching assays as described previously [5] on at least two independent deletion mutants for each candidate gene, and multiple experiments were performed for each mutant (Table 1). As shown in Table 1, two mutants, the czf1Δ/czf1Δ and wor2Δ/wor2Δ knockouts, had a dramatic effect on white-opaque switching, forming opaque colonies much less frequently than did otherwise isogenic wild type (WT) a strains. The CZF1 deletion strain formed opaque sectors and colonies ∼50-fold less frequently than WT a strains. The wor2Δ/wor2Δ mutant was never observed to form opaque colonies, representing a switching frequency at least 180-fold below that of the parent strain. Due to the key role this gene has in white-opaque switching, as described in this paper, we named the gene WOR2. These results implicate both CZF1 and WOR2 in the white-opaque switch; formally, they function as activators of the opaque state.
To verify that the defects in white-opaque switching were attributable to the disrupted genes, we complemented the czf1Δ/czf1Δ and wor2Δ/wor2Δ deletion mutants. Ectopic expression constructs, controlled by the MET3 promoter were introduced into the RP10 locus, as described previously [19]. Both the czf1Δ/czf1Δ pMET3-CZF1 and wor2Δ/wor2Δ pMET3-WOR2 strains were able to form opaque colonies when the MET3 promoter was induced. However, when an empty vector was introduced, or the strains were grown on media that repressed the MET3 promoter (and thus the only copy of CZF1 or WOR2, respectively), the strains remained white. These results confirmed that the loss of CZF1 or WOR2 drastically reduces the ability for the strains to grow as opaque cells.
The deletion mutants lacking either orf19.4972 or HAP3 formed opaque colonies at frequencies comparable to WT a strains and were not studied further. The pho23Δ/pho23Δ mutant switched to the opaque phase approximately six times as frequently as the WT control. If Pho23 works with Rpd3 in C. albicans, as is predicted based on homology in Saccharomyces cerevisiae, this result is consistent with a previous finding that rpd3Δ/rpd3Δ mutants have an increased frequency of interconversion between the white and opaque phases [18]. We did not study Pho23 further, because of its relatively small affect on switching frequencies, and because these effects could well be indirect: in S. cerevisiae, deletion of RPD3 affects transcription levels of approximately 13% of the genome [20].
We next expressed CZF1 and WOR2 ectopically in white cells to test whether either could drive white cells to the opaque form. All ectopic expression constructs described in this study were controlled by the MET3 promoter integrated at the RP10 locus, as previously described [19]. To test if ectopic expression of a given gene causes white-opaque switching, white strains were streaked from frozen stock onto repressing media and grown at room temperature for 1 wk. The strains were then plated for single colonies onto inducing media or repressing media, as a control, and grown for 1 wk at room temperature. The control a strain, with an empty vector (pCaEXP) inserted into the RP10 locus, switched to the opaque phase at the typical low frequency, producing opaque sectors in approximately 0.5% of the colonies on both media conditions (Table 2), indicating that the media conditions used to control the MET3 promoter do not significantly influence the frequency of white-opaque switching.
We found that ectopic expression of CZF1 in WT a cells led to a mass conversion to the opaque phase (Table 2), but only when the MET3 promoter was induced. In contrast, expression of a pMET3-WOR2 construct did not drive the white-to-opaque switching; the cells remained white, based on colony appearance (Table 2) and cell shape (unpublished data). We know that the pMET3-WOR2 construct is functional from the complementation studies described earlier, thus this result indicates that the ectopic expression of WOR2, at least at the level driven by the MET3 promoter, is not sufficient to drive opaque formation in an otherwise WT a strain.
Previous work on the regulation of white-opaque switching identified WOR1 as a master regulator of the white-opaque switch [11–13]. In the next set of experiments, we tested the genetic interactions between WOR1, CZF1, and WOR2 in order to understand how they work together to regulate the switch. As is the case for CZF1 and WOR2, deletion of WOR1 drastically reduces the frequency of opaque formation. Like CZF1, ectopic expression of WOR1 causes mass conversion to the opaque phase in otherwise WT a cells.
We first expressed WOR1 ectopically in a czf1Δ/czf1Δ or wor2Δ/wor2Δ a-cell strain and observed the effects on white-opaque switching. We found that when WOR1 was ectopically expressed in white czf1Δ/czf1Δ mutants, most of the colonies grew in the opaque phase or had opaque sectors (Table 2), although the colonies had a slightly rougher texture than did conventional opaque colonies on inducing media (unpublished data). Inspection of the cells from the opaque colonies revealed elongated cells, typical of opaque cells (Figure 1). In a control experiment, a czf1Δ/czf1Δ mutant with pCaEXP, an expression vector lacking WOR1, was not converted into opaque cells (Table 2).
When we expressed WOR1 ectopically in a white wor2Δ/wor2Δ a strain, we observed that all of the colonies contained cells that had switched to the opaque phase (Figure 1), usually in the form of opaque sectors, though the opaque sectors were slightly lighter in color than those of normal opaques (Table 2). This strain was never observed to form opaque cells when grown on media that repressed expression of the ectopic WOR1. The wor2Δ/wor2Δ strain with pCaEXP, an empty expression vector, also appeared locked in the white phase, whether it was grown on the repressing or inducing media (Table 2).
Next, we tested the effects of ectopic expression of CZF1 in a wor1Δ/wor1Δ a strain. We found that these strains remained locked in the white phase (Table 2); they were indistinguishable from a wor1Δ/wor1Δ mutant. Finally, we tested the ectopic expression of WOR2 in a wor1Δ/wor1Δ a strain, and we found no change in the switching frequency, as compared to a wor1Δ/wor1Δ mutant (Table 2). This result was expected, given that induction of the WOR2 ectopic construct had no effect in a WT background.
Taken together, these results indicate that CZF1 and WOR2 function upstream of WOR1; thus ectopic expression of WOR1 suffices for opaque cell formation whether or not WOR2 and CZF1 are present. However, the converse is not true: deletion of WOR1 cannot be overcome by ectopic expression of CZF1.
Unlike wor1Δ/wor1Δ and wor2Δ/wor2Δ mutants, czf1Δ/czf1Δ mutants do form opaque colonies, albeit infrequently. As described above, ectopic CZF1 expression can induce a switch to the opaque state. To clarify CZF1's role in white-opaque switching, we examined switching in the reverse direction, where opaque cells switch to white cells. When opaque isolates of WT a strains were replated on repressing media, about 16% of the colonies switched back to the white form (Table 1). The rare opaque isolates of czf1Δ/czf1Δ a strains were nearly as stable as WT opaques; upon replating, 23% of the colonies contained white cells.
Because opaque isolates in czf1Δ/czf1Δ strains are rare, we sought to create more opaque czf1Δ/czf1Δ isolates in order to test the stability of the opaque cells lacking Czf1. To do this, we used the pMET3-WOR1 ectopic expression construct to drive czf1Δ/czf1Δ strains to the opaque state, as described above. When the pMET3-WOR1 construct was subsequently repressed in the czf1Δ/czf1Δ opaque a strains, at least 92% of the colonies remained opaque (Table 3). Similarly, a pulse of pMET3-WOR1 in WT white a cells is sufficient to generate stable opaque populations; the ectopic Wor1 expression can be repressed and the strains will largely continue to grow in the opaque phase (Table 3) [11]. These data indicate that, although its presence is important to form opaque cells, Czf1 contributes minimally to the stability (that is, the heritability) of the opaque state, once it has been established.
In parallel with the studies described above, we examined WOR2's role in maintaining the heritability of the opaque state. As described, when WOR1 was ectopically expressed in wor2Δ/wor2Δ mutants, opaque cells formed. When the pMET3-WOR1 construct was then repressed in these cells (Table 3), the majority of the cells reverted to the white form (Table 3). In contrast, WOR2/WOR2 control strains remained in the opaque form for many generations after the pulse of WOR1 expression. These results indicate that Wor2 contributes greatly to the stability of the opaque state, once it has been formed.
Thus far, we have only considered the role of the opaque-enriched transcription factors WOR1, CZF1, and WOR2 in the regulation of the white-opaque switch. However, a fourth regulator, EFG1, which is up-regulated in white cells, is known to participate in white-opaque switching [14,15]. Experiments reported by Sonneborn et al. [14] suggested that depletion of Efg1 induced the formation of opaque cells in some a/α strain backgrounds, but not in others. To clarify these results, we constructed new isogenic homozygous efg1Δ/efg1Δ mutants in a or a/α strains. In the mating type a strain, we found the efg1Δ/efg1Δ mutation caused a majority of the population to switch to the opaque phase; over 98% of the colonies contained opaque sectors (Table 4), with many colonies showing multiple sectors. We also observed a small number of entirely white colonies, indicating that EFG1 is not strictly necessary for growth in the white phase. We also examined the opaque-to-white switching frequency in efg1Δ/efg1Δ mutants; we found that they switched to the white phase ∼80 times less frequently than WT a strains (Table 4). Thus, deletion of EFG1 dramatically increased the likelihood the cells will grow in the opaque phase, confirming previous studies in WO-1, an α strain [15].
In contrast to the previous reports, we never observed opaque colonies or sectors in the efg1Δ/efg1Δ mutant in an a/α strain, despite observing over 3,200 colonies (unpublished data). We obtained the previously published efg1Δ/PCKpr-EFG1 a/α mutant that showed opaque cell formation when remaining allele of EFG1 was repressed [14]. Using PCR to amplify the MTLa1 and MTLα2 genes, we determined that this mutant was an a strain, likely due to spontaneous loss of one copy of Chromosome 5, which carries the MTL locus (Figure 2). Additionally, each of the efg1Δ/efg1Δ mutants that were locked in the white phase was confirmed to be an a/α strains (Figure 2). These results indicate that the loss of Efg1 from a cells causes massive conversion to the opaque state, but this conversion is blocked in a/α cells. Thus Efg1 functions upstream of the a1-α2 block of white-opaque switching.
To understand the genetic interplay between EFG1, CZF1, and WOR2, we created strains that lacked the white-enriched regulator EFG1 and each of the opaque-enriched regulators, CZF1 or WOR2. These double homozygous knockouts were then tested in quantitative switching assays and monitored for the frequency of forming white, opaque, and sectored colonies. Nearly all colonies of the efg1Δ/efg1Δ czf1Δ/czf1Δ mutant were in the opaque phase or contained opaque sectors (Table 4), reflecting the phenotype of the efg1Δ/efg1Δ mutant. When the opaque colonies isolated from efg1Δ/efg1Δ czf1Δ/czf1Δ were replated to test the heritability of the state, only 0.08% of the colonies returned to the white state, in comparison with the normal opaque-to-white switching frequency, where ∼16% of the colonies are white or have white sectors. Thus, the efg1Δ/efg1Δ czf1Δ/czf1Δ opaque cells are approximately 200-fold more stable than WT opaque cells, a stability similar to that of efg1Δ/efg1Δ a mutants (Table 4). Thus, in both forward and reverse switching frequency, the efg1Δ/efg1Δ czf1Δ/czf1Δ double mutants closely resembled the efg1Δ/efg1Δ single mutant.
We also examined the switching behavior of an efg1Δ/efg1Δ wor2Δ/wor2Δ mutant. In this strain, white colonies accounted for ∼99% of the total colonies seen, reflecting the phenotype of the wor2Δ/wor2Δ mutant. We also tested the stability of these rare opaque colonies that were formed by the efg1Δ/efg1Δ wor2Δ/wor2Δ mutant. When replated, these opaque cells proved to be highly unstable; over 98% of the colonies were white or contained white sectors. Thus, in both forward and reverse switching, the efg1Δ/efg1Δ wor2Δ/wor2Δ double mutant resembled the wor2Δ/wor2Δ mutant.
The genetic epistasis data presented above places WOR1 at the center of white-opaque regulation; formally, it is the most “downstream” regulator of opaque formation, as its deletion blocks white-opaque switching in all contexts tested. Moreover, ectopic WOR1 expression suffices to switch white cells to opaque cells when any of the other opaque-enriched transcription factors are deleted. Previous work indicated that Wor1 expression is maintained through an auto-stimulatory positive feedback loop, mediated by Wor1 binding at its own promoter [11]. Expression of the genes encoding Czf1, Wor2, and Efg1 are all regulated by white-opaque switching; in the opaque form, CZF1 and WOR2 are up-regulated, and EFG1 is down-regulated relative to the white form. Thus, in a formal sense all three are regulated by Wor1.
To test whether Wor1 directly regulates CZF1, WOR2, and EFG1, we performed ChIP using an affinity purified antibody (α-Wor1Nterm), raised against a peptide near the N terminus of Wor1. These ChIPs were analyzed genome-wide using microarrays (ChIP-chip): the precipitated DNA was amplified, fluorescently labeled, and competitively hybridized against genomic DNA (input DNA) on custom DNA tiling microarrays containing 60-mer oligonucleotides tiled at 80 bp intervals across the entire C. albicans genome. Two microarrays were hybridized with DNA from two separate immunoprecipitations (IPs) of an opaque WT a strain; a single ChIP was performed in a wor1Δ/wor1Δ (white) a strain as a control.
We examined the Wor1 ChIP data using previously published software that implements a statistical model and integrates data from several neighboring spots along chromosomes to identify IP enrichment peaks [21]. Using standard parameters, this procedure identified 206 peaks of Wor1 enrichment across the genome in opaque cells. These peaks of Wor1 enrichment were confirmed by visual inspection of ChIP-chip data plotted along chromosomes. Of these peaks, 25 also appeared in the ChIP-chip of a wor1Δ/wor1Δ strain, and likely represent cross reactivity or particularly “sticky” proteins; they were removed from further analysis, leaving a set of 181 peaks of Wor1 enrichment. In parallel, we performed a series of ChIP-chip experiments using a different antibody against Wor1 (raised against a peptide at the C terminus of the Wor1 protein, α-Wor1Cterm) and identified 122 peaks of Wor1 enrichment in opaque strains that were not detected in the wor1Δ/wor1Δ control strain. By comparing the sets of peaks of Wor1 enrichment identified using both antibodies, we found that 112 peaks are enriched for Wor1 in opaque cells (but not wor1Δ/wor1Δ strains) using both antibodies. Thus, 112/122 (92%) of the peaks identified using the α-Wor1Cterm were also found using the N-terminal antibody, indicating the set of peaks identified with α-Wor1Cterm is almost entirely a subset of α-Wor1Nterm ChIP-chip data. We found that 112/181 (62%) of the targets identified in using the α-Wor1Nterm antibody were also detected using the α-Wor1Cterm in opaque cells. Because the experiments using α-Wor1Nterm exhibited very little cross-reactivity in wor1Δ/wor1Δ strains and virtually encompassed the set found using α-Wor1Cterm, we chose the targets identified in the α-Wor1Nterm ChIP-chip as our set of high-confidence Wor1 targets for further analysis.
With 181 peaks identified as high confidence Wor1 targets using the α-Wor1Nterm antibody, we turned to the question of identifying the genes potentially regulated by Wor1. We chose to limit the set to the 170 peaks positioned in intergenic regions upstream of at least one open reading frame (ORF); this eliminated three peaks positioned within ORFs and eight peaks positioned between convergent ORFs. Because some of the 170 peaks of Wor1 enrichment lay in the intergenic region of divergently transcribed genes, there are 221 genes potentially regulated by Wor1 (Table S1).
We found clear Wor1 enrichment at the intergenic regions immediately upstream of the CZF1, WOR2, and EFG1 coding sequences (Figure 3). In the ∼7.6 kb of intergenic sequence upstream of CZF1, we found segments that were enriched up to 20-fold for Wor1, as compared to a wor1Δ/wor1Δ control strain. Upstream of the WOR2 coding sequence, we found segments enriched up to 11-fold. Wor1 was enriched up to ∼12-fold in the 10.1 kb upstream of EFG1. We also verified that Wor1 was found upstream of the WOR1 gene using the tiling arrays, showing enrichment up to ∼80-fold in the opaque cells, as compared to the control wor1Δ/wor1Δ strain (Figure 3) [11]. These results confirm that Wor1 is present at its own promoter in opaque cells, and reveal that Wor1 is also present at the promoters of CZF1, WOR2, and EFG1 in opaque a cells.
We also compared the set of 221 genes potentially regulated directly by Wor1 to the set of genes differentially transcribed between white and opaque cells [3]. We found Wor1 enrichment at the intergenic regions upstream of 38 opaque-enriched genes and 20 white-enriched genes (Table S1), suggesting that Wor1 directly controls expression of approximately 15% of the genes regulated by white-opaque switching. These results also suggest that Wor1 may function in opaque cells as both a transcriptional repressor and as an activator. Though there is some ambiguity in ascribing Wor1 regulation at divergently transcribed genes, we estimate that Wor1 also binds more than 100 genes that have not been previously identified as white- or opaque-enriched by transcriptional profiling.
As described above, Wor1 protein is bound at the intergenic DNA upstream of the genes WOR1, EFG1, CZF1, and WOR2. Each of these genes has a remarkably long upstream region of DNA (at least 7 kb in each case), and Wor1 appears to be bound at multiple positions along these regions. From the ChIP-chip experiments, we found that occupancy of large intergenic regions is a general characteristic of Wor1. Analysis of the intergenic regions at all 6,077 gene promoters in the C. albicans genome (excluding those at telomeres) revealed a median promoter length of 623 bp, whereas the median promoter length of the 181 gene promoters bound by Wor1 was 3,390 bp (unpublished data). This preference is especially pronounced when considering the intergenic regions over 10 kb; Wor1 enrichment was seen at 12 of 19 of these intergenic regions. Intriguingly, over half (seven of 12) of these intergenic regions lie upstream of genes encoding sequence-specific DNA binding proteins—WOR1, RFG1, TCC1, WOR2, ZCF37, EFG1, and RME1—suggesting Wor1 may exert much of its control over the white-opaque switch indirectly through other transcriptional regulators.
In this paper, we have dissected the genetic circuitry controlling white-opaque switching in the fungal pathogen C. albicans. White-opaque switching is an epigenetic change between two distinct types of cells, both containing the same genome. The white-opaque switch is crucial for many aspects of C. albicans biology, including interactions with other C. albicans cells (pheromone sensing and mating) and interactions with the host (opportunistic pathogenesis). Our results are summarized in Figures 4 and 5, where the circuitry controlling this switch is diagrammed. This network of positive feedback loops is responsible for the heritability of each state, as well as the frequency of switching between them, and we propose that the structure of this network makes an important contribution to the biology of white-opaque switching.
The default state can be considered the white cell type: most clinical isolates of C. albicans are a/α cells, and they are locked in the white state through a1-α2 repression of WOR1 (Figure 5A). However, even in a and α cells, which are permissive for white-opaque switching, the white cell type still seems to be the default, in that white cells are generally more stable than opaque cells (Figure 5B). For example, opaque cells at 24 °C are stable for many generations, but above 30 °C they become unstable and rapidly switch back en masse to the white form, which is stable under these conditions [1,2]. There are no known environmental conditions that comparably destabilize the white form. In our model, the opaque form is generated when the series of positive feedback loops shown in Figure 4 become excited (Figure 5C). Thus, in opaque cells, WOR1 likely directly induces CZF1 and WOR2 expression, and in turn, CZF1 and WOR2 both activate WOR1. CZF1 does this by repressing a repressor of the opaque state (EFG1), the net effect being a positive feedback loop.
The multiple feedback loops observed in the opaque state are reminiscent of those seen in differentiated animal cells, such as those of the Drosophila eye ([22], reviewed in [23]) and the mammalian myoblast ([24], reviewed in [25]). A series of such feedback loops (as opposed to a single loop) buffers the circuit against transient fluctuations in any single regulatory protein and therefore provides additional stability to the excited form of the circuit. In addition, the nature of the circuit probably defines the switching frequency. For example, deletion of CZF1 decreases the white-to-opaque switching frequency by approximately 10-fold, but has little effect on the backward switching rate. Thus, the primary role of CZF1 seems to be in modulating the switching frequency; in contrast, WOR1 and WOR2 are both required to maintain the opaque state; thus their roles are more integral to the switch itself.
Although the overall logic of the circuit shown in Figure 4 can explain many features of white-opaque switching, there appear to be several unusual features of the circuit components themselves that likely also play important roles in white-opaque switching. For example, our ChIP-chip experiments revealed that Wor1 binding shows a bias toward genes with unusually long upstream intergenic regions—as defined by the distance from the 5′ end of the ORF to the next annotated coding region. This observation suggests that these genes bound by Wor1, which include the four encoding transcriptional regulators that form the interlocking feedback loops (WOR1, EFG1, CZF1, and WOR2) are also controlled by a number of other transcriptional regulators. It is known that the frequency of white-opaque switching can be influenced by environmental cues (e.g., temperature and oxidative stress) [1,26], and it seems plausible that different rates of switching could be “set” by individually adjusting the levels of the regulatory proteins that make up the circuit. For example, since deletion of CZF1 reduces the frequency of white-to-opaque switching 10-fold, regulation of the level of Czf1 by environmental signals could directly control the “forward” switching rate. Another unusual feature of the circuit concerns the wide distribution of Wor1 over much of the upstream regions of WOR1, EFG1, CZF1, and WOR2 (Figure 3), suggesting a highly cooperative transcriptional response to the intracellular levels of WOR1. This, combined with the interlocking positive feedback loops, could be responsible for the switch-like behavior of the system, specifically the failure to readily observe a cell type intermediate between white and opaque in wild-type switching strains.
The switch from the white to the opaque form growth alters transcription of approximately 400 genes. We know that the master regulator Wor1 ultimately controls all of these genes, since deletion of WOR1 locks cells in the white form, and ectopic expression of WOR1 converts white cells en masse to opaque cells. Our ChIP-chip analysis revealed that Wor1 directly regulates approximately 15% of this gene set (20 white-enriched genes and 38 opaque-enriched genes). Since Wor1 is also bound upstream of CZF1, EFG1, WOR2, and 20 additional transcriptional regulators (see Table S1), it seems likely that much of white-opaque switching program is regulated indirectly by Wor1 through its effects on other transcriptional regulators.
An unexpected outcome of the Wor1 ChIP-chip experiments was the presence of Wor1 at a large number of genes that were not identified as white- or opaque-enriched in previous microarray analyses [3]. There are several explanations for this observation. First, Wor1 could control these genes in both white and opaque cells, with their transcription being unaffected by the white-to-opaque transition. We think this explanation is unlikely because Wor1 is up-regulated 45-fold in opaque cells [3], and it seems improbable that this change could have no impact on expression of target genes. To test this idea directly, we performed a Wor1 ChIP-chip experiment in white a cells, and found that Wor1 is not bound at any of these target genes (unpublished data). A second possibility, one that we favor, is that Wor1 may occupy the promoters of these 100 genes in opaque cells, preparing their expression to respond to unknown environmental signals, perhaps those generated by the host. According to this idea, the standard laboratory conditions used for transcriptional profiling would not have included the necessary environmental stimuli, and thus these genes would not have been identified as regulated by the white-opaque switch. This idea suggests there are additional aspects to white-opaque switching which have not been previously recognized.
Finally we note that white-opaque switching does not appear to be a general feature of fungi, even those that are closely related to C. albicans. Indeed it may have arisen during the long association of C. albicans with its warm-blooded hosts. The evolution of a complex circuit composed of interlocking feedback loops is relatively simple to imagine, as it could occur stepwise simply through the acquisitions of cis-acting sequences in genes for transcriptional regulators used for other purposes in the cell. We note that Czf1 also relieves Efg1-mediated repression of hyphal growth under embedded conditions [27], and this genetic relationship has been maintained in the regulation of the white-opaque switch. Thus EFG1 and CZF1 have other key functions in the cell—even in cells that are genetically blocked for white-opaque switching—and their involvement in white-opaque switching could well be a recent adaptation, functioning to modulate the stability of the two states and the frequency of switching between them. The independent evolution of interlocking transcriptional feedback loops in a variety of distinct biological contexts (white-opaque switching in C. albicans, eye development in flies, and muscle development in mammals, for example) suggests they are particularly effective ways of providing, from the same genome, distinctive cell types that can be stably propagated for many generations.
All strains and primers used in this study are listed in Tables S2 and S3, respectively. DNA sequences of C. albicans genes were obtained from the Candida Genome Database (http://www.candidagenome.org/)
Standard laboratory media have been described previously [28]. Synthetic complete media, supplemented with 2% glucose and 100 μg/ml uridine (SCD+Urd) was used to maintain strains in the white and opaque phases at room temperature. For ectopic expression experiments, cells were grown on inducing media (SCD−Met−Cys+Urd) or repressing media (SCD+Met+Cys+Urd) to control expression of the MET3 promoter, as described previously [11,19].
The plasmid containing the pMET3-WOR1 construct (pRZ25) has been described before [11]. To make the pMET3-WOR2 and pMET3-CZF1 constructs, the WOR2 or CZF1 ORFs was PCR-amplified from SC5314 genomic DNA using primers containing BamHI and SphI restriction sites, and cloned into a BamHI/SphI-digested pCaEXP, to create the plasmids pAJ2230 and pAJ2231, respectively.
All strains were derived from SC5314. EFG1, CZF1, or WOR2 was deleted using a modified Ura-blaster protocol [29]. In short, the recyclable URA3-dpl200 marker was PCR-amplified from pDDB57 using long oligonucleotides identical to the sequence immediately flanking each ORF targeted for deletion. The deletion construct was transformed into CHY439 (α1Δα2Δ, Ura−) or CAI4 (a/α, Ura−) and transformants were selected on SD−Ura media. 5-Fluoro-orotic acid was used to counterselect against URA3 marker, and the resulting Ura− isolates were used for subsequent rounds of gene deletion or to create the ectopic expression strains. For each knockout target, at least two homozygous deletion mutants were created from independent heterozygous mutants. When creating double mutants, CZF1 and WOR2 were each deleted in an efg1Δ/efg1Δ (α1Δα2Δ, Ura−) mutant. In the case of the efg1Δ/efg1Δ wor2Δ/wor2Δ mutant, two independent double mutants were created from two independent efg1Δ/efg1Δ homozygous deletion mutants. Each WOR1 allele was deleted from the strain SNY78 (a/α, His−, Leu−, Ura−) using fusion knockout constructs described previously [11]. The resulting strain was grown on sorbose-containing media to generate a/a strains (see [6] and references therein), creating the wor1Δ/wor1Δ (a/a, Ura−) strain.
Ectopic expression constructs pAJ2230, pAJ2231, pRZ25 (described above), or pCaEXP (empty control vector [19]) were linearized to direct integration to the RP10 locus and transformed into Ura− isolates of WT, wor2Δ/wor2Δ, czf1Δ/czf1Δ, or wor1Δ/wor1Δ strains. To create the duplicate ectopic expression strains listed in Table S2, ectopic expression constructs were introduced into independent wor1Δ/wor1Δ or wor2Δ/wor2Δ strains. The czf1Δ/czf1Δ (Ura−) strains used to create the czf1Δ/czf1Δ + pMET3-WOR1 ectopic expression strains are different Ura− loopout isolates generated by 5-fluro-orotic acid counter-selection of the same czf1Δ/czf1Δ (Ura+) strain. The czf1Δ/czf1Δ + pMET3-CZF1 complementation strains were made from the same czf1Δ/czf1Δ knockout strain.
Switching frequencies between the white and opaque phases were determined in plate-based assays, as described previously, with modifications [5]. Strains were streaked from frozen stock onto SCD+Urd and grown at RT for 5–7 d. For each strain, at least five entirely white colonies were resuspended into dH2O, diluted, and plated for single colonies on SCD+Urd. After growth at RT for 1 wk, we examined the colonies and counted the number of switch events (as evidenced by the presence of opaque sectors, or entirely opaque colonies). The same process was used to assess opaque-to-white switching, but the original frozen stocks contained opaque isolates of each strain, and we monitored switching by the presence of white sectors or entirely white colonies. The data shown in Tables 1 and 4 were taken from the same representative experiment and only tested one strain of each genotype. In repetitions of the switching assays (unpublished data), multiple independent deletion mutants of each genotype were tested and yielded results similar to those shown in Tables 1 and 4.
Switching assays in strains containing the pMET3 ectopic expression constructs were performed as described in [11], with modifications. In short, to test if ectopic expression can drive opaque formation, white strains were streaked from frozen stock onto repressing media at RT for 5 d. At least five fully white colonies were replated for single colonies on inducing media (or repressing, as control). After growth at RT for 1 wk, colony phenotypes were recorded. Colonies were resuspended in sterile water and cells were examined by using differential interference contrast microscopy on an Axiovert 200M microscope (Carl Zeiss, http://www.zeiss.de/). All experimental strains, excepting the wor1Δ/wor1Δ + pMET3-WOR2 strains, were tested in at least two repetitions of the switching assay. Data shown in Table 2 are from a single representative experiment, and each strain listed is an independent ectopic expression mutant.
To test if the resulting colony phenotypes were stable after the ectopic expression was repressed, opaque strains (formed by induction of the ectopic expression constructs) were streaked from frozen stock onto inducing media at room temperature. At least five opaque colonies were resuspended in sterile dH2O and replated onto repressing media (or inducing, as control) and grown at RT for 1 wk. Colony phenotypes were recorded. Two independently derived strains were tested for each ectopic expression scenario, and experiments were performed at least twice. Data shown in Table 3 are from a single representative experiment, and each strain listed is an independent ectopic expression mutant.
To determine the mating type of C. albicans strains, we PCR amplified the a and α alleles of each gene located within the MTL locus (PAP, OBP, PIK, MTLa1, MTLa2, MTLα1, MTLα2) [30]. PCR products for every a allele were seen in all strains tested; products for each α allele were seen in all strains except SS4 (unpublished data). PCR products for the MTLa1 and MTLα2 genes are shown in Figure 2.
Overnight cultures (200 ml) were grown in SCD+Urd for approximately 16 h at 25 °C to an OD600 of 0.4. Cells were formaldehyde cross-linked by adding formaldehyde (37%) to a 1% final concentration. Treated cultures were mixed by shaking and incubated for 15 min at room temperature. 2.5 M glycine was added to a final concentration of 125 mM, and treated cultures were mixed and incubated 5 min at room temperature. Cells were pelleted at 3,000 g for 5 min at 4 °C and washed twice with 100 ml of 4 °C TBS (20 mM TrisHCl [pH 7.6], 150 mM NaCl).
Spheroplasting and ChIP were carried out as previously described, with modifications [11,31]. Cell pellets were resuspended in 39 ml of Buffer Z (1 M sorbitol, 50 mM Tris-Cl [pH 7.4]), 28 μl of β-ME was added (14.3 M, final concentration 10 mM), and cells were vortexed. 20 μl of lyticase (Sigma, MO, United States) solution (2 mg/ml in Buffer Z) was added, and cell suspensions were incubated 15 min at 30 °C. Spheroplasted cells were then spun at 3,000 g, for 10 min, at 4 °C and resuspended in 500 μl of 4 °C lysis buffer (50 mM HEPES-KOH [pH 7.5],140 mM NaCl, 1 mM EDTA, 1% Triton X-100, 0.1% sodium deoxycholate) with protease inhibitors. All subsequent ChIP and wash steps were done at 4 °C. DNA was sheared by sonication ten times for 10 s at power setting 2 on a Branson 450 sonicator (http://www.bransonultrasonics.com/), incubating on ice for 2 min between sonication pulses. Extracts were clarified by centrifugation. A 50 μl aliquot of each extract was set aside as ChIP input material.
For the IP, 450 μl of lysis buffer was added to 50 μl of extract, and 5 μl of α-Wor1Nterm antibody was added. α-Wor1Nterm is an affinity-purified antibody generated against a peptide, QVLDKQLEPVSRRPHERER, located near the N terminus of Wor1 (Bethyl Laboratories, http://www.bethyl.com/). The IP was incubated for 2 h at 4 °C, with agitation. Then 50 μl of a 50% suspension of protein A-Sepharose Fast-Flow beads (Sigma, http://www.sigmaaldrich.com/) in lysis buffer was added to the IP and incubated 1.5 h at 4 °C with agitation. The beads were pelleted for 1 min at 3,000 g. After removal of the supernatant, the beads were washed with a series of buffers for 5 min for each wash: twice in lysis buffer, twice in high-salt lysis buffer (50 mM HEPES-KOH [pH 7.5], 500 mM NaCl, 1 mM EDTA, 1% Triton X-100, 0.1% sodium deoxycholate), twice in wash buffer (10 mM Tris-HCl [pH 8.0], 250 mM LiCl, 0.5% NP-40, 0.5% sodium deoxycholate, 1mM EDTA), and once in TE (10 mM Tris, 1 mM EDTA [pH 8.0]). After the last wash, 100 μl of elution buffer (50 mM Tris-HCl [pH 8.0], 10 mM EDTA, 1% SDS) was added to each sample, and the beads were incubated at 65 °C for 15 min. The beads were spun for 1 min at 10,000 g, and the supernatant was removed and retained. A second elution was carried out with 150 μl of elution buffer 2 (TE, 0.67% SDS) and eluates from the two elution steps were combined. For the ChIP input material set aside, SDS (1% final concentration) and 200 μl of TE were added. ChIP and input samples were incubated overnight at 65 °C to reverse the formaldehyde crosslinks. 250 μl of proteinase K solution (TE, 20 μg/ml glycogen, 400 μg/ml proteinase K) was added to each sample, and samples were incubated at 37 °C for 2 h. Samples were extracted once with 450 μl Tris buffer-saturated phenol/chloroform/isoamyl alcohol solution (25:24:1). 55 μl of 4 M LiCl and 1 ml of 100% ethanol (4 °C) were added and the DNA was precipitated for 1 h at 4 °C. The DNA was pelleted by centrifugation at 14,000 g for 15 min at 4 °C, washed once with cold 75% ethanol, and allowed to air dry. The samples were resuspended in 25 μl of TE containing 100 μg/ml RNaseA and incubated 1 h at 37 °C.
ChIPs were also carried out in experiments not shown using affinity-purified antibody generated against a peptide DDAVGNSSGSYYTGT, located at the C terminus of Wor1 (α-Wor1Cterm) (Bethyl Laboratories) [11]. ChIP was performed in WT opaque strains twice using α-Wor1Nterm and three times using the α-Wor1Cterm antibodies. Control ChIPs were performed in the wor1Δ/wor1Δ mutants using α-Wor1Nterm once, and the α-Wor1Cterm was used twice.
ChIP-enriched DNA was amplified and fluorescence labeled as described [32]. Labeled DNA for each channel was combined and hybridized to arrays in Agilent hybridization chambers for 40 h at 65 °C, according to protocols supplied by Agilent (Agilent Technologies, http://www.agilent.com/). Arrays were then washed and scanned, using an Axon Instruments Genepix 4000A scanner.
Approximately 185,000 60-mer oligo probes were designed across the entire Candida genome (contig20 haploid genome assembly) at approximately 80 bp intervals, excluding nonunique regions of the genome (see Protocol S1 for further information). Custom microarrays were manufactured by Agilent Technologies (Agilent Technologies). Array design and ChIP-chip data are available on GEO.
Arrays were blank subtraction normalized, interarray median normalized, and intra-array median normalized using Agilent ChIP Analytics 1.3 software (Agilent Technologies). After normalization, average ratios across replicate arrays (where relevant) were used for further analysis. After normalization, the single array error model was applied across replicate arrays (where relevant), to derive a p-value statistic to represent the probabilities that data at each spot occurred within experimental noise. A segment is a region of adjacent probes containing peaks of Wor1 enrichment, where the enrichment above input is considered to be statistically significant, based on the parameters set in the software. Using the ChIP Analytics software, the Whitehead Neighborhood Model was applied using default parameters as described [21] to map the segments according to their chromosomal positions. When comparing ChIP-chip experiments in WT opaque strains against wor1Δ/wor1Δ strains, or between α-Wor1 ChIP-chip experiments performed in WT opaque strains using the two different Wor1 antibodies, any overlapping segments were eliminated from further analysis.
Within each segment, we used ChIP Analytics software to identify the location of highest Wor1 enrichment (corresponding to the probe with the lowest P[ξ]-value). The positions of peaks were then assessed in relationship to ORFs throughout the C. albicans genome; an ORF was identified as being potentially regulated by Wor1 if there was a segment of Wor1 enrichment within the intergenic region immediately upstream of the given coding sequence.
Candida Genome Database (http://www.candidagenome.org/) accession numbers for the genes discussed in this article are CZF1 (orf19.3127), WOR2 (orf19.5992), HAP3 (orf19.4647), orf19.4972, CSR1 (orf19.3794), and PHO23 (orf19.1759).
|
10.1371/journal.pgen.1002464 | RIC-7 Promotes Neuropeptide Secretion | Secretion of neurotransmitters and neuropeptides is mediated by exocytosis of distinct secretory organelles, synaptic vesicles (SVs) and dense core vesicles (DCVs) respectively. Relatively little is known about factors that differentially regulate SV and DCV secretion. Here we identify a novel protein RIC-7 that is required for neuropeptide secretion in Caenorhabditis elegans. The RIC-7 protein is expressed in all neurons and is localized to presynaptic terminals. Imaging, electrophysiology, and behavioral analysis of ric-7 mutants indicates that acetylcholine release occurs normally, while neuropeptide release is significantly decreased. These results suggest that RIC-7 promotes DCV–mediated secretion.
| Neuropeptides produce prolonged changes in circuit activity that are associated with changes in behavioral states (e.g. mood or appetite); consequently, there is great interest in identifying molecules that are required for neuropeptide secretion. Here we show that a novel neuronal protein RIC-7 promotes neuropeptide secretion in C. elegans but has only subtle effects on neurotransmitter secretion. RIC-7 is conserved in several other nematodes; however, homologous proteins are not found in other sequenced genomes. These results suggest that the machinery responsible for neuropeptide secretion evolved more recently than factors that are required for both neurotransmitter and neuropeptide secretion.
| Neurons secrete both neuropeptides and neurotransmitters. Neurotransmitters, such as acetylcholine (ACh), are secreted by exocytosis of small clear synaptic vesicles (SVs) whereas neuropeptide secretion is mediated by exocytosis of dense core vesicles (DCVs) [1], [2]. The mechanisms leading to DCV and SV exocytosis are similar in many respects. SVs and DCVs both undergo physical docking to the plasma membrane, requiring Munc18 and syntaxin for docking in both cases [3]–[7]. To become fusion competent, SVs and DCVs must both undergo a priming reaction, which is mediated by priming factors (e.g. Munc13 and CAPS) [8], [9]. Exocytosis of SVs and DCVs are both mediated by assembling complexes between vesicular and plasma membrane SNARE proteins [10], [11]. Finally, calcium-evoked fusion of SVs and DCVs are mediated by distinct calcium sensors, which are thought to be different synaptotagmin isoforms [12].
Beyond these similarities, DCVs and SVs exhibit many important differences. DCVs can be found all along the cell body, dendrites and axons of neurons whereas SVs cluster specifically at active zones of synapses [13]. SVs undergo repeated cycles of exo- and endocytosis at synapses, whereas neuropeptides are only packaged into nascent DCVs in the Golgi [14]. Consequently, DCVs cannot undergo local recycling in axons or dendrites. DCVs release their contents over long timescales (>50 ms) while SV exocytosis occurs more rapidly (<20 ms) [13], [15]. Exocytosis of SVs can be evoked by single action potentials while DCV release typically occurs after more prolonged or repeated depolarizations. These differences imply that different molecules are involved in SV and DCV secretion.
To date, very few proteins have been found that are specifically involved in the secretion of one or the other class of vesicles. UNC-31/CAPS (Calcium-dependent Activator Protein for Secretion) is proposed to promote priming of DCVs but not SVs [16]–[18]. However, a subsequent study showed compelling evidence for SV priming defects in CAPS1 and CAPS2 double knockout mice [19], implying that CAPS is also required for SV priming. Similarly, some studies propose that Munc13 primes SVs but not DCVs [18], while others find Munc13 mutants have exocytosis defects for both SVs and DCVs [20], [21]. C. elegans mutants lacking PKC-1, a PKCε ortholog, had significant defects in DCV release but little effect on SV release [20]. Identifying new genes that differentially regulate SV or DCV release will provide new insights into the mechanisms underlying these two forms of secretion.
In C. elegans, the acetylcholinesterase inhibitor aldicarb has been widely used to study neuromuscular signaling in live animals. Aldicarb treatment causes acetylcholine (ACh) to accumulate in the synaptic cleft at NMJs, resulting in an acute paralysis of treated animals. Mutations or RNAi treatments that reduce ACh release confer resistance to aldicarb-induced paralysis, whereas those that stimulate ACh secretion enhance aldicarb sensitivity [22], . We previously showed that neuropeptides also regulate aldicarb responsiveness [24], [25]. Inactivation of genes encoding proneuropeptide processing enzymes [for example, egl-3 prohormone convertase (PC2), egl-21 carboxypeptidase E (CPE), sbt-1 7B2, and nep-1 neprilysin], proneuropeptides (ins-22, ins-31, flp-1, nlp-12), neuropeptide receptors (fshr-1) all cause aldicarb resistance [24], [25]. These results suggested that new genes that are required for DCV secretion could be identified through screens for aldicarb resistant mutants.
In a screen for mutations that suppress the aldicarb hypersensitivity of dgk-1 diacylglycerol kinase (DAGK) mutants, we isolated a new allele of the ric-7 gene. Here we show that ric-7 encodes a novel nematode specific protein that is required for neuropeptide secretion.
To identify new genes required for neuromuscular function, we screened for mutations that suppress the aldicarb hypersensitivity defect of dgk-1 DAGK mutants. One suppressor (nu447) significantly decreased the aldicarb hypersensitivity of dgk-1 mutants. The nu447 mutation mapped close to the ric-7 gene, which was identified in prior screens for aldicarb resistant mutants [22]. We found that nu447 and ric-7(n2657) both correspond to mutations in F58E10.1 gene (Figure 1A), hereafter referred to as the ric-7 gene. The ric-7 locus encodes two isoforms (A and B) that differ only in their first exon. Orthologs of ric-7 are observed in several other nematodes, but homologous genes are not detected in other metazoans. The predicted RIC-7 protein does not contain any previously described structural domains.
Animals homozygous for nu447 or n2657 were resistant to the paralytic effects of aldicarb (Figure 1B). To determine whether RIC-7 functions in motor neurons for aldicarb responsiveness, we constructed a ric-7 transcriptional reporter. The resulting construct expressed GFP in many neurons, including both cholinergic and GABAergic motor neurons (Figure 2A–2B). Transgenes expressing either RIC-7A or B isoforms in all neurons (with the snb-1 promoter), and those expressing RIC-7B in cholinergic neurons (with the unc-17 promoter) rescued the aldicarb sensitivity defect of ric-7 mutants (Figure 1B). By contrast, expressing RIC-7B in GABAergic neurons (unc-25 promoter) or in muscles (myo-3 promoter) did not rescue the aldicarb phenotype of ric-7 mutants (Figure 1B). These results suggest that RIC-7 activity is required in cholinergic neurons for aldicarb responsiveness.
Animals lacking ric-7 also had decreased locomotion rates (Figure 1C). This locomotion defect was fully rescued by ric-7 transgenes expressed in all neurons (using the snb-1 promoter) whereas partial rescue was observed with transgenes expressed in cholinergic or GABAergic neurons (Figure 1C). The morphology of motor neuron axons and NMJs appeared superficially normal in ric-7 mutants (Figure 4C and data not shown), suggesting that these motor defects were unlikely to be caused by changes in neural development.
We expressed GFP-tagged RIC-7 constructs in the cholinergic DA neurons (using the unc-129 promoter). The RIC-7::GFP protein was localized in a punctate distribution in both cell bodies and dorsal cord axons. The majority of RIC-7 puncta co-localized with an SV marker (mCherry-tagged Endophilin) [26], suggesting that RIC-7 is targeted to synapses (Figure 2C). We also compared the distribution of GFP-tagged RIC-7 with a mCherry-tagged neuropeptide, NLP-21. Whereas RIC-7 showed partial overlap with endophilin, nearly complete co-localization was observed with NLP-21 (Figure 2C). Several results suggest that the co-localization of RIC-7 with DCVs was not mediated by physical association of RIC-7 with nascent DCVs. First, RIC-7 synaptic localization was not disrupted in unc-104 mutants (Figure 2D), which lack the KIF1A motor responsible for anterograde transport of SVs and DCVs. Thus, RIC-7 is not co-transported to synapses with immature DCVs or SVs. Second, RIC-7 lacks a predicted signal peptide sequence and GFP-tagged RIC-7 did not produce fluorescence in coelomocytes, both of which indicate that RIC-7 is not translocated into SVs or DCVs (data not shown). These results support the idea that RIC-7 functions in the cytoplasm at presynapses and could play a relatively direct role in regulating SV or DCV secretion.
We did several experiments to test the effects of RIC-7 on the responsiveness of body muscles to neuromuscular agonists (Figure 3). First, the sensitivity of ric-7 mutants to the paralytic effects of the nicotinic agonist levamisole was similar to that of wild type controls (Figure 3A). Second, we recorded body wall muscle currents evoked by application of ACh or the GABA agonist muscimol. In both cases, the amplitude of agonist-evoked current in ric-7 mutant body muscles was not significantly different from that observed in wild type controls (Figure 3B). Third, the fluorescent intensities of GFP-tagged ACR-16 ACh receptor (Figure 3C) and UNC-49 GABAA receptor (Figure 3D) puncta in the nerve cord were unaltered in ric-7 mutants, indicating that the abundance of post-synaptic receptors at NMJs was normal. Thus, the effect of RIC-7 on aldicarb responsiveness is unlikely to be caused by altered agonist responsiveness of body muscles.
The aldicarb resistance phenotype observed in ric-7 mutants could be caused by decreased ACh secretion, increased GABA secretion, or decreased neuropeptide secretion, as aldicarb resistance would be expected in all three scenarios [25], [27], [28]. To assay ACh secretion more directly, we recorded excitatory post-synaptic currents (EPSCs) from body muscles (Figure 4). We found that the rate and amplitude of endogenous EPSCs, i.e. SV fusions evoked by the endogenous activity of motor neurons, in ric-7 mutants were similar to those found in wild type animals (Figure 4A). In addition, the amplitude of stimulus-evoked EPSCs in ric-7 mutants was not significantly different from wild type (Figure 4B). Therefore, baseline ACh secretion occurs normally at cholinergic NMJs in ric-7 mutants.
To further analyze the cholinergic NMJs, we examined the distribution of GFP-tagged Synaptobrevin (GFP::SNB-1) in motor neurons. Changes in the distribution of GFP::SNB-1 are correlated with changes in SV exo- and endocytosis. SNB-1 puncta intensity is correlated with the number of SVs at presynaptic elements [24], [29] whereas puncta density is a measure of synapse density. Neither the fluorescent intensity nor the density of SNB-1 puncta in the cholinergic axons of ric-7 mutants were significantly different from that observed in wild type controls (Figure 4C). Taken together, these results suggest that defects in baseline ACh secretion are unlikely to account for the aldicarb resistance of ric-7 mutants.
Aldicarb resistance could also be caused by defects in neuropeptide signaling [25]; therefore, we next examined ric-7 mutants for changes in neuropeptide signaling. First, we analyzed the aldicarb responsiveness of ric-7 double mutants containing mutations in neuropeptide signaling components. If changes in neuropeptide action contribute to RIC-7's effects on aldicarb responsiveness, then we would expect that ric-7 mutations would occlude the effect of neuropeptide signaling mutations on aldicarb resistance. Consistent with this idea, ric-7 double mutants carrying mutations in either of two proneuropeptide processing enzymes (egl-21 CPE and egl-3 PC2) had aldicarb resistance that was similar to that observed in ric-7 single mutants (Figure 5). These results support the idea that RIC-7 regulation of neuropeptide secretion contributes to the effects of ric-7 mutations on aldicarb responsiveness.
To further address the role of RIC-7 in neuropeptide secretion, we analyzed the secretion of YFP-tagged neuropeptides (Figure 6). For this analysis, we selected two proneuropeptides, NLP-21 and INS-22, which encode FMRFamide related peptides (FaRPs) and an insulin-like growth factor, respectively. When NLP-21::YFP or INS-22::YFP are expressed in the cholinergic DA motor neurons (using the unc-129 promoter), puntate fluorescence is detected in dorsal cord axons and in coelomocytes (Figure 6A and 6C). We previously showed that the axonal puncta fluorescence corresponds to secretory granules containing these proneuropeptides while the coelomocyte fluorescence corresponds to secreted neuropeptides that have been endocytosed [20], [30]. In ric-7 mutants, NLP-21 coelomocyte fluorescence was significantly decreased (∼50%, p<0.001) (Figure 6C–6D) whereas the NLP-21 puncta fluorescence intensity in dorsal cord axons was significantly increased (>2-fold, p<0.001) (Figure 6A–6B). The coelomocyte and axonal NLP-21 fluorescence defects were both rescued by ric-7 transgenes expressed in the DA neurons. Similar changes in axonal puncta fluorescence and coelomocyte fluorescence was observed for a second proneuropeptide (INS-22) in ric-7 mutants (Figure 6). These results suggest that ric-7 mutants have decreased neuropeptide secretion, which results in an accumulation of DCVs in motor axons.
Inactivation of NLP-21 and INS-22 by RNAi does not produce strong locomotion or aldicarb resistance defects [24]; consequently, NLP-21 and INS-22 are unlikely to be the only neuropeptides involved in the Ric-7 locomotion and aldicarb-resistance phenotypes. We recently identified NLP-12 as a neuropeptide that plays a critical role in regulating both locomotion rate and aldicarb-induced paralysis [31]. NLP-12 is expressed in a proprioceptive neuron (DVA) that is activated by body muscle contractions [32], [33]. If Ric-7 aldicarb resistance and locomotion defects are caused by decreased neuropeptide release, we would expect that NLP-12 secretion from DVA would be diminished in ric-7 mutants. We did several experiments to test this idea. First, we analyzed the effect of ric-7 mutations on secretion of YFP-tagged NLP-12 from DVA neurons (Figure 7A–7B). In wild type animals, aldicarb treatment significantly decreased NLP-12 puncta fluorescence in DVA axons (indicating increased NLP-12 secretion), which is most likely caused by activation of DVA stretch receptors by muscle contraction [31]. By contrast, in ric-7 mutants, aldicarb had no effect on NLP-12 puncta fluorescence, indicating that RIC-7 is required for aldicarb-evoked NLP-12 secretion.
Second, if the NLP-12 secretion defect contributes to the Ric-7 aldicarb resistance phenotype, we would expect that RIC-7 expression in DVA neurons would be sufficient to alter aldicarb responsiveness. Consistent with this idea, the aldicarb responsiveness of ric-7 mutants was significantly improved by transgenes expressing RIC-7 in DVA neurons (Figure 7C). This result is consistent with the preceding rescue data (Figure 1B) because the unc-17 promoter (which also rescued the Ric-7 aldicarb defect) is expressed in DVA neurons (data not shown). Collectively, these data suggest that proper aldicarb responsiveness requires RIC-7 function in multiple neuron classes because RIC-7 expression in a single cholinergic neuron (DVA) produced partial rescue of the aldicarb defect whereas complete rescue was obtained following expression in all cholinergic neurons (with the unc-17 promoter) (Figure 1B).
Third, we analyzed evoked ACh release following aldicarb treatment. In wild type animals, a 60 minute pre-treatment with aldicarb significantly increases the total synaptic charge occurring during an evoked response (Figure 7D–7E) [31]. This effect is eliminated in egl-3 PC2 mutants and in nlp-12 mutants [31]. Thus, aldicarb enhancement of evoked ACh release can be utilized to assess changes in endogenous NLP-12 secretion. Aldicarb's effect on evoked ACh release was also eliminated in ric-7 mutants (Figure 7D–7E). Collectively, these results strongly support the idea that RIC-7 acts in DVA neurons to promote secretion of endogenous NLP-12, and that this contributes to the Ric-7 aldicarb-resistance defect.
Changes in GABA secretion could also contribute to the aldicarb resistance observed in ric-7 mutants [22], [27], [28]. Consistent with this idea, ric-7 mutants had defects in defecation behavior that are similar to those observed in mutants with decreased GABA transmission (Figure 8A). Contraction of the intestinal muscles during defecation (i.e. the expulsion step of the defecation motor program) is mediated by excitatory GABAergic input from defecation motor neurons [34]. We found that ric-7 mutants had a significant expulsion defect, which was rescued by ric-7 transgenes expressed in GABA neurons (Figure 8A). These results suggest that ric-7 mutants have a presynaptic defect at GABAergic NMJs involved in defecation.
To assay GABA release at ventral cord NMJs (which are involved in locomotion), we recorded inhibitory post-synaptic currents (IPSCs) from body muscles (Figure 8B). We found that ric-7 mutants had a wild type IPSC rate, while IPSC amplitudes were significantly increased (∼30%, p<0.001). These results suggest that GABA secretion still occurs in ric-7 mutants, albeit in a subtly altered form. Changes in IPSC amplitudes are often caused by changes in post-synaptic sensitivity, e.g. by changing GABA receptor abundance. However, neither the current evoked by applying an exogenous GABA agonist nor the abundance of GFP-tagged UNC-49 GABA-A receptors were altered in ric-7 mutants, suggesting that a post-synaptic defect was unlikely to account for the altered IPSC amplitude (Figure 3B, 3D). Consistent with a pre-synaptic defect, the distribution of GFP-tagged SNB-1 was altered in GABAergic neurons of ric-7 mutants. Although SNB-1 puncta intensity was unaltered, puncta were significantly wider (23% wider, p<0.001) and diffuse SNB-1 axon fluorescence was significantly increased in ric-7 mutants (Figure 8C). Finally, the IPSC and SNB-1 defects were both rescued by transgenes expressing RIC-7 in GABAergic motor neurons (Figure 8B–8C). Collectively, these results support the idea that ric-7 mutants have a pre-synaptic defect that increases GABA secretion, perhaps by increasing the amount of GABA packaged into each SV. Rescue of the IPSC defect (with ric-7 transgenes expressed in GABA neurons) failed to rescue the ric-7 aldicarb defect and weakly rescued the locomotion defect implying that the latter were not caused by altered GABA secretion (Figure 1C).
Mutations disrupting neuropeptide signaling exhibit defecation defects similar to ric-7 mutants [25], [35]. Prompted by these results, we wondered if the Ric-7 defecation and IPSC defects are caused by the neuropeptide secretion defect. In this scenario, we would expect that ric-7 mutations and mutations that impair proneuropeptide processing would not have additive effects on defecation behavior in double mutants. Contrary to this idea, the defecation defects observed in ric-7; egl-21 double mutants were significantly worse than those observed in either single mutant (Figure 8A). In addition, IPSC amplitudes were unaltered in egl-3; egl-21 doubles mutants (Figure S1), suggesting that decreased neuropeptide secretion is also unlikely to explain the Ric-7 IPSC defect. These results indicate that RIC-7 cell-autonomously regulates two forms of secretion, promoting neuropeptide secretion and inhibiting GABA secretion.
Communication between neurons and their targets relies on two classes of neurosecretory vesicles, synaptic vesicles (SVs) and dense-core vesicles (DCVs). The mechanisms governing SV and DCV secretion share many properties (including related SNARE proteins, SNARE binding proteins, and calcium sensors). Despite these similarities, SV and DCV mediated secretion also have significant differences [1], which imply that some proteins will be selectively utilized in one or the other process. Because neuropeptides have dramatic effects on personality, behavior, and metabolism, there is significant interest in identifying molecules that selectively promote DCV secretion. Here we identify RIC-7 as a novel protein that is required for DCV secretion but has relatively subtle effects on SV secretion. Below we discuss the implications of these findings.
The ric-7 gene was identified in a screen for mutations conferring resistance to aldicarb induced paralysis. As RIC-7 lacks identified structural or functional motifs, the mechanisms underlying this behavioral defect were unclear. In principle, aldicarb resistance could arise from several alternative mechanisms, including: altered muscle responsiveness to ACh or GABA, decreased excitatory input from cholinergic motor neurons, increased hyperpolarizing input from GABA motor neurons, or decreased neuropeptide signaling. Our results strongly support the idea that Ric-7 aldicarb and locomotion defects arise from disruption of neuropeptide secretion.
Several results suggest that RIC-7 does not regulate muscle sensitivity to neuromuscular agonists. The currents evoked by applying exogenous ACh and GABA to body muscles, and the expression of nicotinic and GABA receptors in body muscles were both unaffected in ric-7 mutants. Furthermore, ric-7 aldicarb and locomotion defects were not corrected by transgenes expressing RIC-7 in body muscles, whereas rescue was observed for transgenes expressed in cholinergic neurons. Thus, RIC-7 neither acts in body muscles, nor regulates muscle sensitivity to agonists.
Other results indicate that RIC-7 is not required for baseline ACh secretion. In cholinergic motor neurons, the SV protein SNB-1 neither accumulated in synaptic puncta nor in the axonal membrane (as would occur in exocytosis and endocytosis mutants, respectively). The rate, amplitude, and kinetics of endogenous and evoked EPSCs were unaltered in ric-7 mutants. Thus, although RIC-7 expression in cholinergic neurons rescued the aldicarb and locomotion defects of ric-7 mutants, these defects are unlikely to arise from decreased baseline ACh secretion.
In GABA neurons, ric-7 mutants had a presynaptic defect that led to an apparent increase in GABA secretion. This presynaptic defect was manifest by an increase in amplitude of endogenous IPSCs in ric-7 mutants, while the IPSC rate was unaltered. Changes in IPSC amplitudes are often caused by changes in post-synaptic sensitivity, e.g. by changing GABA receptor abundance. However, neither the current evoked by applying an exogenous GABA agonist nor the abundance of GFP-tagged UNC-49 GABA-A receptors were altered in ric-7 mutants, suggesting that a post-synaptic defect was unlikely to account for the altered IPSC amplitude. Consistent with a pre-synaptic defect, SNB-1 axonal fluorescence was significantly increased in ric-7 mutant GABA motor neurons. The defecation, IPSC, and SNB-1 defects were all rescued by transgenes expressing RIC-7 in GABAergic motor neurons. Collectively, these results support the idea that ric-7 mutants have a pre-synaptic defect that increases GABA secretion. Rescue of the IPSC defect did not correlate with rescue of ric-7 aldicarb and locomotion defects, implying that the latter were not caused by altered GABA secretion. In principle, the IPSC defect could cause the Ric-7 defecation defect, for example if increased GABA secretion causes constitutive excitation of the intestinal muscles. Double mutant analysis suggests that decreased neuropeptide secretion is unlikely to account for the Ric-7 IPSC and defecation defects. Consequently, our results are most consistent with the idea that RIC-7 has direct cell autonomous effects on two forms of secretion, promoting neuropeptide secretion and inhibiting GABA secretion. Several other proteins are known to have distinct effects on different neurotransmitter systems. Mouse Munc13-1 knockouts drastically reduce glutamatergic transmission yet have little effect on GABA release [36]. Mouse ELKS2 knockouts have increased GABA release but unaltered glutamate release [37]. Finally, complexin promotes evoked release but inhibits tonic/spontaneous release [38], [39].
Several results suggest that the effects of RIC-7 on aldicarb sensitivity were caused by changes in neuropeptide secretion. First, mutations preventing neuropeptide processing (egl-21 CPE, egl-3 PC2) and ric-7 mutations did not have additive effects on aldicarb responsiveness in double mutants. Second, ric-7 mutants had greatly decreased secretion of YFP-tagged neuropeptides (NLP-21 and INS-22) from cholinergic motor neurons. Third, mCherry-tagged RIC-7 co-localized extensively with a DCV marker (NLP-21), implying that RIC-7 could play a relatively direct role in regulating neuropeptide release. Fourth, ric-7 mutants have decreased aldicarb-evoked secretion of NLP-12 from DVA neurons and lack aldicarb-induced potentiation of evoked ACh release (which is mediated by endogenous NLP-12). And fifth, restoring RIC-7 expression in DVA neurons was sufficient to partially rescue the Ric-7 aldicarb-resistance defect. Taken together, these results strongly support the idea that RIC-7 promotes secretion of endogenous neuropeptides, and that this function plays an important role in the Ric-7 aldicarb resistance and locomotion defects. The Ric-7 aldicarb resistance and locomotion defects are more severe than those observed in mutants lacking NLP-12 or those lacking EGL-3 PC2, indicating that additional RIC-7 functions also contribute to these phenotypes. These additional functions could include promoting secretion of other neuropeptides or novel RIC-7 functions that are not yet defined.
What aspect of DCV secretion is regulated by RIC-7? Decreased neuropeptide secretion in ric-7 mutants could reflect changes in any aspect of DCV biogenesis, transport, docking, priming, calcium-triggering, or fusion. Several results suggest that RIC-7 is not packaged into DCVs, and consequently cannot play a role in pro-neuropeptide processing. RIC-7 lacks a predicted signal peptide sequence, GFP-tagged RIC-7 is not secreted, and RIC-7 delivery to axons is not prevented in unc-104 KIF1A mutants. Collectively, these results suggest RIC-7 is cytoplasmic, and thus cannot play a direct role in neuropeptide processing. DCV biogenesis and transport also occur normally in ric-7 mutants, as neuropeptide fluorescence in motor axons was increased rather than decreased. These results suggest that RIC-7 regulates a step that occurs after DCV transport. Given the prominent colocalization of RIC-7 and DCV markers in axons, we speculate the RIC-7 could identify DCV release sites. Because the sequence of RIC-7 does not provide any clues as to its biochemical function, further experiments will be required to determine a more precise function for RIC-7. Understanding the mechanisms underlying RIC-7's functions will undoubtedly shed light on how DCVs assume their unique properties.
The mechanisms governing SV and DCV secretion share many properties, including: SNARE proteins, SNARE binding proteins (e.g. Munc13, Munc18, and CAPS), and calcium sensors (e.g. Synaptotagmin) [40]. These shared mechanisms are highly conserved across eukaryotic phylogeny, suggesting that these core exocytosis components comprise an ancient process. By contrast, RIC-7 orthologs are observed in other nematodes but homologous genes are not detected in other sequenced genomes. Interestingly, other putative DCV secretion factors have similar patterns of conservation. For example, the Rab27 effector granuphilin is conserved in mammals and flies but not in C. elegans, while a second Rab27 effector melanophilin is present in mammals but absent in flies and worms. These results suggest that the mechanisms distinguishing SV and DCV secretion evolved more recently than the more ancient shared secretion factors.
Strains were maintained at 20°C as described [41]. The wild-type reference strain was N2 Bristol. Descriptions of allele lesions can be found at http://www.wormbase.org. The mutant strains used were: LGIV, egl-21(n476); LGV, egl-3(nr2090), ric-7 (nu447), ric-7(n2657). Acute aldicarb and levamisole assays were performed blind in triplicate on young adult worms as described [42]. The aldicarb (Chem Services) concentrations used ranged from 1 to 2 mM, the levamisole (Sigma) concentration was 200 µM. Locomotion was measured by transferring adults to plates containing fresh (16 h) lawns of HB101 bacteria, letting the worms recover for 30 min and counting the number of body bends of active worms in a 3 min interval.
ric-7(nu447) was isolated in an EMS screen for suppressors of the aldicarb hypersensitive phenotype of dgk-1(nu62) mutants (D.S. and J.K., unpublished data). nu447 was mapped by small nucleotide polymorphism (SNP) mapping, based on its aldicarb resistance phenotype. Analysis of 16 nu447/CB4856 recombinants mapped nu447 to LGV, map unit 5∼6; Another 62 clones further positioned nu447 to 5.86∼6.00 m.u. Seven predicted genes were found in this interval with three of them previously characterized. Sequencing revealed that ric-7(nu447) contained a 830C/T (A277V) point mutation plus a 32 bp deletion (nucleotides 910–941) in F58E10.1b cDNA that shifts the reading frame, leading to a truncated RIC-7 protein that contains 305 amino acids. ric-7(n2657) contained a 621G/A (W207Stop) nonsense mutation in F58E10.1b cDNA. nu447 was backcrossed six times and used for most phenotypic analysis in this study.
All quantitative imaging was done using a Olympus PlanAPO 100×1.4 NA objective and an ORCA100 CCD camera (Hamamatsu). Worms were immobilized with 30 mg/ml BDM (Sigma). Image stacks were captured and maximum intensity projections were obtained using Metamorph 7.1 software (Molecular Devices). GFP fluorescence was normalized to the absolute mean fluorescence of 0.5 mm FluoSphere beads (Molecular Probes). For dorsal cord imaging, young adult worms, in which the dorsal cords were oriented toward the objective, were imaged in the region midway between the posterior gonad bend at the tail. Line scans of dorsal cord fluorescence were analyzed in Igor Pro (WaveMetrics) using custom-written software [43], [44]. For coelomocyte imaging, the posterior coelomocyte was imaged in young adults [20]. Image stacks were captured and maximum intensity projections were obtained using Metamorph 7.1 software (Molecular Devices). For quantitation, the five brightest vesicles were analyzed for each coelomocyte and the mean fluorescence for each vesicle was logged. For each worm, coelomocyte fluorescence was calculated as the mean of the vesicle values in that animal. All p-values indicated were based on student t-tests.
Confocal images were taken using the Olympus FV1000 confocal microscope. Image stacks were captured, and maximum intensity projections were obtained using Metamorph 7.1 software (Universal Imaging).
Electrophysiology was done on dissected adults as previously described [45]. All recording conditions were as described previously [20]. For comparing average electrophysiological values, statistical significance was determined using the Mann-Whitney test or student's t test.
Transgenic strains were generated by injecting either wild type or ric-7(nu447) mutants with the expression construct (10–25 ng/µl) mixed with the co-injection markers, KP#1338 (pttx-3::GFP), KP#1480 (pmyo-2::NLS-mCherry) or KP#1106 (pmyo-2::NLS-GFP), each at 10 ng/µl, using standard methods [46]. The co-localization experiments were done by co-injecting KP#1684 (pric-7::his-24 cDNA::wcherry) with KP#1685 (punc-30::NLS-GFP) or with KP#1686 (punc-17::NLS-GFP), respectively; or by injecting KP#1687 (punc-129::RIC-7 cDNA::GFP) into nuIs252 (Punc-129::mchry::PaGFP::unc-57), or into nuIs470 (punc-129::NLP-21::mCherry), respectively, all at a concentration of 25 ng/µl each. All constructs except KP#1684 were derivatives of pPD49.26 [47].
Two RIC-7 constructs rescued the ric-7(nu447) mutant aldicarb defect: KP#1680 (psnb-1::RIC-7), and KP#1681 (punc-17::RIC-7). Two RIC-7 constructs did not rescue the ric-7(nu447) mutant aldicarb defect: KP #1682(punc-25::RIC-7), and KP#1683(pmyo-3::RIC-7). All four constructs used F58E10.1b cDNA cloned with NheI and KpnI sites.
The previously reported lines used for imaging experiments described are nuIs152 (punc-129::GFP::SNB-1), nuIs183 (punc-129::NLP-21::VENUS), nuIs195 (punc-129::INS-22::VENUS) [20], [24], nuIs283 (pmyo-3::UNC-49::GFP) (J. Bai and J.K., unpublished), nuIs299 (pmyo-3::ACR-16::GFP) [48], nuIs444 (pnlp-12::NLP-12::VENUS) [31] and nuIs376 (punc-25::SNB-1::GFP).
|
10.1371/journal.pntd.0006167 | Emerging trends of Zika apprehension in an epidemic setting | French Guiana is a territory that has a decades-long history of dengue outbreaks and more recently, in 2014, a chikungunya outbreak. Zika virus (ZIKV) emerged in late 2015 and subsequently led to an important outbreak.
A cross-sectional phone survey was conducted among the general population during the outbreak in June 2016 with a total of 1,129 individuals interviewed to assess perceptions, knowledge and behaviors regarding zika infection. The population seemed aware of zika, and perceived the infection as a more serious health threat than other common mosquito-borne diseases. Furthermore, both the perceptions and behaviors related to zika and its prevention were found to vary considerably among different social groups, geographic areas and gender; less educated female participants were found to perceive the disease as more worrisome and were less likely to adopt protective behaviors. Moreover, female population has been particularly responsive to awareness campaigns and rapidly understood the extent of risks associated with ZIKV infection.
These results revealed that ZIKV appeared at the time of the survey as a new health threat that concerns the public more than chikungunya and dengue fever with differences observed among subgroups of population. These results have implications for the development of multifaceted infection control programs, including strategies for prevention and awareness, helping the population to develop an accurate perception of the threat they are facing and encouraging behavior changes.
| Although dengue fever has been a focus of many awareness campaigns in Latin America, very little information is available about beliefs, attitudes and behaviors regarding vector-borne diseases among the population of French Guiana. Following the end of the first chikungunya outbreak and at the initial onset of the first zika outbreak, a quantitative survey was conducted among 1129 individuals aiming to study the emotional, cognitive and behavioral response to the risk of zika infection and assess variations among different groups of population. People from French Guiana were found to perceive zika substantially differently from other Aedes mosquito-borne diseases. Overall, ZIKV appeared at the time of the survey as a new health threat that makes the population more scared than chikungunya and dengue fever. Furthermore, both the beliefs and behaviors related to zika and its prevention were found to vary considerably among different social groups, gender and geographic areas. Education had an impact on perceptions and behaviors among women. Female population has been particularly responsive to awareness campaigns and rapidly understood the extent of risks associated with ZIKV infection. Overall, findings emphasize the importance of developing appropriate and relevant strategies helping population to engage in protective behaviors adapted to the health threat they are facing. Given the importance of the public response and precautionary actions to control the spread of an emergent threat, additional research on risk perceptions and other behavioral determinants is warranted.
| Zika virus (ZIKV) is a Flavivirus transmitted by mosquitoes, primarily Aedes aegypti, the same vector that transmits dengue, chikungunya and yellow fever. During the next half century, ZIKV was considered as an emergent virus with few sporadic and imported cases reported in Africa and Asia until 2007, when a major epidemic occurred in Micronesia (Yap)[1]. A second important epidemic occurred in 2013 in the Pacific, affecting French Polynesia [2].
Since its introduction in the Americas in early 2015 in Brazil, ZIKV has rapidly spread through the continent [3,4]. Although ZIKV infections have not historically been considered as a significant public health concern, during this recent emergence, the virus has been linked to neurological disorders and severe congenital abnormalities. Several recent studies have also highlighted that ZIKV can be transmitted through sexual contact or from mother to fetus [5–7]. Today, it became the first major infectious disease linked to human birth defects, putting at risk pregnant women and women of childbearing age [8–10]. To support national governments and communities in preventing and managing the complications of ZIKV, the WHO launched a global strategic response framework for zika in February 2016. One of its objectives was prevention of health risks associated with ZIKV infection through integrated vector management (IVM) [11]. However, human behavior plays a considerable role in the success or failure of prevention programs and policies implemented by the public health authorities to control the spread of the disease. Since the WHO report about the “behavioral and social aspects of malaria and its control,” human behaviors–and their sociocultural determinants–are increasingly recognized as critical factors contributing to risk for and prevention of mosquito-borne diseases. Furthermore, it has been well-established in the literature that health-protective behaviors are strongly guided by a range of mental representations that individuals and communities develop about health threats over time [12–16]. This construct refers to personal conceptions, beliefs, schemata or imagery about an illness and its diverse consequences. To date, the mental representations of mosquito-borne diseases, and particularly ZIKV infection, have received little attention from the global scientific community. Therefore, the objective of this study was to examine public perceptions associated with this new health threat, with the purpose of informing ongoing intervention practices. More specifically, we wanted to determine (1) whether ZIKV infection is perceived differently from other common mosquito-borne diseases, (2) whether a feeling of emerging infectious diseases “fatigue” can be observed in the French Guiana populations, and (3) whether there existed significant variations in the perception of this emerging health threat among population subgroups, particularly between men and women given the unusual transmission routes and potentially dramatic consequences of ZIKV infection.
French Guiana, an overseas region and department of France located in the Amazonian forest complex, is composed of two main inhabited geographical regions: a central, urbanized area including a coastal strip along the Atlantic Ocean, where a large part of the population lives, and a more remote area along the Surinamese and Brazilian frontiers called the “interior area” (Fig 1).
In French Guiana, Ae. aegypti has been responsible for several major dengue fever outbreaks over the past few decades and a recent outbreak of chikungunya [17,18]. The emergence of ZIKV has been considered of particular concern because the territory has the highest fertility rate in the Americas (3.5 children per woman), with an infant mortality rate (1.2%) that is three times higher than in metropolitan France (0.4%) [19]. On the 22nd of January in 2016, local health authorities launched an official epidemic alert following the rapid spread of ZIKV in the most inhabited part of the territory [20]. The public health response to the outbreak of ZIKV disease included distribution of information about the importance of mosquito bite prevention and the use of physician services for pregnant women with potential symptoms, as well as recommendations by health authorities to delay pregnancy. A massive media campaign was conducted during the outbreak including television and radio spots, prospectus, posters (displayed on the roads and in medical centers) and also newspapers. Messages were about how to prevent mosquito bites or with all zika symptoms listed and there was also specific prospectus for pregnant women. Moreover, all pregnant women were invited to be carefully monitored with a blood sample collected at each trimester of pregnancy then analyzed at the Arbovirus Reference Center at Pasteur Institute of French Guiana [21].
A cross-sectional phone survey about “beliefs, attitudes and practices” among the general population of French Guiana was conducted from June 15–30 in 2016. Eligibility criteria included (i) having a landline or mobile phone, (ii) being at least 18 years old and (iii) indicating consent for participation.
The sample design was based on a random 2-stage selection procedure, stratified according to the type of phone (50% mobile and 50% land line), municipalities and age. Households were randomly selected at the first stage then one participant was randomly selected among adults living in the selected household. The sample calculation included investigation of 1,100 individuals. The survey was conducted by Ipsos, a French consulting firm that used the CATI system (Computer Assisted Telephone Interviews) and CONVERSO software. To reach the largest possible portion of the population, interviews were carried out from Monday to Friday between 10 am and 8 pm and on Saturday between 10 am and 5 pm.
According to the French legislation, the survey protocol and processing of data collection was subject to a declaration to CNIL, the French National Agency responsible for ethical issues and protection of individual data collection under no. 2043940. All the information was collected anonymously, and the participants were informed of their rights to access and rectify personal information.
Data were recorded using a standardized questionnaire administered by phone.
Aside from socio-demographic variables, collected data were grouped into four general categories: 1) environmental variables and exposure to mosquito, 2) perceptions of the illness and risk of contracting ZIKV infection, 3) perceptions and practices of protective behaviors promoted by the public health authorities to control the spread of the disease and 4) self-reported frequency of protective behavior in response to the zika epidemic.
The questionnaire covered of a wide range of topics such as the type of housing, the presence or absence of potential breeding sites, and potential factors associated with breeding sites. Respondents were also asked how frequently they were bitten by mosquitoes (response options: ‘Never’, ‘Seldom’, ‘Sometimes’ or ‘Often’). Participants were also asked whether they were educated about “Aedes mosquitoes”, how frequently they practiced outdoor activities, and during which time of the day mosquito bites occurred. Participants were asked to report the occurrence of an acute febrile illness consistent with presumptive ZIKV infection during the outbreak, and if they had dengue and/or chikungunya virus infection. If the answer was “YES”, we asked them if they had consulted a doctor.
A broad range of personal beliefs was investigated, especially regarding perceptions of the health threat, i.e., qualitative judgments (based on closed-end questions using unordered response options), and quantitative judgments (based on questions using a Likert response scale with a numerical value ranging from 0 to 10) that individuals expressed when asked to evaluate a specific illness and the risk of contracting it [22]. To characterize these perceptions within the population, questions were drawn from the existing methodological literature using the Brief Illness Perception Questionnaire (B-IPQ) [23]. This questionnaire measures six components: the identity- the symptoms that patients associate with the illness; the cause- the personal ideas about etiology; the timeline- the perceived duration of the illness; the consequences- the expected effects and outcome; and the treatment control- the effectiveness of treatment methods for recovery from the illness; and the perceived coherence- whether people think the threat is easy to understand. Complementary questions were adapted from the risk perception literature devoted to transmissible infectious diseases. Worry, perceived severity, perceived exposure, and perceived susceptibility to the disease were also assessed to explore the multiple components of risk perception [24]. Except for the perceived cause and identity of the disease, respondents were asked to use an 11-point Likert response scale ranging from 0 to 10 to characterize their mental and emotional representations of the health threat associated with a variety of vector-borne diseases.
In line with the health beliefs model [25], participants were asked how often they practiced health-protective behaviors as they relate to various vector control methods (e.g., wearing long-sleeved clothing or using repellent) and vector control measures (e.g., covering water receptacles). Participants were then asked whether those behavioral recommendations were effective in preventing mosquito bites (response options: ‘Very ineffective’, ‘Somewhat ineffective’, ‘Somewhat effective’, ‘Very effective’ and ‘Unsure’). To obtain a mean protection score for each respondent, a level of protection score was created as the sum of each mean used.
Statistical analyses were performed using STATA12 software (Stata Corp., College Station, TX, USA) [26] and SPAD8 [27]. Primary sampling units and strata were considered for calculating estimations according to the design effect. Post-stratification weights were used to correct potential biases due to misrepresentation of demographic characteristics, including gender and education level. All estimations were obtained using STATA12 “svy” commands.
Bivariate analyses were conducted using Chi-square tests to compare proportions, and linear combinations tests and regression combined with Wald tests were used to compare mean scores of risk perceptions. The level of statistical significance was set to (p = 0.05). Data were adjusted according to gender and level of education.
In addition, a multiple correspondence analysis and a hierarchical cluster analysis were performed to determine the natural groupings of observations regarding the level of knowledge about zika disease and its issues in order to cluster the population in different groups according to their literacy.
South America and French Guiana layers were drawn using data from OpenStreetMaps (http://www.openstreetmap.org) and mapping operations have been done using QGIS 2.18 software [28].
There was a total of 1,129 participants to the survey. The mean age of participants was 45.6 years old ranging from 18 to 96 years. The original sample showed over-representation of women (64.8% vs 50% in the general population of French Guiana) and well-educated people (38% vs 28.2% in French Guiana). All the results presented later in this article were restated to reflect the post-stratification adjustment.
Over half of the participants reported having knowledge of Aedes mosquitoes (59.1%, 95% Confidence Interval (CI): 54.8–63.3), and 87.9% (95% CI: 84.7–90.5) properly identified zika as a vector-borne disease. The most commonly mentioned symptoms associated with zika disease were fever and myalgia (91%, 95% CI: 87.9–93.4 and 60%, 95% CI: 55.7–64.2, respectively), followed by headache and arthralgia (49.2%, 95% CI: 44.9–53.4 and 44.8%, 95% CI: 40.6–49, respectively). Symptoms mentioned the least frequently were conjunctivitis (8.5%, 95% CI: 6.4–11) and neurologic disorders (1.6%, 95% CI: 0.9–2.8). Most participants were aware of zika transmission issues: 79.3% (95% CI: 75.6–82.6) knew that ZIKV can be transmitted from mother to child, 55.7% (95% CI: 51.5–59.9) knew that ZIKV can be sexually transmitted. Finally, 54.4% (95% CI: 50.1–58.6) declared to be well-informed about zika disease. However, 17.3% (95% CI: 14.-20.9) reported that there was a vaccine against zika. The most popular sources of information about the disease were television, radio and posters (85%, 95% CI: 81.4–87.9, 66%, 95% CI: 61.9–70 and 63.1%, 95% CI: 58.8–67.2, respectively). The proportions of respondents claiming to have been previously infected with zika, chikungunya or dengue were of 8.8% (95% CI: 6.8–11.5), 14.3% (95% CI: 11.4–17) and 20.7% (95% CI: 17.7–24.1), respectively. Among those who reported sudden onset of high fever during the zika outbreak (10.5%, 95% CI: 88–17), 83.8% claimed to have seen a doctor. The WHO recommended a number of IVM actions for individuals to prevent mosquito bites and thus the spread of the virus, thereby reducing their own personal risk. Among respondents who declared being bitten by mosquitoes either often or every day, 40.5% (95% CI: 36.4–44.6) and 66.5% (95% CI: 60.1–72.3), respectively reported taking preventive measures. When asked about the effectiveness of several preventative measures, the most commonly reported were emptying of water from receptacles, use of bed nets, and covering storage containers, whereas the less effective measures were extensive insecticide spraying and the use of a vaporizer for outdoor insecticides. The most frequently reported preventive measures were closing the door and reducing outdoor activities (47.7%, 95% CI: 43.4–52) and using window nets (32.43%, 95% CI: 28.7–36.4). Overall, almost 20% (95% CI: 15.7–22.1) of the population took at least 5 actions to prevent mosquito bites.
As shown in Fig 2, zika received a significantly higher mean score than did dengue and chikungunya as a response to questions about worry, perceived severity, perceived consequences and perceived exposure (p<0.01). Regarding zika treatment, its perceived efficiency was significantly lower than dengue and chikungunya treatments (p<0.01). Moreover, zika appeared to be less understood than chikungunya disease (p<0.01). Nonetheless, ZIKV infection was perceived as more easily avoidable than dengue (p = <0.01).
The distribution of risk perceptions and behaviors among several subpopulations are presented in Table 1. Analysis revealed an association between age, perceived worry, level of understanding, and behaviors, both with a gradient in responses. Respondents between 18 and 25 years old were more worried about zika (p = 0.016), more likely to adopt protective measures (p<0.01), and they claimed to understand the disease less compared to claims of older age groups (p = 0.047). Individuals living in the interior were less worried about zika than those living in the coastal area (p<0.01) and were more likely to adopt protective measures (p = 0.019).
Our results show that previous infection with dengue, chikungunya or zika virus was associated with different perceptions of zika disease. Specifically, individuals reporting previous infection judged the disease less severely (p = 0.05), worried less (p = 0.01), felt more exposed (p = 0.017), had lower estimates for control (p<0.01) and treatment efficiency (p<0.01), and felt more vulnerable in the context of outbreak (p = 0.049).
Finally, respondents with a high level of education characterized the disease as being less severe (p<0.001), as affecting patients to a lesser extent (p<0.001), and as being less avoidable (p = 0.026); highly educated participants also better understood the disease (p<0.01) and judged the treatment as less efficient than respondents with a lower level of education (p<0.001).
As shown in Table 1, women were more afraid of zika than were men. Women were significantly more worried about zika (p<0.001), felt more exposed (p<0.001), and characterized the disease as more severe (p<0.001) and as affecting the patient more than did men (p<0.001). Although scores were higher for women in terms of understanding of the disease and control and treatment efficacy, these differences were not statistically significant. Interestingly, women were worried about disease independently of the level of knowledge, whereas among men, there was an association between knowledge and risk perception (p = 0.03); 65% of men with strong knowledge of the disease were worried vs. 50% among those who were not knowledgeable.
When analyzed according to specific subgroups, women presented nuanced responses with regard to the level of education (Table 2). The impact of education on perceptions was higher among women than men. Low-educated women worried more about the disease (p = 0.02) and were less knowledgeable (p<0.001). Moreover, the level of education was also associated with the adoption of protective behaviors (p<0.01).
Even if the most popular sources of information (television, radio and posters) were reported by the major part of respondents, some variations were observed. Participants aged between 18 and 25 years were more likely to access to social network (p = 0.01) and highly educated participants were more likely to access poster (p = 0.03). Participants who had heard about zika from television (p = 0.010), radio (p<0.001), poster campaign (p<0.001), social network (p = 0.035) and from leaflets (p<0.001) reported to be significantly more informed about zika disease. However, only participants having been aware through leaflets (p = 0.023) were more knowledgeable about zika disease and its issues than those who did not.
Men and women were equally split on the issue of leaving French Guiana in case of pregnancy during a zika outbreak, with 52% and 49% claiming they should have left the territory. Both men and women seemed to be compliant with the recommendations of the WHO, with 85% claiming that they would use a condom during sexual activity. However, approximately 65% of women claimed they would postpone a pregnancy in case of zika outbreak, whereas only 51% of men claimed they would postpone pregnancy. Women indicated that they would worry about undergoing pregnancy during a zika outbreak (75%).
Since its emergence in the Americas in the late 2015, the zika epidemic received considerable attention from the medical, political and lay communities worldwide. During the year 2016, prior to the recent slowdown in transmission of the disease in Latin America, the increasing magnitude of ZIKV outbreaks as well as the severity of their consequences on human health raised the specter of a new and potentially major public health crisis, comparable to those caused by other dramatic re-emerging infectious diseases such as Ebola or SARS [29,30]. Intense public concern was expressed about enhanced transmission of ZIKV which was associated with extensive media coverage, strong institutional attention and high perceptions of health risk [31]. Indeed, since the pioneering works conducted by Slovic & al. in the field of risk psychology [32], it has been well established that diseases caused by invisible, unfamiliar, communicable, potentially highly pathogenic, emerging infectious agents which remains (to a large extent) beyond individual and institutional control are likely to trigger public panic [33].
At both the societal and individual level, one common strategy to cope with unknown and uncontrollable infectious diseases is to relate them to past or existing threats to population health [34]. This social cognitive process, referred to as an “anchoring” effect in social and psychological sciences, consists of actively selecting a small number of familiar events, schemata, images or symbols to describe a new phenomenon. In an epidemic setting, the propensity to use existing or past diseases to understand a new epidemiological phenomenon enables people to make unfamiliar illnesses more familiar, and therefore less threatening [35]. However, it has also been shown that the utilization of some anchors rather than others in the social representation of a new health threat can lead to more or less adaptive responses to the epidemiological event, which may have serious consequences for population health [14,36].
Despite the growing acknowledgment in the scientific community that the cognitive representations held by individuals and groups play a major role in the control and prevention of emerging infectious diseases, little attention has been given to public perceptions of ZIKV infection. To the best of our knowledge, this study is the first to investigate the beliefs, attitudes and practices related to ZIKV among the general population in an epidemic setting. The survey was conducted during the emergence of ZIKV in French Guiana, an endemic region for dengue fever, and nearly one year after a major outbreak of chikungunya, another re-emerging mosquito-borne disease. At the time of the study, the surveillance system estimated a zika prevalence of approximately 3% with 7,830 estimated clinical cases, whereas the study revealed a prevalence of 9%, suggesting an important proportion of clinical cases not covered by monitoring systems.
Our empirical data show that the cognitive representations of ZIKV infection characterizing the French Guiana population were partly anchored on those of other Aedes mosquito-borne diseases, specifically chikungunya. Approximately 90% of the respondents were aware that mosquitoes transmit ZIKV, and a large proportion were aware of the specific transmission issues related to the virus, particularly women. For instance, almost 80% of participants claimed that the virus could be transmitted from mother to fetus. This awareness is likely attributable to the massive media coverage of transmission patterns, as well as institutional campaigns about zika control and prevention. However, we note that despite intense media coverage of this issue, only slightly more than a half of the participants knew that ZIKV can be sexually transmitted. In the same vein, only 8.5% and 1.6% correctly identified conjunctivitis and neurologic disorders as common symptom of ZIKV infection. The poor recognition of typical symptoms associated with zika in our study population suggest some anchoring effects, through which clinical manifestations of this disease were confounded with those of other common mosquito-borne diseases such as dengue fever and chikungunya.
Nevertheless, people from French Guiana were found to perceive zika substantially differently from other Aedes mosquito-borne diseases. Specifically, the study shows that zika was associated with higher mean scores in terms of worry, perceived severity, and consequences of infection. Overall, ZIKV appeared at the time of the survey as a new health threat that concerns the public more than chikungunya and dengue fever. Therefore, these results do not give empirical support to the emerging infectious diseases fatigue hypothesis that was developed in the aftermath of the 2009 H1N1 influenza epidemic [34]. According to this hypothesis, the relative apathy of the Western populations regarding health threats related to emerging infectious diseases can be attributed to the increased frequency of alarmist discourse, which is likely to cause some habituation effects. In other words, the cognitive and affective responsiveness of the public to a re-emerging infectious disease does not seem to be impacted by the repetition of warnings, as long as this new health threat is associated with a significant degree of perceived uncertainty about transmission and potential consequences for human health.
Our results also showed that both the perceptions and behaviors related to zika and its prevention vary considerably among different social groups and geographic areas. Notably, there was a social gradient in the cognitive and affective responses to the risk of ZIKV infection; more educated participants were found to perceive the disease as less severe and controllable and more understandable. Some gender effects were also observed during the epidemic, as there were systematic and significant differences in the way men and women perceive the health threat related to the spread of zika in the region. Female population has been particularly responsive to awareness campaigns and rapidly understood the extent of risks associated with ZIKV infection. This result is not surprising, since (1) ZIKV has been found to cause severe fetal anomalies among babies born to infected women, and (2) the ‘gender gap’ is one of the best-documented phenomena of social and cultural influence in the field of risk perception [37,38].
However, results showed that perceptions and behaviors discrepancies appeared among women. While the level of education had a weak impact on risk perception among men, it was clearly identified as a potential determinant of risk perception among women. Importantly, the level of education was also associated with the adoption of protective measures among women. This finding provides health authorities the opportunity to target and adapt future health messages to less advantaged women. Institutions in charge of epidemic and emergent diseases should thus reach this population visiting mother and child protection centers which are numerous in French Guiana both located in the coastal area and in the interior. Actions should first reassure this population about risks associated with zika and emphasize on the importance of protective measures adoption during an outbreak to encourage behavior changes. In fact, a previous study revealed that the level of comprehension was associated with the adoption of protective behaviors [17].
Findings showed that the main communication media were available to all subgroups of population, even less advantaged people had access to television. However, leaflets seemed to be one of the most informative mean and this is not surprising since a lot of information can be transmitted through this medium, whether a television spot must be short, posters must be able to be seen from a distance and no visual data are broadcasted through radio. A continuous effort is thus needed to efficiently transmit prevention messages and target populations through leaflets and maybe adapt others means of communication to be more efficient.
Limitations of the study are inherent to the design of the study which was based on a phone recruitment process. In fact, the interview lasted 25 minutes and may lead to selection biases with more educated and advantaged participants that would have been more sensitive to public health issues addressed by the study. Moreover, the network coverage is heterogeneous on the territory and some uncovered isolated municipalities representing 7 municipalities (out of 22) could not be investigated. However, these failures concerned isolated and rural villages poorly or not affected by the presence of Ae. aegypti. These areas represent less than 5% of the total population.
The results of our survey indicated that the public perceived zika disease as a more serious potential health threat than other common mosquito-borne diseases, even though a range of elements within cognitive representations of zika were statistically and graphically found to be anchored on those of chikungunya. The survey helped to identify a subgroup of population shaped by specific risk perceptions and behaviors that deserves further attention given the importance of the public understanding and mental representation of illnesses in the adoption of effective protective behaviors. If this assessment seems difficult to conduct in an epidemic setting, high-risk groups identified may be targeted as a priority in case of a new emergence.
|
10.1371/journal.pcbi.1003023 | Evolutionary Capacitance and Control of Protein Stability in Protein-Protein Interaction Networks | In addition to their biological function, protein complexes reduce the exposure of the constituent proteins to the risk of undesired oligomerization by reducing the concentration of the free monomeric state. We interpret this reduced risk as a stabilization of the functional state of the protein. We estimate that protein-protein interactions can account for of additional stabilization; a substantial contribution to intrinsic stability. We hypothesize that proteins in the interaction network act as evolutionary capacitors which allows their binding partners to explore regions of the sequence space which correspond to less stable proteins. In the interaction network of baker's yeast, we find that statistically proteins that receive higher energetic benefits from the interaction network are more likely to misfold. A simplified fitness landscape wherein the fitness of an organism is inversely proportional to the total concentration of unfolded proteins provides an evolutionary justification for the proposed trends. We conclude by outlining clear biophysical experiments to test our predictions.
| The folded form of proteins is only marginally stable in vivo and constantly faces the risk of aggregation, unfolding/misfolding, and other aberrant interactions. For most proteins, the folded form is also the functionally relevant one and forces of natural selection strongly modulate its stability. In vivo, proteins interact with each other on a genome-wide scale. Usually, the interaction of a protein and its binding partners requires both the proteins to be in the folded form and as a result, the interactions tend to shift the population of a protein towards the folded form. Consequently, protein-protein interactions interfere with the evolution of protein stability. Here, we present empirical evidence and theoretical justification for proteins' ability to stabilize the folded form of their interaction partners and allow them to explore the region of the sequence space that corresponds to proteins with less stable structure. We argue that the ‘evolutionary capacitance’ – previously thought to be a property of the chaperone HSP90, a special class of proteins – is a property of all proteins, albeit to a different degree.
| The toxicity due to protein misfolding and aggregation has a considerable effect on the viability of living organisms [1]–. Consequently, cells are under strong selection pressure to evolve thermodynamically stable [6] and aggregation-free protein sequences [7]. The internal region of stable proteins has a tightly packed core of hydrophobic residues. A mutation in the core may disrupt the entire protein structure. Consequently, the core residues are strongly conserved [8], [9]. In contrast, mutations on the surface contribute weakly to the thermodynamic stability of proteins [10] yet surfaces show significant level of conservation [11] owing to protein-protein interactions.
Recent high throughput experiments have established that proteins interact with each other on a genome-wide scale [12]. Such ‘small world’ networks are thought to facilitate biological signaling and ensure that cells remain robust even after a random failure of some of its components [13]. It is thought that evolutionarily, multi-protein complexes are favored over larger size of individual proteins [14] since large proteins are difficult to fold and expensive to synthesize while small interacting proteins can fold independently and then efficiently assemble into large complexes. Individual interaction between proteins can give rise to cooperativity and allostery which results in a finer control over the functional task the protein complex performs. Protein-protein interactions (PPI) are also thought to prevent protein aggregation [15], [16]. Lastly, many proteins can perform promiscuous function in that they can partake in multiple protein complexes. Interestingly, proteins in higher organisms are involved in more interactions and form larger protein complexes compared to more primitive life forms [17].
Here, we hypothesize an additional biophysical advantage for protein-protein interactions. Proteins bound to their interaction partners effectively present a lower monomer concentration inside the cell. Since free monomers are susceptible to misfolding/unfolding and toxic oligomerization, interacting proteins may face a reduced risk towards the same. This reduced risk can be interpreted as interaction-induced stabilization — stabilization due to the protein-protein interaction network — of an otherwise monomeric protein (see Fig. 1 for a cartoon). We propose that by giving proteins an additional stability, each protein in the interaction network acts as an evolutionary capacitor [18], [19] in the evolution of its binding partners: proteins are allowed to explore the less stable regions (regions of low intrinsic stability) of the sequence space as long as they are stabilized by their interaction partners. Inversely, unstable proteins are expected to receive significant additional stability from the interaction network.
Below we outline the empirical evidence for our hypothesis and suggest clear biophysical and evolutionary experiments to test it further.
We present our estimates of the interaction-induced stability (see Methods) and explore the evolutionary interplay between and protein stability using a simplified fitness model for a toy proteome. We test the predictions of the toy model on the proteome of baker's yeast. The fitness model also sheds light on the interplay between protein stability and protein abundance.
Fig. 2 shows the histogram of the estimated interaction-induced stability for cytoplasmic yeast proteins for whom abundance, interaction, and localization data is available (see Methods for the details of the calculations). Note that the average PPI induced stability is and can be as high as . This stabilization is dependent not only on the number of interaction partners of a given protein or the strengths of those interactions but also on the relative abundances of the interaction partners. In fact, the interaction-induced stability of a protein correlates strongly with the relative concentration of its binding partners(Spearman . This suggests a plausible mechanism of stabilization of a protein without changing its sequence viz. via adjusting the expression levels of its interaction partners (see Discussion below).
The estimated values are of the same order of magnitude as the inherent stabilities of proteins, () [9]. Given that random mutations are more likely to destabilize proteins [6], we expect protein-protein interactions to act as secondary mechanisms to stabilize proteins and to interfere with the evolution of protein stability.
To explore the evolutionary consequences of the interaction-induced stability, we investigate a simplified fitness model of a toy proteome consisting of 15 proteins (see Methods, Text S1, and Table S1). Briefly, the fitness of the cell depends only on the total concentration of unfolded proteins in it [20]. During the course of evolution, each protein acquires random mutations that change either a) its inherent stability or b) the dissociation constant of its interaction with a randomly selected interaction partner. Even though protein abundance and protein-protein interactions evolve at the same time scale as protein stability, the former are dictated largely by the biological function of the involved proteins. Incorporating the fitness effects of changes in expression levels and interaction partners in our simple model is non-trivial. Thus, in order to specifically probe the relation between stability and interactions, we do not allow proteins to change their abundance and interaction partners.
In the model, the concentration of unfolded proteins and thus the fitness of the proteome depends on the total stability of individual proteins. While random mutations are more likely to make proteins unstable, protein-protein interactions increase the total stability. In the canonical ensemble description of the evolution of fitness [21], the inverse effective population size (), the evolutionary temperature quantifies the importance of genetic drift. The effective population size modulates the competition between destabilizing random mutations and stabilizing protein-protein interactions.
We find that at higher effective populations, proteins are inherently stable and only the least stable proteins (small ) receive high stabilization from the interaction network (high ). At low effective population, due to genetic drift, proteins are inherently destabilized and protein-protein interactions serve as the primary determinant of the effective stability of proteins. Fig. 3 shows the dependence of average inherent stability (), average interaction-induced stability (), and average total stability () with effective population size. Interestingly, the total stability () of proteins remains relatively insensitive to changes in population size.
We observe that the correlation coefficient between the inherent stability and the interaction-induced stability itself varies with the effective population size. Even though its magnitude decreases, interaction-induced stability becomes more and more correlated with inherent stability as population size increases (See Fig. 4). In real life organisms, interaction-induced stability acts on a need basis for proteins and serve as a secondary stabilization mechanism. In the drift-dominated regime, which is unlikely to be realized in real life organisms (except probably in parasitic microbes with low population sizes), interaction-induced stability becomes the dominant player in the evolution of total stability of proteins [17]. We next examine if this prediction from the toy model holds for real organisms.
Proteome-wide information about the inherent stability of proteins is currently unavailable. Previously, in silico estimates of protein aggregation propensity have been used as proxy for protein stability [22], [23]. We use the TANGO [24] algorithm to estimate protein aggregation propensity. It is known that TANGO aggregation propensity correlates strongly and negatively with protein stability [24]. TANGO has been verified extensively with experiments on peptide aggregation [24] and has been previously used to study the evolutionary aspects of protein-protein interactions [22], [25]. Similar analysis for Aggrescan [26] can be found in Text S1 and Table S3. We find that the aggregation propensity is correlated positively with the interaction-induced stability (Spearman ). As expected [2], the aggregation propensity is negatively correlated with protein abundance (Spearman ). The correlation between and does not depend on this underlying dependence and persists even after controlling for total abundance (partial Spearman ) (See Table S2). This result suggests in the proteome of baker's yeast, protein stability correlates negatively with interaction-induced stability.
The fitness cost of protein aggregation is directly proportional to the amount of aggregate [20]. Thus, the selection forces that make protein sequences aggregation-free act more strongly on highly expressed proteins [1], [2], [22]. Our hypothesis suggests that the proteins that are bound to their interaction partners present a lower concentration of the free monomeric state in vivo (low ) and automatically lower the misfolding/aggregation induced fitness cost, even if highly abundant (high ). The selection forces to evolve an aggregation-free sequence may be weaker for such proteins. Consequently, the aggregation propensity should be principally correlated with the free monomer concentration rather than the total abundance .
Indeed, we observe that the estimated monomer concentration and the aggregation propensity are correlated negatively (Spearman ). Importantly, this correlation is not an artifact of the underlying correlation between the aggregation propensity and total abundance (partial Spearman ). At the same time, the partial correlation coefficient between the aggregation propensity and the total protein abundance controlling for the estimated monomer concentration is minimal (partial Spearman ). In short, the total free monomer concentration of a protein (rather than , its total abundance) might be a better variable to relate to evolutionary and biophysical constraints on the protein.
We have thus far shown that a protein's interaction partners can significantly stabilize its folded state and this stabilization interferes with the evolution of the inherent stability of the protein. We now explore the reverse viz. the evolutionary consequences of the ability of each protein to impart stability to its interaction partners.
The concept of evolutionary capacitor has been previously introduced for the heat shock protein HSP90 [18], [19], which is also a molecular chaperone and a highly connected hub in the PPI network (70 interaction partners in the current analysis). An elevated concentration of HSP90 buffers the potentially unstable variation in proteins, which may allow proteins to sample a wider region of the sequence space, which may often lead to functional diversification [27]. Similar to HSP90, each protein in the interaction network has some ability to stabilize its interaction partners to a certain extent. Consequently, we study the evolutionary capacitance of individual proteins in the context of the interaction network by estimating the effect of protein knockout on ppi-induced stability in silico. Proteins with higher evolutionary capacitance are defined as those with the higher cumulative destabilizing effect on the proteome. We write,(1)For each protein , the sum in Eq. 1 is carried out over all proteins that are destabilized due to its knockout. Here, we assume that the potential of a given protein knockout to generate multiple phenotypes depends on the loss of stability of its interaction partners caused by its knockout. We hypothesize that, similar to unstable proteins requiring HSP90 to fold, the interaction partners of proteins with high capacitance should be unstable. In fact, the capacitance of a protein and the mean aggregation propensity of its interaction partners are strongly correlated (Spearman ). The capacitance is significantly correlated with even after controlling for the abundance of the protein (partial spearman ) and the number of its interaction partners (partial spearman ). This suggests that a protein needs to be present in sufficient quantity and should interact with a large number of proteins in order to effectively act as a capacitor.
We have presented evidence that all proteins can act as an evolutionary capacitor, albeit with variable effectiveness, for their interaction partners. Traditionally, evolutionary capacitors are understood to be chaperones that buffer phenotypic variations by helping misolding-prone proteins fold in a proper structure [19]. Not surprisingly, when we carried out functional term enrichment analysis using gene ontology [28], we found that approximately half of the top 20 capacitors have ‘chaperone’ in their name. The top 20 are also over represented in the chaperone-like molecular function of protein binding and unfolded protein binding () and the biological process of protein folding (). These findings validate our definition of capacitors that were previously identified as chaperones. Interestingly, some of the predicted capacitors do not currently have a protein folding-related functional annotation. These need more experimental investigation (see supplementary File S1 for the list). This suggests that previously identified evolutionary capacitor HSP90 may in fact only be one among the broader set of evolutionary capacitors. Every protein in the interaction network is an evolutionary capacitor for its interaction partners and evolutionary capacitor is a quantitative distinction rather than a qualitative one.
Recently, Fernández and Lynch [17] showed that random genetic drift is the chief driving force behind thermodynamically less stable yet densely interacting proteins in higher organisms [17]. Additionally, protein complexes in higher organisms have more members than in lower organisms [14]. Recently, it was observed that a destabilizing mutation in the enzyme DHFR in E. coli leads to functional tetramerization of the otherwise monomeric enzyme [29] suggesting that protein-protein interactions can at least partially compensate the effect of protein destabilization. lactoglobulin is an aggregation-prone protein generally found as a dimer. It was shown that the specific interactions responsible for the formation of the dimer considerably reduce the risk of protein aggregation [16]. Ataxin-3 is a protein implicated in polyglutamine expansion diseases wherein the functional interactions of the protein reduce the exposure of its aggregation prone interface and thereby decrease its aggregation propensity [15].
Here, we have quantified the interaction-induced stability on a proteome wide scale and hypothesized that the PPI-induced stabilization is a secondary evolutionary advantage of the PPI network; alleviating the selection pressure on proteins in functional multi-protein complexes to evolve a stable folded. A simple model for the fitness of the proteome provided a fundamental justification for the co-evolution of protein stability and protein-protein interactions and made predictions that were tested on the proteome of baker's yeast. In the model, when the effects of natural selection are weak, proteins acquire stability mainly via protein-protein interactions. At a higher population size — in the absence of genetic drift — proteins are intrinsically stable and protein-protein interactions stabilize only those proteins that fail to evolve inherent stability.
We have also presented evidence that all interacting proteins stabilize their binding partners to a certain extent and act as the evolutionary capacitance [19] for their evolution. Interestingly, though some of the top 20 capacitors predicted in this study are known chaperones and are over-represented in GO ontology terms such as protein binding, unfolded protein binding, and protein folding; others do not have any protein folding-related functional annotation and need experimental investigation.
The importance of disordered proteins, especially in the proteomes of higher organisms, cannot be neglected. The proteome of baker's yeast does not have many completely disordered proteins but of the amino acids in the proteins of yeast are predicted to be in a disordered state [30] ( for the proteins considered in this study, see supplementary Text S1 and Fig. S4). Even though the development presented above applied only to an equilibrium between folded and unfolded/misfolded/aggregated protein, it can be easily generalized to disordered proteins. This is because even though the folded unfolded equilibrium is not well defined, similar to well structured proteins, disordered proteins also exist either in a soluble monomeric (instead of the folded state), a misfolded/aggregated, and a complexed state. Many disordered proteins acquire a definite structure when bound to their interaction partners and seldom dissociate to the soluble monomeric [31]. These serve as even stronger candidates for the beneficiaries of interaction-induced stability compared to folded proteins. Consequently, we include both partially disordered proteins and structured proteins in the current analysis of the cytoplasmic proteins.
In cellular homeostasis, the total concentration of any protein can be written as the sum of its free folded monomer concentration , a fraction comprising of insoluble oligomers and unfolded peptide , and as part of all protein complexes containing (See Fig. 5). In our computational model, for simplicity and owing to the nature of the large scale data [34], we restrict protein complexes to dimers [35], thus for all proteins that interact with ,(2)Conservation of mass implies,(3)The concentration of each dimer satisfies the law of mass action,(4)We can write the balance between the three states of the protein, (See Fig. 1), as two equilibrium equations(5)(6)Note that comprises of a collection of biologically unusable states of the protein viz. the misfolded/unfolded and the oligomerized state any of which may convert to/interact with the folded monomeric state . Consequently, the first equilibrium is a collection of thermodynamic equilibriums. The equilibrium constant will thus depend not only on the temperature but also on and . If among the unfolded, misfolded, and the oligomerized states the former dominates the population comprising then, where is the thermodynamic stability of the free monomeric state. Similarly, is given by,(7)and depends not only on the dissociation constants but also the free concentrations of the interacting partners of protein and on the topology of the interaction network in the organism. Here too, we assume that a) only the folded monomeric forms of proteins interact with each other and b) there is no appreciable interaction between the collective unfolded state of protein and any state of any other protein . We have also neglected the role of chaperones in actively reducing the concentration of the unfolded/misfolded/aggregated state by turning it over to the folded state. In fact, some of the chaperones are included in of our mass action equilibrium model and prevent unfolding by sequestering the folded state (see below and the discussion section).
By combining mass conservation (Eq. 3) with Eq. 5 and Eq. 6,(8)In the above development, we have made a crucial assumption that only.
Note that in the absence of interactions, . We identify as the additional decrease in the insoluble fraction due to protein-protein interactions. We define the interaction-induced stability as,(9)
We downloaded the latest set of interacting proteins in baker's yeast from the BIOGRID database [36]. To filter for non-reproducible interactions and experimental artifacts, we retained only those interactions that were confirmed in two or more separate experiments. For the sake of simplicity, we only considered cytoplasmic proteins [37] with known concentrations [38]. This lead to proteins connected by interactions.
The in vivo stability of a protein is a combination of its thermodynamic stability, resistance to aggregation or oligomerization, and resistance to degradation [39]. Note that the interaction-induced stability of a protein depends on the stability of its interaction partners (see Eq. 6, Eq. 7, and Eq. 9). Unfortunately, the exact dependence of the in vivo protein stability on its sequence is unclear and there exist no reliable data or sequence dependent computational estimates for the thermodynamic stability of proteins. Moreover, , and thus (Eq. 6, Eq. 7, and Eq. 9), can be estimated even in the absence of the knowledge of . In our estimates of , we assume that is given simply byHere, is obtained by solving the mass action equations [35] iteratively (see below). This is equivalent to assuming that all the proteins are equally and highly stable ( for all proteins ). The thus calculated serves as the upper limit of interaction-induced stability. In the supplementary materials (Text S1, Fig. S1, Fig. S2, and Tables S4 and S5), we show that different assignments of the equilibrium constants including a simple model of protein stability [40]–[42] do not change the qualitative nature of our observations.
The dissociation constants for protein-protein interactions follow a lognormal distribution with a mean nM [35]. The majority of interactions between proteins are neither too weak nor unnecessarily strong. Common sense dictates that it does not make sense to decrease the dissociation constant between two proteins beyond the point where the abundance limiting protein spends all of its time in the bound state. Motivated by these evolutionary arguments to minimize unnecessary protein production and to avoid unnecessarily strong interactions, Maslov and Ispolatov [35] devised a recipe to assign dissociation constants to individual protein-protein interactions. viz. for interacting proteins and , the dissociation constant . We also explore a few other assignment rules for dissociation constants (see supplementary Text S1, Fig. S3, and Table S6).
We solve for free concentrations iteratively [35]. We start by setting for all proteins and iteratively calculate from(10)till two consecutive estimates of fall within of each other for all proteins.
As noted above, the toxic effects of misfolding and aggregation may be the chief determinant of protein sequence evolution [2], [4], [5]. The dosage dependent fitness effect of misfolded proteins [20] motivates us to introduce a simple biophysical model for fitness of the proteome (See Eq. 11),(11) is the scaling factor. Potentially, can be estimated from fitness experiments by introducing measured quantities of unfolded protein in the cell [20]. We explore the evolution of a hypothetical proteome to investigate the interplay between protein stability and protein-protein interactions.
We believe that protein abundances and the topology of the interaction network are largely dictated by biological function. It is non-trivial to incorporate the fitness effect of changes in gene expression level and the network topology in our simplified model. Thus, to specifically probe the relation between stability and interactions, we concentrate on the effect of toxic gain of function due to misfolding and aggregation on cellular fitness and not include changes in gene expression levels and network topology. In this aspect, our model is in the same spirit as previously proposed models [6], [41]–[48]. The effect of random mutations on average destabilizes proteins and the dynamics of the evolution of thermodynamic stability of proteins can be modeled as a random walk with negative average velocity [6]. We consider the thermodynamic stability as a proxy for the in vivo stability of proteins. We construct the cytoplasm of a hypothetical organism with 15 proteins. The number of proteins is low due to computational restrictions. The proteome is evolved by sampling the dissociation constants from the lognormal distribution while introducing random mutations in proteins that change their stability. At each generation, the fitness is evaluated and the progeny is accepted at a certain evolutionary temperature (defined as the inverse of the effective population size, ) [21]. We run a total of generations for each evolutionary temperature and analyze the organism in the latter half of the evolutionary run (details of the model and a brief description of the population genetics terminology is in supplementary Text S1).
The notion of protein stability relevant to this study is the propensity of a protein to avoid structural transformations that may render it unemployable for biological function. For example, for a small and highly soluble protein, this stability corresponds to the thermodynamic stability of the native state while for a large multi domain protein, it may correspond to the thermodynamic stability of one of its domains against the partially unfolded state. In short, thermodynamic stability of the folded state with respect to the unfolded, partially folded state, and the misfolded state all contribute to the in vivo stability of proteins [39].
Though there is a lack of proteome-wide estimates of thermodynamic stability of proteins, the aggregation propensity can be estimated from the sequence [24], [26] and is known to be correlated with protein stability [24]. In our correlation analysis, we use the estimated aggregation propensity as a proxy for in vivo protein stability and explore the relationship between interaction-induced stability and protein stability. The aggregation propensity was estimated for the same proteins used in the mass action calculation to estimate . We tested the TANGO [24] and Aggrescan [26] to estimate the aggregation propensity of proteins. Previously, TANGO has been used [22], [23], [49] to understand the relation between protein abundance and instability. We show results for TANGO in the main text. Aggrescan results (supplementary Text S1 and Table S3) are quite similar.
|
10.1371/journal.pcbi.1004977 | Machine Learning Meta-analysis of Large Metagenomic Datasets: Tools and Biological Insights | Shotgun metagenomic analysis of the human associated microbiome provides a rich set of microbial features for prediction and biomarker discovery in the context of human diseases and health conditions. However, the use of such high-resolution microbial features presents new challenges, and validated computational tools for learning tasks are lacking. Moreover, classification rules have scarcely been validated in independent studies, posing questions about the generality and generalization of disease-predictive models across cohorts. In this paper, we comprehensively assess approaches to metagenomics-based prediction tasks and for quantitative assessment of the strength of potential microbiome-phenotype associations. We develop a computational framework for prediction tasks using quantitative microbiome profiles, including species-level relative abundances and presence of strain-specific markers. A comprehensive meta-analysis, with particular emphasis on generalization across cohorts, was performed in a collection of 2424 publicly available metagenomic samples from eight large-scale studies. Cross-validation revealed good disease-prediction capabilities, which were in general improved by feature selection and use of strain-specific markers instead of species-level taxonomic abundance. In cross-study analysis, models transferred between studies were in some cases less accurate than models tested by within-study cross-validation. Interestingly, the addition of healthy (control) samples from other studies to training sets improved disease prediction capabilities. Some microbial species (most notably Streptococcus anginosus) seem to characterize general dysbiotic states of the microbiome rather than connections with a specific disease. Our results in modelling features of the “healthy” microbiome can be considered a first step toward defining general microbial dysbiosis. The software framework, microbiome profiles, and metadata for thousands of samples are publicly available at http://segatalab.cibio.unitn.it/tools/metaml.
| The human microbiome–the entire set of microbial organisms associated with the human host–interacts closely with host immune and metabolic functions and is crucial for human health. Significant advances in the characterization of the microbiome associated with healthy and diseased individuals have been obtained through next-generation DNA sequencing technologies, which permit accurate estimation of microbial communities directly from uncultured human-associated samples (e.g., stool). In particular, shotgun metagenomics provide data at unprecedented species- and strain- levels of resolution. Several large-scale metagenomic disease-associated datasets are also becoming available, and disease-predictive models built on metagenomic signatures have been proposed. However, the generalization of resulting prediction models on different cohorts and diseases has not been validated. In this paper, we comprehensively assess approaches to metagenomics-based prediction tasks and for quantitative assessment of microbiome-phenotype associations. We consider 2424 samples from eight studies and six different diseases to assess the independent prediction accuracy of models built on shotgun metagenomic data and to compare strategies for practical use of the microbiome as a prediction tool.
| The human microbiome constitutes the whole set of microbial organisms associated with the human host. It has been shown to be crucial for human health and for the development and maintenance of the immune system and for several metabolic activities [1–3]. Significant effort has been devoted to its characterization in healthy individuals and subjects with a variety of diseases such as inflammatory bowel diseases [4,5], obesity [6,7], and type-2 diabetes [8]. Consequently, the potential use of the microbiome as a diagnostic tool is a promising line of investigation [9]. In addition, even when the findings are not immediately relevant for the clinical setting, identifying associations between the microbiome and specific diseases is essential for follow-up mechanistic studies.
Next-generation DNA sequencing technologies permit comprehensive profiling of the microbial communities from human-associated samples, and have now been sufficiently widely employed to enable meta-analysis for discovering patterns common to independent studies. Meta-analysis has been broadly adopted in other genomics applications, such as for analysis of microarray or RNA-seq data, where multiple studies have been performed for a similar purpose including identifying gene expression signatures of specific human cancers. The general objectives of meta-analysis include proposing new classifiers [10], comparing different classification methods [11], finding a common transcriptional profile [12], and evaluating generalization of prediction models across different studies [13]. In genomics, rigorous meta-analyses are crucial both to validate the findings of each single study, and for providing robust models for clinical purposes.
The most common and cost-effective approach for microbiome characterization to date targets the 16S rRNA gene as taxonomic marker [14]. Meta-analyses and independent validation of such experimental approach have identified differences in microbiome composition or function by body site, age, and disease state [15–17], and have been conducted to determine the most effective techniques for disease classification [18]. More recently, shotgun metagenomics [19] provided expanded resolution to the level of microbial species [20–22] and strain [23], to the fungal and viral kingdoms [24], and to the level of individual genes across the metagenome [25,26]. The decreasing cost of shotgun metagenomics is rapidly increasing the number of available human disease-associated datasets; however, the generalization of resulting prediction models is still unclear.
Improved resolution and lower variability of shotgun metagenomics hold the promise to provide improved generalization of microbial signatures over 16S rRNA sequencing [27]. Meta-analyses on specific host characteristics have been performed (e.g., with respect to host age [28]). The importance of cross-cohort consistency and validation of predictions has also been recognized, with some works assessing the structure of the microbiome in European cohorts [29] and combined European-American cohorts [21, 30]. Some studies focusing on the link between host conditions and microbiome further provided a validation step with respect to other single investigations [31, 32]. Although these works provided a first assessment on the transferability of condition-associated microbiome features across cohorts, no systematic assessments have been performed on clinical outcomes using the full archive of shotgun metagenomic data now publicly available, and no convenient software frameworks for doing so are available in the community.
In this study we uniformly process 2424 shotgun metagenomic samples from eight studies to assess the independent prediction accuracy of models built on metagenomic data and to compare strategies for practical use of the microbiome as a prediction tool. The software framework and the microbiome profiles for thousands of samples are made publicly available.
We evaluated alternative approaches to metagenomics-based prediction tasks, and assessed the strength of microbiome-phenotype associations using publicly available raw sequence data. For this purpose, we developed a machine-learning software framework which uses as features quantitative microbiome profiles, including species-level relative abundances and presence of species- and strain-specific markers (see Methods). Our multi-level validation strategy includes the assessment of microbiome models on single cohorts, across stages of the same study, across different studies, and across target outcomes and conditions (Fig 1). The software and validation framework is publicly available at http://segatalab.cibio.unitn.it/tools/metaml and was applied on a total of 2424 publicly available metagenomic samples from eight large-scale studies (see Table 1 and Methods). All samples were processed with MetaPhlAn2 [21] for quantitative species- and subspecies-level taxonomic profiling after standard sequencing data pre-processing (see Methods).
We first assessed the prediction power of metagenomic data in linking the gut microbiome with disease states. For such purpose, we considered six available disease-associated metagenomic datasets spanning five diseases: liver cirrhosis [33], colorectal cancer [34], inflammatory bowel diseases (IBD) [35], obesity [31], and type 2 diabetes (two distinct studies—[37] and [32]). Each dataset was analyzed independently using cross-validation (denoted as CV in Fig 1), which repeatedly uses part of the samples with associated known phenotype for learning the statistical model, and the remainder for validating the predictions (see Methods). The support vector machines (SVM) [38] and random forest (RF) [39] classifiers were used for this evaluation as they are state-of-the-art approaches and are appropriate for this type of data [18]. We also evaluated Lasso [40] and elastic net (ENet) [41] regularized multiple logistic regression. Neural networks [42] and Bayesian logistic regression [43] represent other possible alternatives not evaluated here.
Prediction performance was evaluated by the area under the curve (AUC) metric, which summarizes true positive and false positive rates and is robust to unequal proportions of each outcome. Using MetaPhlAn2 species abundance [21] as input data produced high accuracy for disease classification (Fig 2), although prediction performance varied considerably between datasets. The most predictable disease state appears to be liver cirrhosis (AUC = 0.945, 95% CI: 0.909–0.981 for the best classifier), followed by colorectal cancer (AUC = 0.873, 95% CI: 0.802–0.944), and IBD (AUC = 0.890, 95% CI: 0.812–0.968). For IBD we considered Crohn and ulcerative colitis patients together due to the low number of cases in the datasets compared to controls (as general rule at least ten samples per class are required for reliable prediction models). Stronger signatures might be found when considering the two conditions separately with adequate sample size, as it has been observed that some bacterial features are specific to Crohn disease only [16]. Confounding factors such as active treatment could of course lead to overestimated prediction capabilities [44, 45], but we adopted here the same contrasting approach used in the original works.
For the other diseases we achieved lower discrimination capabilities, suggesting less dramatic microbial shifts in the patients. For type 2 diabetes, although the two considered datasets have independently sampled and geographically distinct cohorts, we obtained very similar AUC values for both (0.744, 95% CI: 0.688–0.800 and 0.762, 95% CI: 0.651–0.873 for T2D and WT2D, respectively). Prediction of obesity generated the lowest AUC (0.655, 95% CI: 0.576–0.734). Despite a wide range of classification performances, all investigated datasets showed a substantial level of association between disease and the microbiome (Fig 2), with AUC values significantly higher than those obtained by the same classifier applied to the same data with shuffled class labels (p-values ranging from 9.9 × 10−3 for obesity to 5.6 × 10−7 for cirrhosis, S1 Table).
Comparing the accuracy of SVM and RF classifiers, RFs exhibited in all cases similar or better results than SVM. In particular, accuracies differed substantially for three datasets: AUC increased from 0.809 to 0.873 for colorectal, from 0.663 to 0.744 for T2D (difference also supported by statistical significance, p-value 0.011, see S1 Fig), and from 0.664 to 0.762 for WT2D. In two cases, slight improvements were verified: AUC increased from 0.922 to 0.945 for cirrhosis and from 0.862 to 0.890 for IBD. Methodologically, our results thus suggested the use of RFs for disease prediction from species abundances.
We then investigated how feature selection, i.e., the procedure of selecting a reduced subset of relevant discriminative features, impacts the prediction accuracy. To this end, we used the RF classifier that implicitly embeds a feature selection step during the model generation phase (see Methods). Feature selection produced a slight improvement of the AUC in all the cases when the model was generated on a reduced set of species (Fig 3). The advantage of this procedure is twofold. In addition to the increased accuracy, it enables biomarker discovery by detecting the (few) species that are most useful to discriminate between “healthy” and “diseased” subjects. These most discriminative species may be prioritized when performing follow-up and validation analyses, and the reduced complexity of the model potentially enables additional evaluations on low-throughput assays. However, the best accuracies were obtained with still relatively high numbers of species, i.e., more than 60 (S2 Fig). This confirms the complexity of microbial ecosystems where the combination of few species is probably not sufficient to characterize the microbiome associated with complex diseases.
We then investigated the use of strain-specific markers, as opposed to species-level taxonomic abundance, by applying the same classification and cross-validation methods to strain-specific microbial features generated by MetaPhlAn2. Given their strain-specificity, adopting markers as features let us also test the hypothesis that complex diseases are associated with the presence of specific strains or subspecies rather than only species-level abundances. Consistent with this hypothesis, better predictions were obtained from markers (Fig 2) than species abundance, with differences that were statistically significant for one dataset (S1 Fig). This was obtained using SVM with linear kernel, which in this context is more practical to use than RF and SVM with more complex kernels due to the very high dimensionality (~100K features) of the data. Focusing on SVM, markers gave statistically significant improvements with respect to species abundances in half of the datasets (S1 Fig). Moreover, RF in combination with feature selection (RF-FS:Emb) achieved satisfactory classification results, i.e., average accuracies were usually worse than SVM but with no statistically significant difference (S1 Fig) even using a very limited portion (<0.2%, S2 Fig) of the investigated markers. The biomarker discovery step here is of particular interest because it permits identification of a limited set of strain-specific markers potentially directly involved in the association with disease.
We also considered alternative approaches to feature selection based on Lasso and ENet (see Methods). Applying Lasso or ENet as pure classifiers, which implicitly incorporates the feature selection and classification steps, did not give satisfactory results, with AUC worse than RF or SVM for both species abundance and marker features (S3 Fig). Better accuracies were obtained by using them for feature selection only, followed by RF or SVM classification. However, both Lasso and ENet feature selection in general worsened the performance of RF and SVM without prior feature selection. Finally, ENet worked better than Lasso, although it was associated with more time-consuming tuning of its free parameters on a two-dimensional grid.
Feature selection can also be used for biomarker discovery, and several tools have been developed specifically for this task in metagenomics [46–48]. The approach proposed here (RF with embedded feature selection) focuses on the set of features with the most discriminating power rather than on strictly statistical assessments [46] or statistical assessment coupled with effect size [47]. The implemented tool automatically plots the most relevant species (or markers) with the importance factor (see Methods) along with the average relative abundance (or average presence) associated with the different considered classes. We observed a reasonable level of overlap between the detected species and markers, as the most discriminative markers tended to represent strains of the most discriminative species. Interestingly, for all the considered datasets (Fig 4 and S4 Fig) the importance factor attributed to each species (or marker) was not well correlated with its average relative abundance (or presence) in the samples (maximum correlation of 0.49 for the T2D dataset, S5 Fig). In several cases, we detected relevant species with partial prevalence but highly discriminative potential between “healthy” and “diseased” subjects. For example, Peptostreptococcus stomatis resulted the most discriminative species in the colorectal dataset with an average relative abundance in the samples less than 0.15%.
In the cirrhosis dataset, the most relevant taxonomic abundances were enriched in diseased patients. The top features were especially related to the Veillonella (Veillonella spp., Veillonella dispar, Veillonella parvula, and Veillonella atypica) and Streptococcus genera (Streptococcus anginosus and Streptococcus parasanguinis) in addition to Haemophilus parainfluenzae, which is consistent with findings of the original study [33]. Species belonging to Veillonella and Streptococcus are typical colonizers of the oral cavity, but they are often overgrown in the small intestine in patients affected by liver cirrhosis, thus suggesting the invasion of the gut from the mouth in these patients [33]. Moreover, species such as Veillonella spp., V. dispar, V. atypica, and S. anginosus were already associated with opportunistic infections [33]. Also the H. parainfluenzae pathogen may arrive to the gut from the oral cavity [33]. In the colorectal dataset we identified five major species: P. stomatis, Fusobacterium nucleatum (both enriched in diseased patients) and Streptococcus salivarius (depleted in diseased subjects) as found in the original study [34], in addition to Parvimonas spp. and Parvimonas micra.
We then compared the discriminative species across datasets through hierarchical clustering (S6 Fig). We found some species that were distinctive of one disease only as it is the case for P. stomatis, P. micra and Gemella morbillorum in colorectal cancer, multiple Veillonella species in cirrhosis, and, partially, Bifidobacterium bifidum and Lachnospiraceae in IBD. Interestingly, F. nucleatum was highly discriminant both in colorectal cancer and cirrhosis, suggesting the presence of a similar dysbiosis niche for this organism. Overall, the discriminative species for the two diabetes datasets and the obesity dataset had lower weights, consistent with the lower classification performances achieved with them. Moreover, the pattern of discriminative species for these two datasets clustered together (S6 Fig), suggesting similar dysbiotic configurations of the gut microbiome for obesity and type-2 diabetes. Some species were also found in the set of top discriminative features for all the studies, in particular S. salivarius, S. anginosus, V. parvula, Roseburia intestinalis, and Coprococcus comes. These species might thus be biomarkers of general dysbiosis or ecological community stress in non-healthy states, and should be recognized as such in future disease-microbiome association studies.
We extended the cross-validation analysis by evaluating the predictability for non-disease based classification problems. Gender discrimination (S7 Fig, part a) exhibited in general low classification accuracy with an AUC close or less than 0.6 for most of the considered datasets. However, statistically significant discrimination was verified in some cases (AUC equal to 0.662 and 0.796 for skin and IBD dataset, respectively, both p < 0.05 by permutation test with shuffled labels), which may suggest some gender-dependent differences in the human microbiome as highlighted by recent studies [49]. High classification accuracy in body site prediction in the Human Microbiome Project (HMP) dataset (AUC = 0.96), is consistent with previously reported large differences in the microbiome composition among different body areas [1], and provided validation of the proposed tool for multi-class classification problems (S8 Fig). The confusion matrix revealed moderate misclassification between nasal and skin body sites, which may be due to nasal samples being taken from the anterior nares (external part of the nostrils), and thus having relatively similar biochemical characteristics compared to skin samples from the retroauricular crease.
The cross-validation studies discussed in previous sections permitted evaluation of the predictability of different disease states from the human microbiome. However, they are not necessarily a good proxy to evaluate the generalization of the prediction model to independent validation samples, a scenario more relevant to a clinical setting but that has been scarcely investigated. Specifically, how do prediction models perform when applied to samples generated in an independent clinical and laboratory study? We address this question for several problems of increasing complexity (denoted as CStaV in Fig 1).
We first considered the cirrhosis dataset, in which the samples were acquired in two distinct stages named “discovery” and “validation” (Fig 5). The generalization of the model was evaluated by (i) generating the model on the samples of the training (TR) stage and (ii) applying it on the test (TS) stage. For comparison, we also report the cross-validation results obtained on each specific stage. In general, we found that the model was transferred properly from one stage to the other. In fact, RF applied on species abundance produced an AUC value on the discovery stage that was only slightly decreased from 0.936 (for cross-validation) to 0.919. For the validation stage we actually obtained an increase from 0.958 (for cross-validation) to 0.972, and the marker-based predictions achieved slightly better but overall consistent values (Fig 5). Finally, we note that the AUC achieved on each specific stage were in line with the AUC exhibited by cross-validation using the entire set of samples (0.945).
A similar analysis was done on the T2D dataset, in which samples were collected in two different stages (stageI and stageII, Fig 6). We verified sufficient generalization of the model across the two stages, although we observed a decrease in accuracy relative to cross-validation. AUC for RF on species abundance decreased from a cross-validation value of 0.737 (0.735 for marker presence) to 0.661 (0.639) for stageI, and from 0.743 (0.771) to 0.686 (0.672) for stageII. In general, the results obtained on the cirrhosis and T2D datasets provide reasonably good generalization of the model when applied across disease stages, i.e., to independent samples/batches from the same study. This implies that the samples, although associated with different subjects and acquired at different time points, share common characteristics such the population of study, sample collection approach, DNA extraction protocol, sequencing technology, and analysis strategy [19].
Cross-study validation (denoted as CSV in Fig 1) is a more difficult standard of validation than cross-stage validation, in that training and validation are performed in completely independent studies targeting the same disease. We focused on type-2 diabetes, for which two distinct datasets are available (i.e., T2D and WT2D). The two datasets presented very different population characteristics as T2D targeted Chinese subjects while WT2D enrolled European women. Still, we observed generalization from one study to the other, (Fig 6), although cohort effects clearly affected the results. For validation on the T2D dataset, the AUC for RF on species abundance decreased from a cross-validation value of 0.744 (0.747 for marker presence) to 0.569 (0.566) when the model was constructed on the WT2D dataset. Similarly, for validation on the WT2D dataset, AUC decreased from a cross-validation value of 0.762 (0.739) to 0.664 (0.622). Different results were achieved by transferring the model to WT2D from the two different experimental stages of the T2D dataset. We obtained an AUC of 0.585 (0.595) and 0.689 (0.637) by transferring the model from T2D_stageI and T2D_stageII, respectively, indicating that T2D_stageII was more similar than T2D_stageI to WT2D. This similarity was consistent with integrative correlation [50] between the feature relative importance scores obtained on the considered stage of T2D and those on WT2D (S2 Table). The features of T2D_stageII were more correlated to WT2D than were T2D_stageI features, in agreement with the prediction accuracies.
Cross-study validation of T2D classification was improved by adding gut microbiome samples from the healthy subjects of four other datasets, i.e., cirrhosis, colorectal, HMP, and IBD, to the training data. While we included all the control groups as “healthy”, there is the potential for health problems among some control subjects. However, it is standard practice in case-control studies to exclude known disease conditions from control groups, so we can assume that, even in the worst case, just a few diseased patients may be included in the controls and these may be mostly due to undiagnosed cases. In this setting we tested the generalization of the model across cohorts (Fig 7A) by generating the models on all the available samples apart those associated with the dataset considered for testing, a "leave-one-dataset-out" cross-study validation [51] (denoted as lodoCSV in Fig 1). Interestingly, we obtained improved discrimination for T2D when control samples from multiple independent studies were added to training sets, with a high cross-validation AUC score in predicting type-2 diabetes on the entire set of samples (0.837/0.806 for species abundance and marker presence using RF, respectively). These values were in fact higher than the AUC obtained by merging all the T2D and WT2D samples into a single set and cross-validating them (0.743/0.736). This cross-validation accuracy was reduced when we tested the generalization of the model to the two T2D datasets (from 0.743/0.736 to 0.655/0.653 and 0.709/0.679 for T2D and WT2D, respectively), which confirmed a non-complete generalization of the model across cohorts. Interestingly, such values obtained by including healthy samples from other cohorts were again higher than for models constructed only on the T2D or WT2D datasets (Fig 6). Thus, including healthy samples from independent cohorts was effective at improving the detection of T2D status. Finally, we evaluated generalization on the healthy samples of the four other datasets (prediction assessed in terms of overall accuracy–OA, right part of Fig 7A). In such cases we verified high accuracy (i.e., OA close to 1 for all the considered datasets), confirming correct prediction for most of the control samples. Addition of independent healthy samples to training sets was also performed for gender prediction (S7 Fig, part b), also resulting in increased accuracy, although the discrimination capabilities remained generally low. Overall, these results strongly suggest that the inclusion of samples of healthy individuals from unrelated cohorts is beneficial in disease-targeted investigations, especially when the prediction task has to be generalized to new cohorts.
We compared the cross-validation accuracies that we obtained (Fig 3) with results reported in the original papers, when available. For cirrhosis, our best AUC value was 0.963, higher than the cross-validation result reported in [33] (AUC = 0.838). Slight improvements were also verified for colorectal cancer (AUC = 0.881 against the 0.84 reported in [34]). The best AUC for discrimination of IBD patients in the IBD dataset was 0.914, while a similar analysis was not performed in the original paper [35].
For the other datasets (i.e., obesity, T2D, and WT2D), the original works used a two-step procedure that tends to overstate discrimination accuracy: i) first a statistical test was applied on the entire set of samples to select the most discriminative features, then ii) the model was generated on this set of features and the prediction accuracies were estimated directly on the training set or through a cross-validation approach. This approach overestimates accuracy metrics such as AUC because supervised feature selection is applied on the same data used to evaluate the model, a problem referred by the machine learning community as overfitting [52]. When we adopted the same overfitting-prone procedure, our cross-validation accuracy estimates (especially using marker features) were higher than the original ones for all datasets (S9 Fig), but as discussed these are overestimations of the actual discriminative power of the models.
Conversely, the overfitting-prone method resulted in much worse performance when the model was transferred to different cohorts. For example, the results reported in [32] showed an AUC equal to 0.83 when cross-validating on WT2D, which decreased significantly to 0.66 when the model was transferred from the T2D dataset. For the same dataset, we estimated an AUC of 0.785 (Fig 3) and 0.701 (Fig 6) by (non-overfitted) cross-validation and cross-study validation, respectively. Non-overfitted models in general exhibit cross-validation accuracies that are lower, but better represent the ultimate goal of generalization of the model to independent cohorts.
We stress that the use of a strict, complete cross-validation/cross-study validation approach is necessary in metagenomics. For cross-validation this requires that for each fold, all training steps (including feature selection, model selection, and model construction) are applied on a set of samples that are not overlapping with the samples used for model evaluation/testing. This, together with reducing confounding factors such as antibiotic usage, is necessary for non-overfitted and non-overestimated assessment of the prediction capabilities of metagenomic data.
We finally tested the hypothesis that the distinction between the “healthy” and disease-associated gut microbiome can be generalized to diseases for which training information is not available. For this purpose, we considered all gut samples from the disease-associated datasets for a total of 903 samples (Fig 7B). Here the class “diseased” included patients affected by the set of disparate diseases discussed above. The cross-validation analysis on the entire set of samples exhibited satisfactory results (AUC = 0.821 for the best model; most discriminative features are reported in S10 Fig, part a, although in this scenario the model may in reality classify each type of disease separately from the others. More interesting are the results of cross-study and cross-disease prediction. In such cases the disease associated with the testing cohort was not present in the datasets used to generate the model. Although the obtained AUC were lower than the disease-specific cross-validation results reported previously in Fig 3, we still verified in all cases a certain level of generalization of the model. In particular, the AUC varied between 0.628 (for T2D) and 0.872 (for cirrhosis). This represents an intriguing result that can be associated to the task of modelling the features of the “healthy” microbiome for use as a dysbiosis prediction model for syndromes where few or no training samples are available. As expected, several disease-specific species such as G. morbillorum, B. bifidum and P. micra were not among the most discriminative, the diseases with which they are correlated are not in the training set (S10 Fig, part b). Conversely, species discriminative for multiple diseases (S. salivarius, S. anginosus, V. parvula, R. intestinalis, and C. comes) are even more relevant here, confirming that these species are associated to a general non-healthy microbiome state rather than to specific host conditions (especially S. anginosus, which is the most relevant feature for four of the five testing datasets). Overall, this suggests that the dysbiosis-associated microbiome is partially distinct from the healthy microbiome regardless of the specific disease under investigation. This also confirms that study-specific confounding factors [44, 45] are only partially affecting the estimation of the classification performance. These species identified as associated to general microbiome dysbiosis should be considered in future microbiome studies as non-specific responses to dysbiosis rather than as organisms directly involved in the pathogenesis of the disease under study.
We uniformly processed shotgun metagenomic microbiome data for 2424 samples from 8 studies of 6 disease types, and used cross-validation, cross-study validation, and cross-disease validation to evaluate the accuracy of candidate methods of predictive modelling of disease states. We make recommendations of best approaches and non-overfitted practices for using the microbiome as a prediction tool and discuss species and strain-level biomarkers we identified for single and combined datasets. While in this manuscript we focused on taxonomic information, metagenomic functional data such as gene or gene-family abundance data [53] can be exploited in a similar way to conduct a more advanced function-based analysis. Future work will be devoted to exploring more advanced machine learning strategies to further improve classification performance.
In general, cross-validation revealed good prediction capabilities, however classification results varied considerably between prediction tasks. In some cases, the ability to predict disease in undiagnosed cases may be overestimated due to the presence of confounding factors such as active antibiotics treatment. The influence of confounding factors on human microbiome has been scarcely investigated in the literature [54, 55], but recent studies [44, 45] highlight this problem and question the study design of some works.
Cross-study validation involved evaluating the transferability of prediction models between completely independent patient cohorts. We verified generalization across studies, although transferred models were in some cases less accurate than models tested by within-study cross-validation. Interestingly, the inclusion of healthy (control) samples from independent cohorts in training sets was effective for improving the transferability of predictions. We emphasize that considering cross-study performance, instead of the more traditional cross-validation approach, is necessary to understanding prediction capabilities from metagenomic data [56]. Furthermore, avoiding overfitting is crucial for transferring models between different cohorts. Finally, we obtained promising results in the ambitious task of modelling the features of the “healthy” microbiome for use as a dysbiosis prediction model for syndromes where few or no training samples are available. Importantly, this setting is not affected by confounding factors on the target dataset since target samples are not used to build the model. The identified biomarkers for the “healthy” versus “dysbiosis”-associated microbiome are also very important for future microbiome studies of new diseases, because if the same biomarkers are appearing as discriminatory they should be regarded as general dysbiotic organisms rather than microbes directly in involved in the disease under investigation.
Compared to the considerable amount of work done on learning methods for 16S rRNA studies, our contribution emphasized two strengths unique to the shotgun metagenomic approach. First, we showed that improved performance can be achieved using strain-level genomic features (i.e., markers) that are not available from 16S rRNA studies. This is also a confirmation that many disease phenotypes are likely linked with microbial genes and factors that are not “core” components of microbial species, but rather are encoded in variable genomic portions that are strain- or subspecies-specific. Second, despite the common potential biases in DNA extraction, shotgun sequencing is considered more consistent across studies than 16S rRNA sequencing for which different variable regions and primer choices are available [57–59], and thus quantitative microbial signature are inherently less difficult to transfer across cohorts and populations. From a more technical viewpoint, learning analysis in shotgun metagenomes presents distinct challenges due to the very high dimensionality of the dataset when considering strain-level markers (~100K features), requiring different considerations in machine learning than for 16S rRNA datasets. Altogether, we provide the first validated toolbox for disease prediction across studies using shotgun metagenomics.
This study provides a publicly available software framework and uniformly processed microbiome profiles for thousands of samples, to facilitate follow-up studies and evaluation of new methods for classification of disease and other states using metagenomic data. This tool allowed us to assess the predictive power of the microbiome features with respect to disease states and transferability across independent datasets. On a final note, we notice that meta-analyses like the one we performed here were recently regarded as “research parasitism” [60], because we analyse data produced by other laboratories. The analysis of cross-study predictions, and identification of a dysbiotic microbiome, would not be possible any other way. We hope that not only will these results inform future clinical microbiome studies of disease, but that they promote data transparency and re-use as key components of scientific progress.
We developed a computational tool for metagenomics-based prediction tasks based on machine learning classifiers (i.e., support vector machines (SVMs), random forests (RFs), Lasso, and Elastic Net (ENet)). The tool uses as features quantitative microbiome profiles including species-level relative abundances and presence of strain-specific markers. The framework is fully automatic, including model and feature selection, permitting a systematic and non-overfitted analysis of large metagenomic datasets. Two main kinds of analysis are implemented, i.e., cross-validation (to evaluate the prediction strength of metagenomic data) and cross-study (to evaluate the generalization of the model between different studies). Additionally, the most relevant features are detected for biomarker discovery tasks. Finally, a set of tools is provided to evaluate classification performances in different ways including i) evaluation metrics such as overall accuracy (OA), precision, recall, F1, and area under the curve (AUC); ii) receiver operating characteristic (ROC) curve plots; iii) confusion matrices; iv) plots of the most relevant features in addition to average relative abundances; and v) heatmap figures.
The MetAML (Metagenomic prediction Analysis based on Machine Learning) tool is open-source and available online at http://segatalab.cibio.unitn.it/tools/metaml. All the species-level taxonomic profiling and marker presence and absence data generated by MetaPhlAn2 and used in this paper are available at the same address.
The developed tool incorporates four classification approaches (i.e. SVM, RF, Lasso, and ENet) which have been extensively applied in many different fields including computational biology and genomics [18]. The classifiers were implemented using the scikit-learn python package [61].
SVMs aim at finding the hyperplane that maximizes the margin between the samples in different classes [38], a strategy with many theoretical and practical advantages [62]. Although they are intrinsically linear, they can be extended to the non-linear case by mapping data into a higher dimensional feature space by means of a kernel function. In this work, a radial basis function (RBF) kernel was considered for classifying species abundances, while a linear kernel was adopted for markers due to the sparsity of marker-based profiling. In both cases, the best regularization parameter C (both for linear and RBF kernel) and the width parameter γ (only for RBF kernel) were chosen in {2−5, 2−3, …, 215} and {2−15, 2−13, …, 23}, respectively, using a 5-fold stratified cross-validation approach. In cross-validation, samples are first randomly subdivided into k subsets (folds) of equal size. In particular, we use here stratified cross-validation, in which folds are made to preserve the percentage of samples of each class. A single subset is then used for the testing the model, and the remaining k−1 subsets are used for training. The whole process is repeated k times, with each of the k subsets used once as the testing set. Finally, the results on the k testing folds are averaged to produce a single accuracy evaluation. The parameters that maximize the accuracy (or another metric of choice) are finally chosen. SVMs are binary classifiers and, in this work, extension to multi-class classification problems was obtained through the one-against-one approach [63]. Moreover, class posterior probabilities of each sample were estimated from the predicted labels in the binary case using the Platt formulation [64], which, in the multi-class case, was extended as per [65].
RFs are an ensemble learning method which constructs a large number of decision trees at training time and outputs the class that is the mode of the classes of the individual trees [39]. The free parameters of such classifier were set in this work as follows: i) the number of trees was equal to 500; ii) the number of features to consider when looking for the best split was equal to the root of the number of original features; iii) the quality of a split was measured using the gini impurity criterion. Although a better estimation of such parameters may be obtained through cross-validation, no significant variations were verified by empirical evaluation. We note that RFs can intrinsically deal with binary and multi-class classification problems and give estimation of class probabilities. Moreover, they implicitly provides a list of the features sorted in terms of relative importance. Feature importance was computed in our case using the strategy usually referred to as “gini importance” or “mean decrease impurity” [66]. These importance values were exploited to perform an embedded feature selection strategy (denoted as RF-FS:Emb) implemented as follows: i) RF was applied on the whole set of available features; ii) features were ranked in terms of importance; iii) RF was re-trained on the top k-th features, by varying k in the set {5, 10, 20, 30, 40, 50, 60, 70, 80, 90, 100, 125, 150, 175, 200}; iv) the number of features that maximized the accuracy was chosen as the optimal number; v) the final model was generated by training RF on this reduced set of features.
Lasso [40] and ENet [41] are generalized linear modelling approaches that incorporate feature selection and regularization to increase prediction accuracy from high-dimensional and collinear predictors. Lasso is based on a multiple logistic regression trained with L1-norm penalized likelihood, while both L1 and L2 norms are penalized in ENet. In this work, we exploited them in two main ways: i) directly applying Lasso or Enet as pure classifiers by training a regression model on the binary classification problem; and ii) using Lasso or ENet for feature selection and then applying SVM or RF on the selected features. In both cases, best regularization parameters were estimated using a 5-fold stratified cross-validation approach. For Lasso this implied to chose the alpha parameter in {10−4, …, 10−0.5} with values evenly spaced on a logarithmic scale. For ENet, along with alpha the L1_ratio parameter was chosen in [0.1, 0.5, 0.7, 0.9, 0.95, 0.99, 1.0].
Two main kinds of analysis were performed in this work, i.e., cross-validation and cross-stages/studies. For cross-validation studies, prediction accuracies were assessed by 10-fold cross validation, repeated and averaged on 20 independent runs. We underline that model selection and feature selection are done using only the training set thus avoiding overfitting problems. In the cross-stages/studies case, all the samples of the first stage/study are considered for training and thus used to generate the classification model including the model selection and feature selection steps. The generalization of the model is evaluated by applying it on the samples of the independent stage/study. In all the cases, the results obtained on the original classification problem were compared with those obtained by a random classifier (denoted in the paper as SVM-Shuffled and RF-Shuffled). For such purpose, we applied the same setting after shuffling randomly the labels of all the samples.
Several different metrics were taken into account to evaluate classification performances. First, we considered the OA, which is the percentage of correctly predicted samples. From the confusion matrix three main metrics were computed: i) the precision (i.e., the number of correct positive samples divided by the number of samples predicted as positive); ii) the recall (i.e., the number of correct positive samples divided by the total number of positive samples); iii) the F1 score, which is the harmonic mean of precision and recall, i.e., F1 = 2*(precision*recall)/(precision+recall). These three metrics (which range in [0, 1], where 1 indicates the best case) can be computed for each class separately. For brevity, we report in the paper only the average values: after calculating the metrics for each class, their average values, weighted by the number of samples per class, are computed. For binary classification problems, class posterior probabilities were used to plot the ROC curve, which represents the true positive rate (i.e., the recall) against the false positive rate (i.e., the number of wrong positive samples divided by the total number of non-positive samples). From the ROC curve, we computed the widely-used AUC statistic, which can be interpreted as the probability that the classifier ranks a randomly chosen positive sample higher than a randomly chosen negative one, assuming that positive ranks higher than negative. The AUC ranges in [0.5, 1], where 0.5 corresponds to random change.
In the comparison among classifiers, prediction accuracy was assessed by 10-fold cross-validation, repeated and averaged on 20 independent runs. The same folds were used for all classifiers, i.e. training and test sets were identical for each classifier. In this way, the difference in performance of two classifiers could be calculated directly as the difference in AUCs (or any other metric) within each test fold. Mean difference and standard error were calculated for each 10-fold CV, then averaged across the 20 repetitions for smoothing. 95% confidence intervals on the difference in AUC performance of two classifiers were calculated using the t-distribution with df = 9, i.e.:
95%CI:120110∑j=120∑i=110(AUC1ij−AUC2ij)±2.26×σj10
where AUC1ij and AUC2ij are the AUC of two classifiers in fold i of repetition j, and σj is the standard deviation of the AUC1ij−AUC2ij across i = 1…10 folds in repetition j. Similarly, p-values were obtained from the t-statistic obtained with mean difference and standard error smoothed over the 20 repetitions:
t=110120∑j=120∑i=110(AUC1ij−AUC2ij)120∑j=120σj10
using two-tailed t-test with df = 9, noting that the AUC differences were approximately normally distributed.
In terms of feature selection, we reported the list of the 25 most important features found by RFs. For each feature, we considered also the relative importance score, which is a real number in the range [0, 1] with features that sum to 1. Feature selection is done for each run independently, and we report the average results.
We initially considered a total of 2571 publicly available metagenomic samples (from eight main studies/datasets) that were reduced to 2424 after pre-processing and curation (see next sections). These are all the human-associated shotgun metagenomic studies with more than 70 samples and read length bigger than 70nt available as of January 2015. Six studies were devoted to the characterization of the human gut microbiome in presence of different diseases. Cirrhosis included 123 patients affected by liver cirrhosis and 114 healthy controls [33]. Colorectal consisted of a total of 156 samples, 53 of which were affected by colorectal cancer [34]. IBD represented the first available large metagenomic dataset and includes 124 individuals, 25 were affected by inflammatory bowel disease (IBD) [35]. Obesity included 123 non-obese and 169 obese individuals [31]. Two distinct studies were instead related to the alteration of the microbiome in subjects with type 2 diabetes (T2D). In the T2D dataset, 170 Chinese T2D patients and 174 non-diabetic controls were present [37]. The WT2D focused on European women and included 53 T2D patients, 49 impaired glucose tolerance individuals and 43 normal glucose tolerance people [32]. Among these six datasets, two of them comprise two independent stages. For cirrhosis, 181 and 56 samples were collected during the so defined discovery and validation phases, respectively. Similarly, for T2D, 145 and 199 samples were acquired during the first (stageI) and second (stageII) stages, respectively. Additionally, two studies focused on healthy subjects and not strictly related to the gut microbiome were also taken into account. HMP included samples collected from five major body sites (i.e., gastrointestinal tract, nasal cavity, oral cavity, skin, and urogenital tract). A subset of these samples were described in [1]. Finally, skin was composed by 291 samples acquired from several different skin sites [36].
The entire analysis was done by taking into account two types of features: species-level relative abundances and presence of strain-specific markers. These features were extracted from the metagenomic samples using MetaPhlAn2 [21] with default parameters. Species abundances are real numbers in the range [0, 1] that sum up to 1 within each sample, while markers assume binary values. Species abundance and marker presence profiles are characterized by very different numbers of features: in the hundreds for species abundance, and hundreds of thousands for markers (the exact numbers of features for each dataset are detailed in Fig 2). Before applying MetaPhlAn2 the samples were subject to standard pre-processing as described in the SOP of the Human Microbiome Project [1] without however the step of human DNA removal as these publicly available metagenomes were deposited free from human DNA contamination. Additionally, we removed reads with length less than 90 nucleotides. For the IBD and obesity datasets the minimum length was set to 70 and 75, respectively, as these cohorts were sequenced with shorter read-lengths. Few samples did not pass the minimum length requirement and were thus discarded.
The experimental evaluation can be summarized into five main steps: 1) cross-validation analysis was done on the six disease-association datasets for evaluating the capabilities of metagenomic data for disease classification; 2) cross-stage studies were performed on the cirrhosis and T2D datasets in order to test the generalization of the model on independent collection batches from the same study; 3) in terms of T2D, the analysis was extended by taking into account also samples from completely distinct cohorts; 4) cross-studies were also done to model the features of the “healthy” gut microbiome for use as a dysbiosis prediction model for syndromes where few or no training samples are available; 5) cross-validation and cross-study analysis were applied to deal with different classification problem such as gender and body site discrimination. We note that all the investigated classification problems, excluding the body site discrimination, represented binary classification problems. Moreover, most of the analysis was done in terms of disease classification, in which the objective was to discriminate between “healthy” and “diseased” subjects.
The MetAML (Metagenomic prediction Analysis based on Machine Learning) software is open-source, written in Python and available online at http://segatalab.cibio.unitn.it/tools/metaml together with all the data used and discussed in this work.
|
10.1371/journal.pntd.0005910 | Induction of allopurinol resistance in Leishmania infantum isolated from dogs | Resistance to allopurinol in zoonotic canine leishmaniasis has been recently shown to be associated with disease relapse in naturally-infected dogs. However, information regarding the formation of resistance and its dynamics is lacking. This study describes the successful in-vitro induction of allopurinol resistance in Leishmania infantum cultured under increasing drug pressure. Allopurinol susceptibility and growth rate of induced parasites were monitored over 23 weeks and parasite clones were tested at selected time points and compared to their parental lines, both as promastigotes and as amastigotes. Allopurinol resistance was formed in strains from two parasite stocks producing a 20-fold rise in IC50 along three distinct growth phases. In addition, characteristic differential clustering of single nucleotide polymorphisms (SNP) was found in drug sensitive and resistant parasite clones. Results confirm that genetic polymorphism, as well as clonal heterogeneity, contribute to in-vitro resistance to allopurinol, which is likely to occur in natural infection.
| Visceral leishmaniasis caused by the parasite Leishmania infantum is a neglected tropical disease transmitted from animal hosts to humans by sand fly bites. This potentially fatal disease affects thousands of people annually and threatens millions who live in disease risk areas. Domestic dogs are considered as the main reservoir of this parasite which can also cause a severe chronic canine disease. Allopurinol is the main drug used for long term treatment of this disease but it often does not eliminate infection in dogs. We have recently demonstrated that allopurinol resistant parasites can be isolated from naturally infected dogs that have developed clinical recurrence of disease during allopurinol treatment. In this study we aimed to see if resistance can be induced in susceptible parasite strains isolated from sick dogs by growing them in increasing drug concentrations under laboratory conditions. The changes in allopurinol susceptibility were measured and the impact of drug on parasite growth was monitored over 23 weeks. Induction of resistance was successful producing parasites 20-folds less susceptible to the drug. The pattern of change in drug susceptibility suggests that a genetic change is responsible for the increased resistance which is likely to mimic the formation of resistance in dogs.
| Visceral leishmaniasis caused by Leishmania infantum is a life threatening disease, affecting humans in Europe, Asia, North Africa and Latin America, as well as domestic dogs which are the main reservoir for this infection [1, 2]. We recently reported the detection of disease relapse in infected dogs associated with allopurinol resistant parasite strains [3]. Allopurinol is the main drug used for long-term control of the canine disease, and since resistant parasites may enhance transmission to humans and other dogs [4], this finding is alarming. In this study, we aimed to improve our understanding of the formation of resistance to allopurinol by following an in-vitro model of resistance induction in susceptible isolates under increasing drug pressure and examining the susceptibility to allopurinol of several clones from the same time point, in both promastigote and amastigote stages.
Two allopurinol susceptible L. infantum isolates, obtained prior to drug treatment from dogs at time of first diagnosis of clinical disease, were used in the study; MCAN/IL/2011/NT4 and MCAN/IL/2011/NT5. Both dogs were males; presented weight loss, skin lesions, enlarged lymph nodes, mild anemia and elevated serum globulin levels. Dog NT4 was also azotemic. Isolation, culture procedure and IC50 testing were done as previously described [3]. Briefly, allopurinol susceptibility was determined using a promastigote viability test, following 72 h incubation in increasing drug concentrations. Each test was repeated twice.
Resistance was induced in cultures designated NT4.L and NT5.L in a stepwise manner; beginning with the original isolates and every 2–6 days thereafter, 5*106 promastigotes were transferred into 5 mL of complete M-199 medium containing increasing allopurinol concentrations, starting at 100 μg/mL with 50 μg/mL increments. Thus, each step was defined as the period between two successive subcultures. Average growth rate for each step was calculated as the increase in parasite concentration during the step divided by its length in days (average step growth rate–ASGR). Once parasites were able to grow at 900 μg/mL allopurinol, they were maintained in it for at least two additional months. Allopurinol IC50 of cultures was tested every 7–14 days and culture samples were cryopreserved. Controls of each isolate cultured without allopurinol, designated NT4 and NT5, were maintained and tested in parallel. All culture medium components manufacturers and lot numbers were kept constant for the duration of the experiments.
Drug induced cultures from selected time points were thawed and single clones were isolated using the hanging drop method, adapted from Evans and Smith [5]. Briefly, 0.5 μL samples were taken of each culture adjusted to contain 2*103 parasites per mL. Samples were inspected microscopically, and those containing individual promastigotes were subcultured in 200 μL of culture medium until a stable clonal culture was established. IC50 was established for each revived frozen culture, as well as for 5–10 of its clones. Allopurinol susceptibility was also studied in intracellular amastigotes developed for each induced strain and its clones at one time point, as previously described [3]. Briefly, DH-82 cells were infected with promastigotes from thawed samples of the drug induced strains and 5 respective clonal cultures. Infected cells were treated with either 0 or 300 μg/mL allopurinol for 72h, followed by counting of intracellular parasites per 100 DH-82 cells on Giemsa stained preparations, and calculation of the percent inhibition caused by the drug. Drug-free control cultures were also thawed and tested at the specific time points.
Six clonal strains derived from cultures NT4.L and NT5.L as described above, presenting low (n = 2) and high (n = 4) allopurinol IC50 values were chosen for whole genome sequencing (WGS). These included clone 1 of NT4.L on day 28 (NT4.L.s, see S1 Table) and clones 3 and 4 from day 104 (NT4.L.r1 and NT4.L.r2, respectively); clone 2 of NT5.L from day 28 (NT5.L.s) and clones 1 and 5 of day 86 (NT5.L.r1 and NT5.L.r2, respectively).
DNA for WGS was extracted from 2*108 mid log-phase promastigotes of each of the six strains described above. Following centrifugation at 1500 rpm for 10 minutes, supernatant was discarded and promastigotes were suspended in 250μL phosphate buffered saline. DNA was then extracted using the Illustra blood genomicPrep Mini Spin KIT (GE Healthcare, UK) according to manufacturer’s instructions and included RNAse treatment (RNAse A, Sigma-Aldrich, St. Louis, MO). Quantity and quality of DNA was tested using the NanoDrop 2000 (Thermo Scientific, Wilmington, DE), followed by visualization in 1% agarose gel with ethidium bromide. Fragmentation was done using the Covaris shearing (Covaris S2, Covaris, Woburn, MA) set to the size target at 400bp for library preparation. Libraries were made using the TruSeq DNA kit (Genomic DNA Sample Prep Kit, FC-102-1004, Illumina, San Diego, CA).
Sequencing was done using 100 bases paired ends reads, on an Ilumina HiSeq2000 platform, with the TruSeq SBS Kit (TruSeq SBS v3, FC-401-3001, Illumina, San Diego, CA) and TruSeq PE Cluster Kit (TruSeq PE Cluster Kit v3, PE-401-3001, Illumina, San Diego, CA), at the DNA LandMarks Laboratory (St.-Jean-sur-Richelieu, Canada).
Raw reads were subjected to a cleaning procedure using the FASTX Toolkit (http://hannonlab.cshl.edu/fastx_toolkit/index.html, version 0.0.13.2). Trimming read end nucleotides with quality scores under 30 was done using the fastq_quality_trimmer, and removal of read pairs was done if reads in the read pair had less than 70% base pairs with quality score under or equal to 23, using the fastq_quality_filter.
The L. infantum JPCM5 genome with chromosomes 1–36 was used as a reference genome (European nucleotide archive, BioProject PRJNA12658, FR796433—FR796468). Cleaned paired-end reads, obtained after processing and cleaning of the 6 samples were mapped to the reference genome using the Bowtie2 program version 2.0.0 with default parameters [6]. SNP analysis was done using the Picard (http://broadinstitute.github.io/picard/) and GATK UnifiedGenotyper (version 2.5–2) [7] programs. The SNP’s calling was done in reference to the L. infantum JPCM5 genome. The degree of similarity between SNP of the six induced resistant strains was studied by a maximum likelihood analysis with bootstrapping (N = 100), using the PhyML 3.0 software [8].
Comparing IC50 values between induced and control cultures for matching time points was done using the t-test. Tukey HSD and Wilcoxon tests were used to compare IC50 values within each drug-cultured isolate at different time points. The Tukey HSD test was used to compare IC50 and percent inhibition values within strains and respective clones on each time point. Correlations between promastigote IC50 values and amastigote percent inhibition values for respective clones were described for NT4.L and NT5.L using a linear, logarithmic or polynomial trend lines.
The animal care protocol used in this study was approved by the Hebrew University’s Institutional Animal Care and Use Committee (IACUC); approval no. MD-08-11476-2, following the USA NIH guidelines.
A starting allopurinol concentration of 100 μg/mL was chosen because the IC50 values of the parent isolates were 105 and 93 μg/mL for NT4 and NT5, respectively [3]. During the experiment, drug concentration in the culture medium quadrupled by day 22 and maximal drug level tested was 900 μg/mL allopurinol from days 60–71 and on, about 10 folds higher than the initial IC50 values of NT4.L and NT5.L, and close to the average IC50 value found for resistant clinical isolates previously [3] (Fig 1A and 1B).
Allopurinol susceptibility kinetics showed three distinguishable phases in both isolates. In the initial phase (up to day 28 in NT5.L and day 60 in NT4.L), IC50 values had not changed significantly compared to the initial value, and were comparable to those measured for the controls at most points. In the second phase, a significant 4 folds increase in IC50 was seen, only in drug exposed cultures of both isolates (t-test, P <0.05). Peak IC50 values recorded were 2225 μg/mL at day 93 for NT5.L and 2209 μg/mL at day 94 for NT4.L. In the third phase, a decline of over 2 folds in IC50 compared to peak values was measured for both drug-cultured isolates (Tukey HSD and Wilcoxon tests, P <0.05). Control cultures NT4 and NT5, not exposed to drug pressure, fluctuated in IC50 over time, however peak values did not exceed 405 μg/mL.
Using ASGR values, two distinct phases were detected in both induced cultures, marked by a sharp change in ASGR. In the first phase, growth rate was decreased compared to the original isolate, and slower growth lasted approximately to day 39 for NT5.L (allopurinol concentration 550 μg/mL) or day 60 for NT4.L (allopurinol concentration 700 μg/mL). In the second phase, growth rate significantly increased by 2 folds or more (Fig 1C and 1D, t-test, P <0.001). Difference in frequencies of steps length between the two growth phases was tested, and no significant difference was found between the two phases (χ2 test, P = 0.3945), confirming that promastigotes growth rates were indeed increased in phase two for both cultures.
Frozen drug-cultured isolate samples from different time points were revived, clones were isolated and their promastigote IC50 compared with that of the respective parent sample (Fig 2A and 2B).
IC50 values of clones presented a 2–4 fold variation at all time points, with significant differences detected between clones (Tukey HSD test, P <0.05, S1 Table). Interestingly, values of parent cultures were within the 95% confidence intervals created by values of their respective clones. As found for promastigotes, testing of intracellular amastigotes also demonstrated in most cases a significant difference between susceptible control strains and resistant induced strains and clones (Fig 2C, S2 Table). Inter-clonal variation demonstrated in amastigotes was smaller than seen in promastigotes. However, this can be in part due to limitation of the assay that prevents using drug concentrations of over 300μg/mL allopurinol. R square values for correlation between promastigote IC50 values and amastigote percent inhibition values ranged between 0.86–0.97 for NT4.L and 0.7–0.79 for NT5.L, for linear and polynomial model, respectively.
WGS resulted in cleaned paired-end reads of 17–38*106 reads per sample, high assembly rate of 98.13–98.52% and coverage of x99-x218. Maximum likelihood analysis including all SNP’s found (including 9,969 positions, S3 Table) resulted in the resistant and susceptible clonal strains dividing into two distinct clusters (Fig 3).
Leishmania infantum resistance to allopurinol may pose a combined veterinary and public health threat. Drug resistance can result in infected dogs having an uncontrolled high parasite load and being parasitemic for longer periods, increasing both the impact of the disease on the canine host and the potential for transmission via sandflies to humans [9, 10]. Valuable molecular and biochemical information can be obtained by analyzing resistant field isolates. The in-vitro generation of drug resistance is a useful complementary tool for elucidating the mechanisms of resistance formation, especially when a genetic basis is suspected and sought [11–15]. As a first step in exploring resistance to allopurinol we constructed an in-vitro promastigote model that allowed monitoring the progression of resistance development under drug pressure over time, with less variables and complexity compared to an intracellular amastigote based model. This same approach, when applied in studies of antimonials [16] and miltefosine [17], resulted in the identification of genetic changes found also in resistant amastigote strains [18, 19]. In the present study, we applied drug pressure of up to 10 times higher (900μg/mL) than the initial IC50 of the two induced promastigote cultures, a drug level compared to the average IC50 level found previously for resistant clinical isolates (996±372μg/mL) [3]. This experimental setup has induced or selected for a considerable increase in allopurinol resistance, resulting in IC50 levels of up to 20 folds higher than initial level, comparable to levels measured for allopurinol-resistant parasites isolated from dogs that experienced clinical disease relapse [3]. Three distinct stages were discerned by monitoring growth rates and IC50 values during the induction of resistance. Initially, following introduction of drug pressure parasite growth rates decreased. This was due either to adaptation to the culture medium or the additional stress put on by the drug. Although an increase in IC50 values accompanied the decreased growth rate, it was found to be non-significant and shared by both control and test cultures. Therefore, this increase may also represent an adaptation to the culture medium and to the purine sources in particular, affecting the uptake or metabolism of the purine analog allopurinol [20–22]. Following this period of adaptation the growth rates of the isolates cultured with drug at least doubled in parallel to a significant increase in their IC50. Since the peak IC50 values were measured slightly after maximum drug concentration was reached, this IC50 increase was most likely due to genetic adaptations, where the drug pressure selected variants that carried advantageous polymorphisms [15, 23, 24]. The existence of an inherent basis for resistance is supported also by the relatively high correlation found between drug susceptibilities of promastigotes and amastigotes of respective strains. The dynamics of resistance formation seen here fits the suggested model for appearance of pathogen drug resistance in infectious diseases following treatment, which includes emergence, establishment, increase and equilibrium of mutations promoting growth and survival [25]. Noteworthy is the significant 2–3 folds decline in IC50 values following the peak values seen in both NT4.L and NT5.L. This decline occurred when the drug concentration in the medium was maintained constant at maximal levels, leaving IC50 values still significantly higher compared to most time points during the adaptation phase. This phenomenon might reflect the result of intra-clonal mechanisms such as negative sign epistasis between mutations or may be caused by inter-clonal interaction in mixed cultures in response to prolonged exposure to high drug concentrations, such as clonal interference [25, 26]. In support of the latter explanation, the IC50 values of clones generated from a culture under drug pressure at individual time points revealed significant heterogeneity, both when promastigotes or amastigotes were tested. As a rule, clonality is well recognized in Leishmania and was suggested to play a role in its evolution [15]. Albeit the heterogeneity in IC50 values within cultured populations, both cultures (NT4.L, NT5.L) demonstrated very similar patterns during the development of allopurinol resistance, both with respect to IC50 values and growth rates. This suggests that the process of adaptation to drug pressure may have been similar in both independent cultures.
In conclusion, this study describes the successful induction of allopurinol resistance in L. infantum, under drug pressure. The model may facilitate studies on the mechanisms, pathways and genetics of allopurinol resistance in parasite populations, as well as identification and monitoring of resistance in clinical isolates.
|
10.1371/journal.pgen.1002733 | Stimulation of Host Immune Defenses by a Small Molecule Protects C. elegans from Bacterial Infection | The nematode Caenorhabditis elegans offers currently untapped potential for carrying out high-throughput, live-animal screens of low molecular weight compound libraries to identify molecules that target a variety of cellular processes. We previously used a bacterial infection assay in C. elegans to identify 119 compounds that affect host-microbe interactions among 37,214 tested. Here we show that one of these small molecules, RPW-24, protects C. elegans from bacterial infection by stimulating the host immune response of the nematode. Using transcriptome profiling, epistasis pathway analyses with C. elegans mutants, and an RNAi screen, we show that RPW-24 promotes resistance to Pseudomonas aeruginosa infection by inducing the transcription of a remarkably small number of C. elegans genes (∼1.3% of all genes) in a manner that partially depends on the evolutionarily-conserved p38 MAP kinase pathway and the transcription factor ATF-7. These data show that the immunostimulatory activity of RPW-24 is required for its efficacy and define a novel C. elegans–based strategy to identify compounds with activity against antibiotic-resistant bacterial pathogens.
| Infections with antibiotic-resistant bacterial pathogens are increasing at an alarming rate, and there are very few new therapies currently being developed. We have identified a small molecule that protects the nematode C. elegans from bacterial infection by stimulating the host immune response, not by directly interfering with bacterial growth, which is the mechanism employed by all currently available antibiotics. In this study, we investigate the mode of action of this small molecule and demonstrate that it stimulates the expression of immune effector molecules via evolutionarily conserved immune regulators. These studies illustrate the potential of targeting host immunity for the treatment of bacterial infections and the advantages that C. elegans offers both to identify small molecules that target key cellular processes and to study their mechanism of action.
| Studies in the model nematode Caenorhabditis elegans have greatly expanded our understanding of development, neurobiology, host-pathogen interactions, and many other aspects of metazoan biology. Here we show that C. elegans-based assays enable the identification of immunostimulatory compounds, which can be employed together with genetic analyses to interrogate innate immune signaling pathways.
Previously, we demonstrated that C. elegans can be used in bacterial infection assays to identify novel antimicrobials [1]–[3]. Fifteen to twenty C. elegans animals fit comfortably in the wells of standard 384-well assay plates and the assay can be automated using image analysis software, which enables such studies to be conducted in high-throughput [1]. Using this system, we tested 37,214 compounds and identified 119 small molecules that prolonged the lifespan of nematodes infected with the Gram-positive human bacterial pathogen Enterococcus faecalis [1].
We hypothesized that the C. elegans-based screen for novel anti-infectives would identify small molecules that cure nematodes by stimulating the host innate immune response, in addition to compounds that block microbial virulence or directly inhibit bacterial growth [1]. In nature, nematodes consume bacteria for food and have evolved sophisticated innate immune mechanisms within their intestinal epithelium to defend against ingested pathogens [4]. C. elegans mount specific immune responses toward both bacterial and fungal pathogens using immune signaling mediators that are strongly conserved throughout evolution [5]–[10]. Principal among these regulators is the NSY-1/SEK-1/PMK-1 Mitogen Activated Protein (MAP) kinase pathway, orthologous to the ASK1 (MAP kinase kinase kinase)/MKK3/6 (MAP kinase kinase)/p38 (MAP kinase) pathway in mammals [10]. In C. elegans, the p38 MAP kinase pathway acts cell autonomously in the intestine [11] to coordinate the expression of immune effectors such as C-type lectins and genes that may encode antimicrobial peptides [8]. Recently, Shivers et al. found that the transcription factor ATF-7, an ortholog of mammalian ATF2/ATF7, is phosphorylated by the p38 MAP kinase PMK-1 and is also required for defense against ingested bacterial pathogens [12].
In this study, we describe a small molecule named RPW-24 that strongly stimulates the innate immune response of C. elegans in a manner that confers a survival advantage for nematodes during bacterial infection. We show that the activity of this compound is partially dependent on the C. elegans p38 MAP kinase cassette and ATF-7. These data demonstrate that C. elegans can be used in facile in vivo screens to identify compounds with desirable biological activities.
In a previous study, we screened 37,214 small molecules for those that prolonged the lifespan of C. elegans infected with the Gram-positive bacterial pathogen E. faecalis as a means to identify novel antimicrobials [1]. Of 119 compounds identified, 31 did not have any structural relationship to known antimicrobials and ten of these small molecules were effective against E. faecalis in the C. elegans infection model at doses that did not inhibit growth of the pathogen in an in vitro growth assay [1]. In contrast, all currently available antibiotics interfere with some aspect of bacterial growth or metabolism. In the C. elegans-based assay, traditional antibiotics, such as tetracycline, ciprofloxacin, ampicillin, and vancomycin, cured E. faecalis-infected nematodes only at doses several fold higher than the in vitro minimum inhibitory concentration (MIC) for bacterial growth [2]. We therefore hypothesized that a subset of these 31 small molecules conferred a survival advantage to infected worms by either stimulating the host immune response of C. elegans or by interfering with virulence factor production in the bacteria.
We reasoned that a small molecule, which demonstrated curing activity against diverse bacteria without directly affecting bacterial growth, would be a candidate immunostimulator. It would be less likely for such a compound to be a specific inhibitor of bacterial virulence determinants given the complex and presumably pathogen-specific nature of these factors. We therefore screened the 31 compounds described above that did not have any structural relationship to known antimicrobials for those that prolonged the lifespan of nematodes infected with Pseudomonas aeruginosa, a Gram-negative pathogen, at doses lower than the in vitro MIC for P. aeruginosa. We found that eight of the 31 small molecules demonstrated in vivo efficacy against nematodes infected with P. aeruginosa (Table 1, Figure 1A and Figure S1) at doses several fold lower than the in vitro MIC for these compounds against P. aeruginosa. Indeed, four of the tested compounds did not affect the growth of P. aeruginosa at any concentration tested (Table 1).
To ask if any of these eight compounds could activate the transcription of a putative immune effector, we used transgenic C. elegans animals carrying a transcriptional GFP reporter for the gene F35E12.5 [13]. F35E12.5 encodes a protein that contains a CUB-like domain and is involved in the transcriptional response towards several bacterial pathogens, including P. aeruginosa [8], [13], [14]. Interestingly, one of the four compounds that exhibited P. aeruginosa curing activity in the C. elegans infection assay without affecting growth of the pathogen (RPW-24, Figure 1A) activated the F35E12.5::GFP reporter in a dose dependent manner (Figure 1B). We therefore decided to focus on RPW-24 and explore its mode of action in more detail.
We found that RPW-24 promoted survival of P. aeruginosa-infected animals in a dose-dependent manner using an agar-based assay, the typical way that C. elegans infection assays are carried out in most laboratories (Figure 1C). C. elegans treated with 7, 35 and 70 µM of RPW-24 (but not 0.7 µM) were significantly resistant to P. aeruginosa infection compared to animals treated with the solvent control (DMSO)(Figure 1C). To determine if RPW-24 directly affects the growth of P. aeruginosa, we monitored in vitro bacterial growth in the presence 70 µM RPW-24 and DMSO. We conducted these experiments in the same media used for the nematode killing assay (Figure 1D) or in standard bacterial culture media (Luria broth, data not shown), but in the absence of C. elegans. In both cases, we found that 70 µM RPW-24 did not affect the growth rate of P. aeruginosa (Figure 1D and data not shown). In the C. elegans- P. aeruginosa pathogenicity assay, we observed that the intestines of the DMSO-treated animals were markedly distended and packed with P. aeruginosa cells 40 hours after infection (Figure S2), consistent with previous observations [9], [15]. In contrast, animals treated with RPW-24 had non-distended intestines, a morphology that was strikingly different from the DMSO controls (Figure S2).
In summary, RPW-24 is efficacious against both Gram-positive [1] and Gram-negative bacteria (Figure 1C) at concentrations that do not inhibit bacterial growth (Figure 1D, Table 1) and activates the transcription of a putative immune effector, F35E12.5 (Figure 1B). RPW-24 treatment also dramatically reduced the intestinal burden of P. aeruginosa 40 hours after infection compared to DMSO controls (Figure S2). These data suggest that RPW-24 affects the host immune response of C. elegans in a manner that confers a survival advantage against infection with diverse bacterial pathogens.
To further study the effects of RPW-24 on C. elegans, we used Affymetrix whole genome GeneChips to generate transcriptome profiles of wild-type nematodes following exposure to either 70 µM RPW-24 or the solvent control DMSO in liquid culture media in the absence of bacterial pathogens for 16 hours at 15°C (Figure 2A). We found that RPW-24 induced a remarkably robust transcriptional response that involved only a small fraction (∼1.3%) of the genes of the C. elegans genome (Figure 2A). 269 genes were upregulated three-fold or greater (P<0.025), 125 of which were induced more than 50-fold during RPW-24 exposure (Table S1). The most highly upregulated gene was expressed more than 3500-fold higher in compound-exposed worms. For confirmation, we used qRT-PCR to analyze 10 genes that exhibited varying degrees of induction and expression levels in the microarray analysis (Figure 2A). We found that the transcriptional changes observed in the transcriptome profiling analysis directly correlated with the values obtained by qRT-PCR for all 10 genes tested (Figure S3). We also observed no difference in C. elegans gene induction whether the RPW-24 exposure occurred in liquid or on solid media (Figure S3). Taken together, these data suggest that RPW-24 strongly induces gene transcription in the absence of pathogen exposure.
Examination of the 269 genes that were induced greater than three-fold revealed that RPW-24 causes induction of many genes previously shown to be involved in the C. elegans transcriptional response to pathogenic bacteria. We found that 70 of the 269 genes induced greater than 3-fold by RPW-24 were also activated during infection with P. aeruginosa (a gene set characterized by Troemel et al. [8])(Figure 2B), which is significantly more than the 3.3 gene overlap expected by chance alone (P<2.7×10−16). Among these induced genes are several gene classes that have been implicated in antimicrobial defenses, such as CUB-like domain-containing genes, ShK-like toxins, C-type lectins and small molecule kinases (Table S1) [8]. This result is particularly interesting considering that RPW-24 is active against P. aeruginosa in the worm infection model but does not affect growth of the pathogen in vitro.
The transcriptome profiling analysis demonstrated that RPW-24 induces the transcription of putative immune effectors while C. elegans animals are feeding on Escherichia coli OP50, a relatively non-pathogenic food source for nematodes. We wondered if RPW-24 would further enhance the induction of these genes when C. elegans is infected with a bacterial pathogen that activates expression of these effectors. We therefore used qRT-PCR to test the level of induction of ten putative immune effectors (Figure 2A) during P. aeruginosa infection in the presence and absence of RPW-24. This panel included five genes that contain a CUB-like domain (C32H11.1, F35E12.5, F08G5.6, C29F3.7, and K08D8.5), two ShK-like toxins (F49F1.6 and C14C6.5), one antibacterial lysozyme (lys-7), one C-type lectin (clec-67) and one metallothionein (mtl-1). Eight of these ten genes were activated by P. aeruginosa in the absence of RPW-24 (Figure 2C). Interestingly, the induction levels of five of these eight genes (C32H11.1, F08G5.6, F35E12.5, C29F3.7, and F49F1.6) were markedly increased in the presence of both RPW-24 and P. aeruginosa (Figure 2C). RPW-24 also caused the induction of one gene that was strongly repressed during P. aeruginosa infection (mtl-1) and another whose transcription was unaffected by P. aeruginosa exposure (lys-7). RPW-24 did not change the level of induction for clec-67, C14C6.5 or K08D8.5 in the presence of P. aeruginosa. Taken together, these data demonstrate that RPW-24 both enhances the expression of a subset of putative immune effectors that are normally activated during P. aeruginosa infection and causes the induction of others that are not activated by P. aeruginosa. We do not, however, think that the induction of any single gene is itself responsible for the efficacy of RPW-24 since others have shown that C. elegans immune effectors likely function redundantly to defend against P. aeruginosa infection [8]. The finding that putative immune effectors are robustly induced by RPW-24 during bacterial infection provides a potential explanation why this compound provides protection against P. aeruginosa-mediated killing.
Because RPW-24 caused the induction of genes normally upregulated during bacterial infection, we reasoned that C. elegans immune gene activation by RPW-24 is important for its anti-infective activity and depends on conserved defense response pathways. Previous studies have identified a requirement for three C. elegans signaling pathways in the defense against P. aeruginosa infection: the p38 MAP kinase signaling cassette [10], the G-protein coupled receptor FSHR-1 [16], and the bZIP transcription factor ZIP-2 [17]. Using loss-of-function mutants, we asked whether the activity of any of these pathways is required for the efficacy of RPW-24 in treating C. elegans infected with P. aeruginosa. Indeed, we found that the magnitude of RPW-24-mediated lifespan prolongation of pmk-1(km25) null mutants infected with P. aeruginosa was significantly attenuated compared to the lifespan extension observed in compound-treated wild-type animals (Figure 3) (P<0.001 for the difference in the lifespan prolongation between pmk-1(km25) and wild-type animals in two biological replicates). In contrast to pmk-1(km25) mutants, fshr-1(ok778) loss-of-function mutants demonstrated significant lifespan extension by RPW-24 compared to pmk-1(km25) animals, despite the fact that fshr-1(ok778) animals are also hypersusceptible to P. aeruginosa infection (Figure 3) (P<0.001 for the difference in the lifespan prolongation between pmk-1(km25) and fshr-1(ok778) in two biological replicates). Likewise, the lifespan of zip-2(tm4067) animals infected with P. aeruginosa was also significantly prolonged by RPW-24, and to a greater degree than pmk-1(km25) (P<0.01 for the difference in two biological replicates). These data demonstrate that the activity of PMK-1, more so than FSHR-1 or ZIP-2, is important for RPW-24 to extend the lifespan of nematodes infected with P. aeruginosa. Morever, the observed reduction in the ability of RPW-24 to extend the lifespan of pmk-1(km25) animals is not likely to be secondary to decreased overall fitness of immunocompromised animals infected with a bacterial pathogen because RPW-24 prolongs the lifespan of fshr-1(ok778) animals, a strain that is also markedly hypersusceptible to P. aeruginosa infection (Figure 3).
The p38 MAP kinase PMK-1 functions as part of a conserved signaling cassette to regulate host innate immune responses, which involves upstream activation by the MAP kinase kinase SEK-1, the MAP kinase kinase kinase NSY-1 and the Toll-Interleukin-1 receptor TIR-1. We tested the ability of RPW-24 to extend the lifespan of nematodes with mutations in each of these genes: sek-1(km4), nsy-1(ag3), and tir-1(qd4). As with the pmk-1(km25) mutants, we found that the curing activity of RPW-24 was attenuated in each of these mutant strains compared to the wild-type control (Figure 3), suggestive of a role for the entire TIR-1/NSY-1/SEK-1/PMK-1 signaling cassette in the activation of defense responses following exposure to RPW-24.
The infection assays described above show that the p38 MAP kinase pathway is required for RPW-24 to prolong the lifespan of C. elegans infected with P. aeruginosa, suggesting that RPW-24 induces the expression of immune effectors regulated by this cascade. Previous work has shown that PMK-1 coordinates the transcription of CUB-like genes, ShK toxins and C-type lectins, gene classes which are also induced by RPW-24 [8]. Indeed, we found that 14 of the 86 genes whose basal expression depends on PMK-1 [8] were induced by exposure to RPW-24, which is 14.6 fold more than expected by chance alone (P = 0.002)(Table S1). To test whether RPW-24 activates immune effectors in a PMK-1-dependent manner, we used qRT-PCR to determine the induction levels of six putative immune effectors in pmk-1(km25) null mutants exposed to DMSO and RPW-24 in the absence of pathogen. Wild-type worms were used as the control. The set of six putative immune effectors (F35E12.5, F08G5.6, clec-67, C32H11.1, F49F1.6 and mtl-1) were chosen from the panel of ten genes described above on the criterion that they were robustly upregulated by RPW-24 (Figure 2A, Figure S3). Figure 4A shows that PMK-1 affects the basal expression levels of the RPW-24-induced genes F35E12.5, F08G5.6, clec-67 and F49F1.6 (defined as the relative expression of the gene in pmk-1(km25) mutants compared to its expression in wild-type animals)(Figure 4A). PMK-1 is also required for the RPW-24-mediated induction of clec-67 and C32H11.1 (defined as the fold difference in gene expression in the presence and absence of RPW-24 in wild-type or pmk-1(km25) mutant animals), and perhaps F49F1.6 and mtl-1, although the differences in induction of these later two genes did not reach statistical significance (Figure 4B). Importantly, the absolute expression levels of all six of these genes following RPW-24 exposure were significantly lower in pmk-1(km25) animals than wild-type controls (Figure 4A). Taken together with the P. aeruginosa infection assays in pmk-1(km25) mutants (Figure 3), these data suggest that PMK-1-dependent immune effectors mediate part of the protective effect of RPW-24 in P. aeruginosa-infected C. elegans.
The phenotypic and genetic data presented above show that the p38 MAP kinase pathway is important for the RPW-24-induced modulation of C. elegans immune responses during P. aeruginosa infection. As described above, we found that 70 µM RPW-24 caused a striking increase in GFP production in the F35E12.5::GFP transcriptional reporter (Figure 1B). We thus reasoned that an RNAi screen could be used to find downstream regulator(s) of the RPW-24-induced immune response by identifying the genetic dependence of F35E12.5::GFP activation.
The basal regulation of F35E12.5 requires PMK-1, but its induction by RPW-24 occurs in a PMK-1-independent manner (Figure 4A and 4B). We therefore anticipated that a reverse genetic RNAi screen aimed at identifying transcription factors required for the RPW-24-mediated induction of F35E12.5::GFP would identify genetic regulators that act either downstream of or in parallel to the PMK-1 pathway. We used a feeding RNAi library containing bacterial clones that produce double stranded RNA (dsRNA) designed to individually knockdown the expression of 393 transcription factors in C. elegans, corresponding to 30–50% of the transcription factors in the C. elegans genome [18] and screened for RNAi clones that abrogated the RPW-24-mediated induction of F35E12.5::GFP. Among 393 screened, we found that a single clone, corresponding to the transcription factor ATF-7, caused a striking reduction of F35E12.5::GFP expression when nematodes were either growing on their normal laboratory food source (E. coli OP50) or infected with P. aeruginosa (Figure 5A). ATF-7 was previously shown to function downstream of PMK-1 in the regulation of immune response genes during P. aeruginosa infection [12].
qRT-PCR analysis confirmed that the reduction of F35E12.5::GFP expression was due to the knockdown of ATF-7 and not the consequence of non-specific transgene silencing. Specifically, the absolute level F35E12.5 expression following RPW-24 exposure was markedly reduced in two atf-7(lof) mutants [atf-7(qd137) and atf-7(qd22 qd130) [12]] compared to the control strain (Figure 5B), as were the levels of three other putative immune effectors (F08G5.6, clec-67 and C32H11.1). Moreover, the basal levels of F35E12.5, F08G5.6 and clec-67 were reduced to a similar degree in the atf-7(lof) mutants and the pmk-1(km25) animals, consistent with the previously described role for PMK-1 and ATF-7 in the basal regulation of immune effectors (compare Figure 4A and Figure 5B) [12]. Interestingly, we found that ATF-7 was also required for the full induction of five of six putative immune effectors (F35E12.5, F08G5.6, clec-67, C32H11.1, and F49F1.6)(Figure 5C), including two genes (F35E12.5 and F08G5.6) that were induced by RPW-24 independently of PMK-1 (Figure 4B).
To determine if the activity of ATF-7 is important for the efficacy of RPW-24, we tested the ability of RPW-24 to prolong the lifespan of atf-7(lof) mutants exposed to P. aeruginosa. The magnitude of lifespan extension conferred by RPW-24 was reduced in both the atf-7(qd137) and the atf-7(qd22 qd130)(Figure S4A and S4B, respectively) mutants compared to the control strain. We therefore conclude that RPW-24 stimulates the C. elegans immune response genes in a manner that involves both the p38 MAP kinase cassette PMK-1 and the conserved transcription factor ATF-7, consistent with the placement of ATF-7 downstream of PMK-1 [12]. Expression analysis of RPW-24-induced genes (Figure 5C) suggests that in addition to functioning downstream of PMK-1, ATF-7 receives inputs from a PMK-1 independent pathway to coordinate the induction of putative immune effectors (such as F35E12.5 and F08G5.6) and the RPW-24-mediated resistance to P. aeruginosa infection. This conclusion is based on the finding that the RPW-24-mediated activation of F35E12.5 and F08G5.6 is PMK-1 independent (as opposed to their basal level of expression), but is at least partially dependent on ATF-7 (compare Figure 4B with Figure 5C). That is, the basal levels of expression of F35E12.5 and F08G5.6 are both PMK-1 and ATF-7 dependent, whereas the fold induction of these genes following RPW-24 is not affected in the pmk-1(km25) mutant, but is reduced by at least half in the atf-7(lof) mutants. The biological significance of this PMK-1-independent transcriptional activator is not known.
It is also interesting to note that RPW-24 exhibited an attenuated, but significant (P<0.001) and reproducible ability to rescue the atf-7(lof) mutants. Therefore, the anti-infective activity of RPW-24 may involve another immune signaling pathway, independent of both the p38 MAP kinase pathway and ATF-7. This conclusion is consistent with the observation that RPW-24 also modestly, but significantly (P<0.001), extends the lifespan of pmk-1(km25), sek-1(km4), nsy-1(ag3), and tir-1(qd4) mutant animals infected with P. aeruginosa (Figure 3), and the fact that the induction of F35E12.5 and F08G5.6 is not completely abrogated in the atf-7(lof) mutants. An alternate explanation is that RPW-24 affects virulence factor production by P. aeruginosa or exerts a subtle effect on growth of the pathogen.
In addition to inducing a preponderance of genes involved in the transcriptional response to pathogenic bacteria, RPW-24 caused the upregulation of genes involved in the detoxification of small molecules (Table S1). The microarray analysis revealed that 58 of the 269 genes activated 3 fold or more and 31 of the 57 genes activated 50 fold or more were UDP-glucuronosyltransferases (UDPs), cytochrome P450s (CYPs), glutathione-s-transferases (GSTs) or short-chain dehydrogenases (SDRs). These gene classes play integral roles in the Phase I and II detoxification of both endobiotic and xenobiotic toxins in both nematodes and mammals [19], [20]. Indeed, seven of ten CYPs, two of three UDPs, one of four GSTs, three of three carboxylesterases, and three of six C-type lectins were induced both by RPW-24 and exposure to five xenobiotic toxins [21]. Taken together, these data suggest that RPW-24 induces xenobiotic detoxification pathways in C. elegans.
To determine if RPW-24 adversely affects wild-type nematodes growing in the absence of pathogen, we first used a behavioral assay designed to study the aversion response of C. elegans to xenobiotic toxins [22]. The addition of some poisons to the center of small lawns of non-pathogenic E. coli causes C. elegans animals to leave the lawn, presumably to minimize toxin exposure. Interestingly, we observed a significant aversion response to 70 µM RPW-24 (Figure 6A and 6B). After 16 hours, 51% of the nematodes had left the lawn containing RPW-24, whereas only 7% of animals left a control lawn (P = 0.002; Figure 6B). Next, we conducted a lifespan assay on nematode growth media supplemented with either DMSO or varying concentrations of RPW-24 and found that RPW-24 shortened C. elegans lifespan in a dose-dependent manner (Figure 6C), with 70 µM RPW-24 resulting in a 24% reduction in median lifespan. Interestingly, we observed lifespan shortening only at compound concentrations that rescued C. elegans from P. aeruginosa infection [7, 35 and 70 µM, but not 0.7 µM]. We also found that 70 µM RPW-24 slowed the development of animals when they were exposed at the first larval stage (L1)(Figure 6D).
In C. elegans, oxidative stress is a potent inducer of phase II detoxification genes. Moreover, the p38 MAP kinase PMK-1 regulates the cellular response to oxidative and arsenite stress, but through the transcription factor SKN-1, not ATF-7 [12], [23]–[25]. We therefore wondered whether RPW-24 might confer protection against arsenite stress. We found, however, that 70 µM RPW-24 did not protect nematodes exposed to 5 mM sodium arsenite for 16 hours at 20°C (Figure S5). In fact, the toxicity of RPW-24 and sodium arsenite were synergistic in this assay, resulting in nearly 100% mortality of wild-type nematodes exposed to both compounds (Figure S5).
While RPW-24 confers a survival advantage for nematodes infected with pathogenic bacteria, these data demonstrate that it is toxic to C. elegans growing under standard laboratory conditions. Whether this toxicity is a consequence of direct effects of the compound on C. elegans or hyper-activation of the immune system by RPW-24 is unknown.
We hypothesized that a C. elegans-based screen for novel anti-infectives would identify small molecules that act by stimulating the host immune response. Four lines of evidence allow us to conclude that RPW-24 protects C. elegans from bacterial infection by inducing the production of putative immune effectors via an evolutionarily conserved immune pathway. First, RPW-24 confers a survival advantage for nematodes infected with the Gram-positive bacteria E. faecalis and the Gram-negative bacteria P. aeruginosa at doses that do not inhibit in vitro growth of these pathogens. Second, whole genome microarray analysis demonstrates that RPW-24 induces a transcriptional response in C. elegans in the absence of pathogen exposure. We observed that 26% of the genes induced 3-fold or greater are also induced by infection with P. aeruginosa. Moreover, we showed that the RPW-24-mediated induction of several putative immune effectors is enhanced in the context of P. aeruginosa infection. Third, using C. elegans animals deficient in immune pathway signaling in pathogenesis assays and gene expression analyses, we demonstrated that the TIR-1/NSY-1/SEK-1/PMK-1 cascade in C. elegans is required for RPW-24 to exert its full effect. Lastly, an RNAi screen indicated that the transcription factor ATF-7 controls the RPW-24-mediated induction of a putative immune effector gene. The ability of RPW-24 to promote survival of P. aeruginosa-infected nematodes is partially dependent on this transcription factor. ATF-7 has been shown to function downstream of the MAPK PMK-1 to activate the expression of immune effectors [12]. Here we show that ATF-7 also activates the expression of genes independently of the PMK-1 p38 MAPK pathway. In summary, these data strongly suggest that immune gene activation by RPW-24 is required for its protective effects. While it is theoretically possible that RPW-24 confers protection for C. elegans by also inhibiting virulence factor production in both E. faecalis and P. aeruginosa, we feel this is less likely given that clear effects of RPW-24 on the nematode and the evolutionary diversity between these bacterial pathogens.
An unexpected observation in our transcriptome profiling analysis was that RPW-24 caused the dramatic induction of genes involved in the detoxification of small molecules, including UDP-glucuronosyltransferases (UGTs), cytochrome P450s (CYPs), glutathione-s-transferases (GSTs) and short-chain dehydrogenases (SDRs). In mammals, these gene classes act together to detoxify xenobiotic small molecules via two successive reactions [19]. Phase I reactions involve the addition of chemically reactive functional groups to the toxins and are predominantly mediated by CYPs and SDRs. In Phase II, UGTs and GSTs add side groups that increase the solubility of small toxic molecules, which aides in their excretion. These mediators of detoxification are highly conserved throughout evolution and are present in C. elegans [20], [26], [27]. Their marked induction by RPW-24 suggests that this compound may be recognized as a toxin. Consistent with this hypothesis, RPW-24 caused a strong behavioral avoidance phenotype, shortened nematode lifespan, and delayed the development of nematodes growing on non-pathogenic bacteria.
We have identified a low molecular weight molecule that can potently activate the innate immune response of C. elegans. The target of this compound is not known, nor are the mechanisms that act upstream of the p38 MAP kinase cassette to trigger the RPW-24-mediated immune activation in C. elegans. Indeed, it also unclear how any of the C. elegans immune pathways, including the p38 MAP kinase cassette, are activated during bacterial infection. Recently, several investigators have shown that the nematode monitors the integrity of cellular processes as a means to detect pathogen invasion, and to trigger defense responses and behavioral avoidance phenotypes. McEwan et al. [28] and Dunbar et al. [29] each found that the inhibition of translation by a bacterial toxin induced a protective immune response in C. elegans that was dependent on the p38 MAP kinase and ZIP-2 pathways. Similarly, Melo et al. showed that disruption of core cellular processes, such as translation, mitochondrial respiration and proteasome function, by bacterial toxins induced a behavioral avoidance phenotype [22]. Thus, it is possible that the toxic effects of RPW-24 trigger immune response pathways and a behavioral aversion response in an analogous manner. Alternatively, RPW-24 itself could directly activate immune pathways in C. elegans.
In this study, we have demonstrated the utility of using small molecules in conjunction with classical epistasis analysis and RNAi screens to dissect immune signaling pathways in C. elegans. We therefore hypothesize that RPW-24 can be used as a tool in additional genetic analyses both to identify the target(s) of this small molecule and to determine mechanism by which the p38 MAP kinase pathway is activated by RPW-24, which may offer insights into how C. elegans detects bacterial pathogens.
The World Health Organization has declared that antimicrobial-resistant pathogens are one of the three greatest threats to human health. Exacerbating this problem is the striking absence of novel antibiotics in the development pipeline. In 2008, there were only 16 antimicrobial compounds in late stage clinical trials and only one of these agents had a novel mechanism of action [30]. Furthermore, all of these agents target some aspect of bacterial replication or metabolism. Indeed, it has been suggested that the obvious bacterial targets amenable for antimicrobial drug design have been exhausted [31]. Identifying host-acting small molecules that modulate innate immune responses is a promising approach to identify novel antimicrobial therapies [32]. In theory, such agents should place minimal selection pressure on bacteria to acquire resistance determinants and could have broad-spectrum antimicrobial activity. Agonists of the mammalian Toll-like and NOD-like receptors are among the most promising compounds that are being explored for this purpose [32], [33]. We propose that C. elegans-based compound screening assays can be used to mine large chemical libraries for additional small molecule anti-infectives that act by stimulating host immune defenses. Whether or not these host-acting small molecules will be effective for the treatment of bacterial infections in mammals, genetic dissection of the C. elegans signaling pathways activated by such compounds may suggest strategies for the development of new classes of antimicrobial therapies.
C. elegans were maintained and propagated on E. coli OP50 as described [34]. The C. elegans strains used in this study were: N2 Bristol [34], pmk-1(km25) [10], sek-1(km4) [10], nsy-1(ag3) [10], tir-1(qd4) [11], atf-7(qd137) [12], atf-7(qd22 qd130) [12], fshr-1(ok778) [16], zip-2(tm4067) [17] glp-4(bn-2) [35], AU78 [agIs219 (pT24B8.5::GFP::unc-54-3′UTR pttx-3::GFP::unc-54-3′utr)] [11] and AY101 [acIs101[pDB09.1(pF35E12.5::GFP); pRF4(rol-6(su1006))] [13].
Slow-killing P. aeruginosa solid media infection assays were performed as previously described [15] with some modifications. A single colony of P. aeruginosa PA14 was innoculated into 3-mL of Luria-Bertani (LB) media and allowed to incubate at 37°C for 14 to 15 hours. 10 µL of this culture was added to 35-mm tissue culture plates containing 4 mL of slow kill agar supplemented with 1% DMSO and the indicated conentration of RPW-24. Others have shown that this concentration of DMSO has little effect on growth or development of C. elegans [36], [37]. Plates were incubated for 24 hours at 37°C and 24 hours at 25°C. C. elegans lifespan assays were conducted on nematode growth media (NGM) supplemented with the indicated concentraion of RPW-24 and seeded with OP50. The sensitivity of RPW-24-treated wild-type animals to oxidative stress was determined using 5 mM sodium arsenite following a previously described protocol [12]. For the infection, lifespan and sodium arsenite assays, 0.1 mg/mL 5-fluorodeoxyuridine (FUDR) was added to the media 1 to 2 hours before the start of the assay to prevent progeny from hatching. Approximately 50 L4 staged nematodes were picked to each of three or four assay plates per experimental condition. Animals were scored as live or dead on a daily basis by gently touching them with a platinum wire. Worms that crawled onto the wall of the tissue culture plate were eliminated from the analysis. The P. aeruginosa killing assays were conducted at 25°C. The lifespan and sodium arsenite assays were performed at 20°C. The sample sizes for each of these experiments are given in Table S2. For the experiments with the atf-7(lof) mutants, we used an N2-derived strain carrying the acIs219 transgene (AU78) as the control strain because this transgene is also present in the atf-7 mutant strains [12]. The C. elegans liquid media infection assay used to screen the 31 compounds for those with activity against P. aeruginosa-infected nematodes was developed in our laboratory and was conducted in either 384 or 96 well plates using glp-4(bn-2) animals (Kirienko, NV and Ausubel FM, unpublished data). Growth of P. aeruginosa in the presence of RPW-24 was determined by inoculating 1.0×104 bacteria in liquid slow-kill media containing either 70 µM RPW-24 or DMSO and allowing the culture to grow at 37°C in a roller drum. At the indicated time points, 10 µL of the culture was removed and CFUs were determined by plating serial dilutions.
The propensity of wild-type C. elegans to leave a lawn of bacteria supplemented with RPW-24 was assayed using a previously described protocol [22]. Briefly, 6-well tissue culture plates containing NGM were seeded with concentrated E. coli OP50. 70 µM RPW-24 or an equal volume of DMSO was added to the center of the E. coli lawn and allowed to dry. 70 to 90 young L4 animals were added to the center of the lawns and animals were scored as either on or off the lawn after 16 hours incubation at room temperature. To determine if RPW-24 slowed the development of wild-type animals, 70 µM RPW-24 or DMSO was added to NGM plates and seeded with OP50. L1 staged animals, synchronized by hypochlorite treatment, were added to these plates and allowed to incubate at 20°C for 65 hours. Developmental stages of 50 animals per treatment group were determined by microscopic examination of the gonad.
N2 animals were synchronized by hypochlorite treatment. Arrested L1s were plated on 10 cm NGM plates seeded with E. coli OP50 and grown at 20°C until the late L4 larval stage. Animals were incubated for 15 hours at 15°C in 2 mL of liquid S-basal complete medium [38] containing 70 µM RPW-24 or DMSO and supplemented with E. coli OP50. The final concentration of DMSO in both samples was 1%. RNA was extracted from three biological replicates using TRI Reagent (Molecular Research Center) according to the manufacturer's instructions and purified using an RNeasy column (Qiagen). RNA samples were prepared and hybridized to Affymetrix full-genome GeneChips for C. elegans at the Harvard Medical School Biopolymer Facility (Boston, MA) following previously described protocols [5], [8] and instructions from Affymetrix. Data were analyzed using GenePattern version 2.0 software using GC-RMA and quantile normalization [39]. Conditions were compared using GenePattern to determine the fold change between conditions for each probe set and to generate a P value using a modified t-test. Probe sets were considered differentially expressed if the fold change was 3-fold or greater (P<0.025).
Animals of the indicated genotype were treated and RNA was extracted as described for the microarray analysis. For the experiments with the atf-7(lof) mutants, we used strain AU78 as as the control strain [12]. For gene expression analysis of nematodes on solid media, 70 µM RPW-24 or DMSO was added to 20 mL nematode growth media in 10 cm petri dishes seeded with E. coli OP50. For qRT-PCR studies of nematodes infected with P. aeruginosa, 20 mL of slow killing media was added to 10 cm petri dishes containing either DMSO or 70 µM RPW-24. Plates were seeded with either 250 µL of E. coli OP50 or 50 µL P. aeruginosa diluted in 200 µL LB, each from overnight cultures. The plates were incubated for 24 hours at 37°C and 24 hours at 25°C. Old L4/young adult animals were added to the assay plates and incubated at 25°C for eight hours. RNA was reverse transcribed to cDNA using the Retroscript kit (Ambion). cDNA was analyzed by qRT-PCR using a CFX1000 machine (Bio-Rad) and previously published primers [8]. All values were normalized against the control gene snb-1, which has been used previously in qRT-PCR studies of C. elegans innate immunity [5], [8], [9], [12], [40]. Analysis of the microarray expression data revealed that the expression of snb-1 did not vary under the conditions tested in the experiment. Fold change was calculated using the Pfaffl method [41].
The RNAi screen of 393 transcription factors was conducted using RNAi clones from the Ahringer and Vidal RNAi libraries and an established protocol [42]. The RNAi clone for atf-7 was developed by Shivers et al [12]. Briefly, overnight cultures of feeding RNAi clones were added to each well of a 96-well RNAi plate and allowed to grow at room temperature overnight. 40–60 L1 staged acIs101 animals, which express a F35E12.5::GFP transgene, were added to each well and allowed to grow for two days at 20°C until they were at the L4 or young adult stage. Animals were then washed from the RNAi plates into S-basal complete media [38] containing 70 µM RPW-24. All experiments with feeding RNAi used a gfp RNAi as the negative control, which resulted in no visible GFP expression in acIs101 transgenic animals. acIs101 animals treated with the empty vector L4440 and exposed to 70 µM RPW-24 was used as the positive control. Animals were scored for GFP expression following photograph of each well using an Image Xpress Micro microscope (Molecular Devices Corporation, Sunnyvale, CA).
Nematodes were mounted onto agar pads, paralyzed with 10 mM levamisole (Sigma) and photographed using a Zeiss AXIO Imager Z1 microscope with a Zeiss AxioCam HRm camera and Axiovision 4.6 (Zeiss) software.
Differences in survival of C. elegans animals infected with P. aeruginosa on slow-killing assay plates were determined with the log-rank test. To determine if the increase in survival conferred by RPW-24 treatment was different in one population compared to another [for example, in wild-type versus pmk-1(km25) animals], we examined the difference in the effect of RPW-24 treatment on the hazard in each group using a Cox proportional hazard model (Stata11, Stata, College Station, TX). Fold changes in the qRT-PCR analyses and the differences in survival in the arsenite assays were compared using unpaired, two-tailed student t-tests. When comparing microarray datasets, the overlap expected by chance alone was determined in 50 groups of randomly selected C. elegans genes using Regulatory Sequence Analysis Tools (http://rsat.ulb.ac.be/rsat/), a technique that has been used for similar analyses [43]. P values were determined using chi-square tests.
Accession numbers for the genes and gene products mentioned in this paper are given for Wormbase, a publicly available database that can be accessed at http://www.wormbase.org. These accession numbers are pmk-1 (B0218.3), nsy-1 (F59A6.1), sek-1 (R03G5.2), atf-7 (C07G2.2), zip-2 (K02F3.4), fshr-1 (C50H2.1), skn-1 (T19E7.2), C32H11.1, F35E12.5, F08G5.6, C29F3.7, K08D8.5, F49F1.6, C14C6.5, lys-7 (C02A12.4), clec-67 (F56D6.2) and mtl-1 (K11G9.6). The microarray dataset can be downloaded from the National Center for Biotechnology Gene Expression Omnibus (GEO; http://www.ncbi.nlm.nih.gov/geo). The accession number for these data is GSE37266.
|
10.1371/journal.pcbi.1003171 | Task-Based Core-Periphery Organization of Human Brain Dynamics | As a person learns a new skill, distinct synapses, brain regions, and circuits are engaged and change over time. In this paper, we develop methods to examine patterns of correlated activity across a large set of brain regions. Our goal is to identify properties that enable robust learning of a motor skill. We measure brain activity during motor sequencing and characterize network properties based on coherent activity between brain regions. Using recently developed algorithms to detect time-evolving communities, we find that the complex reconfiguration patterns of the brain's putative functional modules that control learning can be described parsimoniously by the combined presence of a relatively stiff temporal core that is composed primarily of sensorimotor and visual regions whose connectivity changes little in time and a flexible temporal periphery that is composed primarily of multimodal association regions whose connectivity changes frequently. The separation between temporal core and periphery changes over the course of training and, importantly, is a good predictor of individual differences in learning success. The core of dynamically stiff regions exhibits dense connectivity, which is consistent with notions of core-periphery organization established previously in social networks. Our results demonstrate that core-periphery organization provides an insightful way to understand how putative functional modules are linked. This, in turn, enables the prediction of fundamental human capacities, including the production of complex goal-directed behavior.
| When someone learns a new skill, his/her brain dynamically alters individual synapses, regional activity, and larger-scale circuits. In this paper, we capture some of these dynamics by measuring and characterizing patterns of coherent brain activity during the learning of a motor skill. We extract time-evolving communities from these patterns and find that a temporal core that is composed primarily of primary sensorimotor and visual regions reconfigures little over time, whereas a periphery that is composed primarily of multimodal association regions reconfigures frequently. The core consists of densely connected nodes, and the periphery consists of sparsely connected nodes. Individual participants with a larger separation between core and periphery learn better in subsequent training sessions than individuals with a smaller separation. Conceptually, core-periphery organization provides a framework in which to understand how putative functional modules are linked. This, in turn, enables the prediction of fundamental human capacities, including the production of complex goal-directed behavior.
| Cohesive structures have long been thought to play an important role in information processing in the human brain [1]. At the small scale of individual neurons, temporally coherent activity supports information transfer between cells [2]. At a much larger scale, simultaneously active cortical areas form functional systems that enable behavior [1]. However, the question of precisely what type of cohesive organization is present between the constituents of brain systems—especially at larger scales—has been steeped in controversy [3], [4]. Although low-frequency interactions between pairs of brain areas are easy to measure, the simultaneous characterization of dynamic interactions across the entire human brain remained challenging until recent applications of network theory to neuroimaging data [5]. These efforts have led to enormous insights, including the establishment of relationships between stationary functional brain network configuration and intelligence [6] as well as relationships between altered brain network organization and disease [7]. In this paper, we extend this approach to a non-stationary situation: the change of network activity across the brain as a new skill is acquired.
Acquisition of new motor skills alters brain activity across spatial scales. At the level of individual neurons, this induces changes in firing behavior in the motor cortex [8]. At the level of large-scale areas, this induces changes in the interactions between primary motor cortex and premotor areas, and these changes can influence the amount of learning [9]. Previous studies have demonstrated that pairwise interactions between some of these premotor regions, as measured by the magnitude of coherence between low-frequency blood-oxygen-level-dependent (BOLD) signals , strengthen with practice [10]. Furthermore, complex contributions by non-motor systems such as prefrontal cortex are involved in the strategic control of behavior during learning [11]. These findings reveal some of the changes in local circuits that occur with learning. However, there remains no global assessment of changes in brain networks as a result of learning. In this paper, we seek to find cohesive structures in global brain networks that capture dynamics that are particularly relevant for characterizing skill learning that takes place over the relatively long time scales of minutes to hours of practice.
To address these issues, we extract a set of functional networks from task-based functional Magnetic Resonance Imaging (fMRI) time series that describe functional connectivity between brain regions. We probe the dynamics of these putative interactions by subdividing time series into discrete time intervals (of approximately two minutes in duration; see Fig. 1A) during the acquisition of a simple motor skill. Subjects learned a set of 10-element motor sequences similar to piano arpeggios by practicing for at least 30 days during a 6-week period. The depth of training was manipulated so that 2 sequences were extensively practiced (EXT), 2 sequences were moderately practiced (MOD), and 2 sequences were minimally practiced (MIN) on each day. In addition, subjects performed blocks of all of the sequences during fMRI scanning on approximately days 1, 14, 28, and 42 of practice. Using the fMRI time series, we extract functional networks representing the coherence between 112 cortical and subcortical areas for each sequence block.
To characterize brain dynamics, we represent sets of functional networks as multilayer brain networks and we identify putative functional modules—i.e., groups of brain regions that exhibit similar BOLD time courses—in each 2–3 minute time window. Such cohesive groups of nodes are called “communities” in the network-science literature [12], [13], and they suggest that different sets of brain regions might be related to one another functionally either through direct anatomical connections or through indirect activation by an external stimulus. A community of brain regions might code for a different function (e.g., visual processing, motor performance, or cognitive control), or it might engage in the same function using a distinct processing stream. Characterizing changes in community structure thus makes it possible to map meaningful dynamic patterns of functional connectivity that relate to changes in cognitive function (e.g., learning).
We employ computational tools for dynamic community detection [14], [15] for multilayer representations of temporal networks [16] and summarize our findings using diagnostics that quantify three properties of community structure. (See Materials and Methods for their definitions and Ref. [17] for evidence supporting the utility of these diagnostics in capturing changes in brain dynamics over 3 days of learning.) To measure the strength of functional modularization in the brain and quantify the extent of compartmentalization of putative functional modules, we maximize a quality function called multilayer modularity to obtain a partition of the brain into communities. (The associated maximum value of is known as the maximum modularity.) A high value of indicates that the pattern of functional connectivity in the brain can be clustered sensibly into distinct communities of brain regions that exhibit similar time courses. We also compute the number of communities (i.e., putative functional modules) in partitions of the multilayer networks. A large value of indicates that there are a large number of distinct temporal profiles in BOLD activations in the brain. To measure the temporal variability of community structure, we compute the flexibility of each region , as this quantifies the frequency that a brain region changes its allegiances to network communities over time. A high value of flexibility indicates that a region often changes community affiliation.
Our results demonstrate that the temporal evolution of community structure is modulated strongly by the depth of training (as reflected in the total number of practiced trials). We also show that the temporal variability of module allegiance varies across brain regions. Sensorimotor and visual cortices form the bulk of a relatively stiff temporal core in which module affiliations change little over a scanning session, whereas multimodal association areas form the bulk of a relatively flexible temporal periphery in which module affiliations change frequently. The separation between the temporal core and temporal periphery predicts individual differences in extended learning. We combine these methods for identifying a temporal core and periphery with a notion of core-periphery organization that originated in the social sciences [18] to show that the organizational structure of functional networks in 2–3 minute time windows correlates with the organizational structure of the brain's temporal evolution: densely connected regions in individual time windows tend to exhibit little change in module allegiance over time, whereas weakly connected regions tend to exhibit significant changes. Taken together, our results suggest that core-periphery organization is a critical property that is as important as modularity for understanding and predicting cognition and behavior (see Fig. 1B).
Community structure changes with the number of trials practiced, independent of when the practice occurred in the 6 weeks. In Fig. 2, we show multilayer modularity (, a measure of the quality of a partition into communities), the number of communities, and mean flexibility (, a measure of the temporal variability in module allegiance) as a function of the number of training trials completed after a scanning session. See Materials and Methods for the definitions and Table 1 for the relationship between the number of trials practiced and training duration and intensity. After an initial increase from 50 to 200 trials practiced, multilayer modularity decreases with an increase in the number of trials practiced, suggesting that community structure in functional brain networks becomes less pronounced with learning. Both the number of communities and the flexibility of community structure increase with the number of trials practiced, which is consistent with an increased specificity of functional connectivity patterns with extended learning.
Given our demonstration that there exists a temporal core in dynamic brain networks, it is important to ask what role such core regions might play in individual network layers of the multilayer network [17]. While the roles of nodes in a static network can be studied in multiple ways [28], [29], we focus on describing the geometrical core-periphery organization— which can be used to help characterize the organization of edge strengths throughout a network —to compare it with the temporal core-periphery organization discussed above. The geometrical core of a network is composed of a set of regions that are strongly and mutually interconnected. Measures of network centrality can be useful for identifying nodes in a geometrical core because such measures help capture a node's relative importance within a network in terms of its immediate connections, its distance to other nodes in the graph, or its influence on other nodes in the graph [30], [31].
Drawing on studies of social networks [18], we examine geometrical core-periphery organization in networks extracted from individual time windows by testing whether core nodes are densely connected to one another and whether peripheral nodes are sparsely connected to one another. Rather than proposing a strict separation between a single core and single periphery, we assess the role of a node along a core-periphery spectrum using a centrality measure known as the (geometrical) core score , which was introduced in Ref. [30]. Network nodes with high values are densely connected to one another, whereas nodes with low values are sparsely connected to one another. The method in Ref. [30] uses a two-parameter function to interpolate between core nodes and peripheral nodes. One parameter (which is denoted by ) sets the sharpness of the boundary between the geometrical core and the geometrical periphery. Small values of indicate a fuzzy boundary, whereas large values indicate a sharp transition. The second parameter (which is denoted by ) sets the size of the geometrical core. Smaller values of correspond to smaller cores. We can quantify the fit of the transition function that defines the set of core scores to the data using a summary diagnostic that is called the -score (see Materials and Methods for definitions ). Large values of indicate a good fit and therefore provide confidence that one has uncovered a good estimate of a network's core-periphery organization.
In Fig. 4A, we show a typical -score landscape in the parameter plane. This landscape favors a relatively small core and a medium value of the transition-sharpness parameter. To choose sensible values of and for studying core-periphery organization, we examine the distributions of the relative frequencies of and values that maximize the -score for each network layer, participant, scanning session, and sequence type (see Fig. 4B). We use the mean values of these distributions ( and ) to assign a core score to each node. In Fig. 4C, we show the shape of the “mean core” that we obtain using these parameter values. This figure demonstrates that the typical (geometrical) core-periphery organization in the networks under study is a mixture between a discrete core-periphery organization, in which every node is either in the core or in the periphery, and a continuous core-periphery organization, in which there is a continuous spectrum to describe how strongly nodes belong to a core. In these networks (which usually possess a single core), the majority of nodes do not belong to the core, but those nodes that do (roughly of the nodes) have a continuum level of association strengths with the core.
In some cases, we identified multiple competing cores, which we found by using simulated annealing to explore local maxima of the -score rather than only identifying a global maximum. Because of this stochasticity in the methodology for examining core-periphery organization, we performed computations with the chosen parameter values ( and ) 10 times and used the solution with the highest -score out of these 10 iterations for each network layer, subject, scanning session, and sequence type.
An interesting question is whether geometrical core-periphery organization remains relatively constant throughout time or whether the organization changes with learning. We observed that regions that have a high geometrical core score in the first scanning session and in EXT blocks were likely to have high geometrical core scores in later scanning sessions and in MOD and MIN blocks. (See the Text S1 for supporting results on the reliability of geometrical core-periphery organization.) In light of this consistency, we calculate a mean geometrical core score for each node by taking the mean over all blocks in a given scanning session (1, 2, 3, and 4) and sequence type (EXT, MOD, and MIN). The variance of the mean geometrical core score over nodes in a network then gives an indication of the separation between the mean core and periphery. As we show in Fig. 5, we find that the variance of the mean geometrical core score over trials decreases as a function of learning. A high variance of the mean core scores over nodes indicates a greater separation between the mean core and periphery as well as a high consistency of the core score of each node over trial blocks. If a node's core score is inconsistent over trial blocks, then the mean core score for each node in the network is expected to be similar and thus one would expect the variance of the core scores over nodes to be small. A low variance in the mean geometrical core score over trial blocks therefore suggests either little separation between the core and periphery or an increased variability in core scores of a given node over trial blocks.
Given the geometrical core-periphery organization in the individual layers of the multilayer networks and the temporal core-periphery organization in the full multilayer networks, it is important to ask whether brain regions in the temporal core (i.e., regions with low flexibility) are also likely to be in the geometrical core (i.e., whether they exhibit strong connectivity with other core nodes, as represented by a high value of the geometrical core score). In Fig. 6, we show scatter plots of the flexibility and core score for the 3 training levels (EXT, MOD, and MIN) and the 4 scanning sessions over the 6-week training period. We find that the temporal core-periphery organization (which is a dynamic measurement) is strongly correlated with the geometrical core score (which is a measure of network geometry and hence of network structure). This indicates that regions with low temporal flexibility tend to be strongly-connected core nodes in (static) network layers. In Fig. 6, we show that the relationship between temporal and geometrical core-periphery organization occurs reliability across training depth, duration, and intensity. In Fig. 7, we show that this relationship can also be identified robustly in data extracted from individual subjects.
We have shown how the mesoscale organization of functional brain networks changes over the course of learning. Our results suggest that core-periphery organization is an important and predictive component of cognitive processes that support sequential, goal-directed behavior. We summarize our findings in Fig. 7, which demonstrates that poor learners tend to have poorer separation between core and periphery (as indicated by straighter, shorter spirals in the figure) and that good learners tend to have greater separation between core and periphery (curvier, longer spirals). Our findings also demonstrate that during the generation of motor sequences, the brain consists of a temporally stable and densely connected set of core regions complemented by a temporally flexible and sparsely connected set of peripheral regions. This functional tradeoff between a core and periphery might provide a balance between the rigidity necessary to maintain motor function by the core and the adaptivity of the periphery necessary to enable behavioral change as a function of context or strategy.
In the Text S1, we provide supporting results that indicate (i) that our findings are not merely a function of variation in region size and (ii) that they cannot be derived from the underlying block design of the experimental task. We also show in this supplement that (arguably) simpler properties of brain function—such as the regional signal power of brain activity, mean connectivity strength, and parameter estimates from a general linear model—provide less predictive power than core-periphery organization.
The notion of a core-periphery organization is based on the structure (rather than the temporal dynamics) of a network [30]. Intuitively, a core consists of a set of highly and mutually interconnected set of regions . In this paper, we have described what is traditionally called “core-periphery structure” using the terminology geometrical core-periphery organization. (It is geometrical rather than topological because the networks are weighted.) This intuitive notion was formalized in social networks by Borgatti and Everett in 1999 [18]. Available methods to identify and quantify geometrical core-periphery organization in networks include ones based on block models [18], -core organization [32], and aggregation of information about connectivity and short paths through a network [33]. Unfortunately, many methods that have been used to study cores and peripheries in networks have binarized networks that are inherently weighted, which requires one to throw away a lot of important information. Even the recently developed weighted extensions of -core decomposition [34] require a discretization of -shells, which have been defined for both binary and weighted networks [35]. Importantly, -core decomposition is based on a very stringent and specific type of core connectivity, so this measure misses important core-like structures [30], [36]. A well-known measure called the “rich-club coefficient” (RCC) [37] considers a different but somewhat related question of whether nodes of high degree (defined as some threshold value ) tend to connect to other nodes of high degree. (The RCC is therefore a form of assortativity.) The RCC has also been extended to weighted networks [38], but it still requires one to specify a threshold value of richness to enable one to ask whether “rich” nodes tend to connect to other “rich” nodes.
The aforementioned limitations notwithstanding, several of the measures discussed above have recently been used successfully to identify a structural core of the human brain white-matter tract network, which is characterized not only by a -core with a high value of the degree (in particular, ) [34] and rich club [39], [40] but also by a knotty center of nodes that have a high geodesic betweenness centrality but not necessarily a high degree [36]. A -core decomposition has also been applied to functional brain imaging data to demonstrate a relationship between network reconfiguration and errors in task performance[41].
A novel approach that is able to overcome many of these conceptual limitations is the geometrical core-score [30], which is an inherently continuous measure, is defined for weighted networks, and can be used to identify regions of a network core without relying solely on their degree or strength (i.e., weighted degree). Moreover, by using this measure, one can produce (i) continuous results, which make it possible to measure whether a brain region is more core-like or periphery-like; (ii) a discrete classification of core versus periphery; or (iii) a finer discrete division (e.g., into 3 or more groups). In addition, this method can identify multiple geometrical cores in a network and rank nodes in terms of how strongly they participate in different possible cores. This sensitivity is particularly helpful for the examination of brain networks for which multiple cores are hypothesized to mediate multimodal integration [42]. In this paper, we have demonstrated that functional brain networks derived from task-based data acquired during goal-directed brain activity exhibit geometrical core-periphery organization. Moreover, they are specifically characterized by a straightforward core-periphery landscape that includes a relatively small core composed of roughly 10% or so of the nodes in the network.
In this paper, we have introduced a method and associated definitions to identify a temporal core-periphery organization based on changes in a node's module allegiance over time. We have defined the notion of a temporal core as a set of regions that exhibit fewer changes in module allegiance over time than expected in a dynamic-network null model. Neurobiologically, the temporal core contains brain areas that show consistent task-based mesoscale functional connectivity over the course of an experiment , and it is therefore perhaps unsurprising that their anatomical locations differ from nodes in the )-core [34] and RCC [39], [40] of the human white matter tract network. Our approach is inspired by the following idea: although the brain uses the function of a small subset of regions to perform a given task (i.e., some sort of core ), a set of additional regions that are associated more peripherally with the task might also be activated in a transient manner. Indeed, several recent studies have highlighted the possibility of a separation between groups of regions that are consistently versus transiently activated during task-related function [43], [44], and they have demonstrated that correlations between such regions can be altered depending on their activity [43], [45].
Given the very different definitions of the geometrical and temporal cores, it is interesting that nodes in the temporal core are also likely to be present in the geometrical core. Importantly, the notions of temporal and geometrical core are complementary, and they are both intuitive in the context of brain function. A set of regions that is coherently active to perform a task (i.e., is in the geometrical core) must remain online consistently throughout an experiment (i.e., be in the temporal core), whereas a set of regions that might be activated less coherently (i.e., is in the geometrical periphery) can be utilized by separate putative functional modules over time (i.e., it can be in the temporal periphery). This interpretation is consistent with the notion that the anatomical locations of the core and periphery are task-specific. Should brain activity during other tasks also exhibit core-periphery organization, then the core and periphery of these other task networks could consist of a different set of anatomical regions than those observed here. A comparison of dynamic community structure and associated mesoscale organizational properties across brain states elicited by other tasks is outside of the scope of the present study. However, such a study in a controlled sample with similar time-series length and experimental task structure (e.g., trial lengths, block lengths, and rest periods) would likely yield important insights.
Community structure and core-periphery organization are two types of mesoscale structures, and they can both be present simultaneously in a network [30], [36]. Moreover, both modular and core-periphery organization can in principle pertain to different characteristics of or constraints on underlying brain function. In particular, the presence of community structure supports the idea of the brain containing putative functional modules, whereas the presence of a core-periphery organization underscores the fact that different brain regions likely play inherently different roles in information processing. A symbiosis between these two types of organization is highlighted by the findings that we report in this manuscript: the dynamic reconfiguration of putative functional modules can be described parsimoniously by temporal core-periphery organization, demonstrating that one type of mesoscale structure can help to characterize another. Furthermore, the notion that the brain can simultaneously contain functional modules (e.g., the executive network or the default-mode network) and regions that transiently mediate interactions between modules is consistent with recent characterizations of attention and cognitive control processes [46].
It is increasingly apparent that functional connectivity in the brain changes over time and that these changes are biologically meaningful. Several recent studies have highlighted the temporal variability [47]–[50] and non-stationarity [51] of functional brain network organization, and both of these features are apparent over short time intervals (less than 5 minutes in fMRI; less than 100 s in EEG) [47]–[49]. Although temporal variability in functional connectivity was seen initially as a signature of measurement noise [51], recent evidence suggests instead that it might provide an indirect measurement of changing cognitive processes. Thus, it might serve as a diagnostic biomarker of disease [51], [52]. Moreover, such temporal variability appears to be modulated by exogenous inputs. For example, Barnes et al. [53] demonstrated using a continuous acquisition “rest-task-rest” design that endogenous brain dynamics do not return to their pre-task state until approximately 18 minutes following task completion. Similar results that consider other tasks have also been reported [54]. More generally, the dynamic nature of brain connectivity is likely linked to spontaneous cortical processing, reflecting a combination of both stable and transient communication pathways [48], [49], [55].
In this study, we observed that properties of the temporal organization of functional brain networks (e.g., on day 1 of this experiment) can be used to predict extended motor learning (e.g., on the following 10 days of home training on a discrete sequence-production task). Our findings are consistent with two previous studies that demonstrated a predictive connection between both dynamic [17] and topological [56] network organization and subsequent learning. (Note that we use the term topological because Ref. [56] considered only unweighted networks.) Reference [17] focused on early—rather than extended—learning of a cued sequence-production motor task (rather than a discrete one) and found that network flexibility on the first day of experiments predicted learning on the second day and that flexibility on the second day predicted learning on the third day. Reference [56] investigated participants' success in learning words of an artificial spoken language and found that network properties from individual time windows could be used to predict such success [56]. Together with the present study, these results highlight the potential breadth of the relationship between network organization and learning. The presence of such a relationship has now been identified across multiple tasks, over multiple time scales, and using both dynamic and topological network properties.
Our study has focused on large-scale changes in dynamic community structure that are correlated with learning. Finer-scale investigations that employ alternative parcellation schemes [57]–[62] with greater spatial resolution or alternative neuroimaging techniques such as EEG or MEG [55] with greater temporal resolution might uncover additional features that would enhance understanding of functional network-based predictors of learning phenomena.
Throughout this paper, we have referred to feature similarities (which we estimated using the magnitude squared coherence) between pairs of regional BOLD time series as functional connectivity [63]. As appreciated in prior literature [64]–[66], the interpretation of functional connectivity must be made with caution. Coherence in the activity recorded at different brain sites does not necessitate that those sites share information with one another to enable cognitive processing, as they could instead indicate that those two sites are activated by the same third party (either another brain region or an external stimulus). In this paper, we do not distinguish between these two possible drivers of strong inter-regional coherence. Future studies could employ multiple estimates of statistical associations in the form of diagnostics [67]–[69] and/or models [70], [71] that might uncover other sets of interactions that could predict the observed coherence structure and hence the observed behavior.
We performed all data analysis and statistical tests in Matlab. We performed the dynamic community detection procedure using freely available Matlab code [99] that optimizes multilayer modularity using a Louvain-like locally greedy algorithm [100].
|
10.1371/journal.pgen.1006538 | Rfx2 Stabilizes Foxj1 Binding at Chromatin Loops to Enable Multiciliated Cell Gene Expression | Cooperative transcription factor binding at cis-regulatory sites in the genome drives robust eukaryotic gene expression, and many such sites must be coordinated to produce coherent transcriptional programs. The transcriptional program leading to motile cilia formation requires members of the DNA-binding forkhead (Fox) and Rfx transcription factor families and these factors co-localize to cilia gene promoters, but it is not clear how many cilia genes are regulated by these two factors, whether these factors act directly or indirectly, or how these factors act with specificity in the context of a 3-dimensional genome. Here, we use genome-wide approaches to show that cilia genes reside at the boundaries of topological domains and that these areas have low enhancer density. We show that the transcription factors Foxj1 and Rfx2 binding occurs in the promoters of more cilia genes than other known cilia transcription factors and that while Rfx2 binds directly to promoters and enhancers equally, Foxj1 prefers direct binding to enhancers and is stabilized at promoters by Rfx2. Finally, we show that Rfx2 and Foxj1 lie at the anchor endpoints of chromatin loops, suggesting that target genes are activated when Foxj1 bound at distal sites is recruited via a loop created by Rfx2 binding at both sites. We speculate that the primary function of Rfx2 is to stabilize distal enhancers with proximal promoters by operating as a scaffolding factor, bringing key regulatory domains bound by Foxj1 into close physical proximity and enabling coordinated cilia gene expression.
| The multiciliated cell extends hundreds of motile cilia to produce fluid flow in the airways and other organ systems. The formation of this specialized cell type requires the coordinated expression of hundreds of genes in order to produce all the protein parts motile cilia require. While a relatively small number of transcription factors has been identified that promote gene expression during multiciliate cell differentiation, it is not clear how they work together to coordinate the expression of genes required for multiple motile ciliation. Here, we show that two transcription factors known to drive cilia formation, Foxj1 and Rfx2, play complementary roles wherein Foxj1 activates target genes but tends not to bind near them in the genome, whereas Rfx2 can’t activate target genes by itself but instead acts as a scaffold by localizing Foxj1 to the proper targets. These results suggest not only a mechanism by which complex gene expression is coordinated in multiciliated cells, but also how transcriptional programs in general could be modular and deployed across different cellular contexts with the same basic promoter configuration.
| Animal gene expression is typically mediated by cell type specific transcription factors, acting through consensus binding sites present in distal enhancers and proximal promoters. Such factors act by opening chromatin,[1,2] by facilitating the deposition of histone modifications,[3] by employing local interactions with other factors,[3,4] or some combination of these. While the details of how discrete sites in the genome accumulate transcriptional machinery are just coming into focus, the mechanism by which these sites bind to and coordinate gene expression at a given promoter is less well known. Recent work hints that this coordination is mediated by 3-dimensional chromosomal architecture.[5–10] More specifically, genomic regions with high levels of spatial self-interaction, termed topological domains or topologically-associated domains (TADs), have been proposed to encourage promoter-enhancer interactions or transcriptional activation more generally.
Work on discrete transcriptional programs has demonstrated the integration of multiple transcription factors at individual promoter elements.[11] The production of specialized cilia in animals requires one such transcriptional program: these structures are deployed in a variety of cell types in different anatomical locations but appear to arise in these different settings using a similar transcriptional architecture.[12,13] For example, the presence of both Fox and Rfx transcription factors at a handful of cilia gene promoters has been associated with robust gene activation in the mammalian airway and also Drosophila neurons.[14,15] However, cilia are composed of hundreds of gene products, and it is not clear how many of them require cooperation between Fox and Rfx factors at promoters or at distal regulatory elements, to what extent other factors are involved, or if changes in 3-dimensional genome structure accompany their specification.
Here we analyze how Foxj1 and Rfx2 cooperate, both at the level of sequence specificity and chromosomal architecture, within Xenopus skin progenitors to promote gene expression during multiciliated cell (MCC) differentiation. First, we use extensive RNAseq to capture a robust transcriptome of MCC genes. Next, we use tethered conformation capture, a variant of HiC, to map the organization of topologically-associated domains within epithelial progenitors. We use ChIPseq of the cohesin component Rad21 to confirm these domains and map the location of active regions (H3K4me3 and H3K27ac) and genes expressed during MCC differentiation to them. We further analyze Foxj1 binding using ChIPseq and find that its binding, along with Rfx2, is a good predictor of MCC gene expression as compared to the binding of other major MCC transcription factors such as Myb and E2f4. We further show that Foxj1 stabilization at genomic sites, especially MCC promoters, frequently requires Rfx2. Finally, by examining 3D chromatin interactions, we show that Rfx2, Foxj1, and the promoters of MCC genes lie at the anchor endpoints of DNA loops, and that these interactions are stronger in progenitors converted entirely into multiciliated cells. We propose a model in which dimerization of Rfx proteins recruits or stabilizes distal enhancers to promoters as a scaffolding factor, enabling distantly bound factors such as Foxj1 to promote stable transcription of cilia genes during terminal MCC differentiation.
The MCC is a major cell type in the Xenopus larval skin that forms in the early embryo along with mucus-secreting cells and proton-secreting ionocytes. The differentiation of these cell types can be readily analyzed using RNAseq[16] analysis on skin progenitors isolated away from the embryo. To identify gene expression associated with distinct differentiation programs, we analyzed isolated progenitors where cell fate was altered in a predictable way. Blocking Notch in these progenitors (by expressing a DNA binding mutant of suppressor of hairless, dbm[17], labeled here as “Notch-”) leads to a marked increase in the number of MCCs and ionocytes that arise within an inner layer and move into the outer epithelium (Fig 1B and 1C). Activating Notch, (by expressing the intracellular domain of Notch, icd[18], labeled here as “Notch+”) completely suppresses MCC and ionocyte formation. Ectopic expression of Multicilin[19] in skin progenitors along with active Notch completely rescues MCC differentiation, and moreover, converts most skin progenitors into MCCs, including those in the outer layer normally fated to become mucus-secreting cells. Inhibiting Multicilin activity using a dominant-negative mutation (termed dnMulticilin [19]) blocks the formation of MCCs, but spares other cell types such as ionocytes. Finally, for an additional comparison, we ectopically expressed Foxi1 in these progenitors, a transcription factor that promotes ionocyte differentiation alone. Epithelial progenitors were isolated from injected embryos at stage 10, and subjected to RNAseq analysis as the various differentiation programs unfolded (Fig 1C). To maximize the cell-specific signal in these data sets, we compared conditions in which the change in cell number was great: for example, progenitors with activated Notch (using Notch-icd) have virtually no MCCs, whereas progenitors co-injected with activated Notch and Multicilin have large numbers without a confounding ionocyte population. MCC and ionocyte number is also markedly increased by blocking Notch signaling (using dbm), whereas progenitors co-injected with dbm and dnMulticilin have virtually no MCCs but still have increased ionocytes. A complete comparison of the effects of these manipulations is shown in S1 Table. Clustering of genes with the greatest variance (see RNAseq informatics section in Methods) in these extensive RNAseq data sets revealed that the largest cluster of genes by far was the one that changed in association with MCC differentiation (Fig 1F; this group of conditions is detailed in S1 Table and shown in S1 Fig; the cluster of genes themselves is in S2 Table). By comparison, this cluster of genes associated with MCC differentiation dwarfed that associated with ionocyte differentiation (induced with Foxi1 as a control, Fig 1F, S1A Fig, S3 Table), with Notch signaling (Fig 1F, S1B Fig, S4 Table), or with embryonic age (Fig 1F, S1C and S1D Fig, S5 and S6 Tables). These data emphasize the relatively large number of genes coordinately upregulated during the formation of this specialized epithelial cell type.
We next generated a reliable list of genes upregulated during MCC differentiation based on differential expression (p adjusted for multiple testing < 0.05) between pairs of conditions exhibiting the greatest change in MCC number (Fig 1B and 1E: dbm vs. Notch-icd, Notch-icd + Multicilin vs. Notch-icd, and dbm vs dbm + dnMulticilin). A conservative list of 950 genes was derived from the oldest, 9-hour timepoint (813 genes if we collapse the L and S forms based on orthology assignment); this list changed in all three comparisons (p<0.05, Fig 1E), and provides a high-confidence collection of the genes upregulated during MCC differentiation, which we designate “core MCC genes” (S7–S9 Tables). To assess this claim, we compared our core MCC gene list to those generated using other approaches, including Foxj1 overexpression in Zebrafish,[20] FACS sorting of Foxj1+ cells from mammalian airway epithelia,[21] or knockdown of Rfx2 in Xenopus.[22] These gene lists, as well as our own, were plotted using our data based on their level of expression (using normalized counts) versus the fold-change that occurred in our RNAseq analysis as MCC number increased. In each case, the published lists lacked a large fraction of genes on our list (red dots in S2B–S2D Fig, also see S11–S13 Tables for a comparison of overlap between these sets), and the genes unique to our list for each comparison were strongly enriched for cilia GO terms (S3A, S3C and S3E Fig). Conversely, the published list included a large number of genes that did not increase with more MCCs in our data, and thus absent from our list (grey dots in S2B–S2D Fig), but these genes were either not strongly enriched for any GO terms or enriched for more general terms (S3B, S3D and S3F Fig e.g., “biological process”). Altogether, these comparisons suggest that the core MCC list more accurately identifies genes upregulated during MCC differentiation, and thus serves as a solid foundation for the subsequent analysis described below.
We next related the MCC core genes to the 3-dimensional genome, in particular to areas of high self-interaction known as topologically-associated domains (TADs). To obtain 3D chromosomal interactions across the entire genome, we performed tethered conformation capture (TCC)[23] on epithelial progenitors isolated from X. laevis embryos, analyzing both wild-type tissue (unmanipulated progenitors containing a mixture of outer cells, ionocytes, and multiciliated cells), and after inducing MCC differentiation using Multicilin (resulting in progenitors consisting of MCCs). Most of the interactions in our data sets were within chromosomes (92%-87% of totals, respectively, S4 Fig), indicating a low level of interchromosomal interactions (high levels of these are thought to reflect spurious interactions[24]). The active and inactive compartments in the wild-type and Multicilin-injected samples were compared by computing Pearson’s correlation matrices[25] (Fig 2A). This analysis failed to detect loci with a negative correlation coefficient and very few with correlation coefficients below 0.5 (268 bins out of a possible 50,030, S14 Table), suggesting that while these two tissues are significantly different in their transcriptional profiles (Fig 1), there was very little in the way of large-scale changes, such as compartment switching between the two conditions.[7,25] Genes found in the few regions that were poorly correlated are not implicated in MCC differentiation (S14 Table), while those on our core MCC list, for example, wdr16, were highly correlated between the two conditions (Fig 2A).
At higher resolution, 3D interaction maps reveal TADs, structures thought to facilitate complex transcriptional regulation.[7,9,26,27] To determine the positions of TADs in the X. laevis genome, we pooled the reads from the two TCC experiments and calculated a directionality index,[26] resulting in 7,249 domains (Fig 2A and 2B). As TAD boundaries are reportedly enriched for CTCF and cohesin, we performed ChIPseq on the cohesin component Rad21[26,28] (Fig 2A, S5 Table) and found striking enrichment of this protein at our called domain boundaries, providing independent validation that they were called accurately (Fig 2A and 2C). We also found Rad21 peaks not associated with TADs (Fig 2A, ~81% of total peaks); reports by others suggest that both Rad21 and CTCF are frequently outside TADs and are most likely to mark TAD boundaries when cobound.[29] We annotated the positions of transcriptional start sites (TSS’s) relative to TADs and found enrichment at the boundaries (Fig 2C), in agreement with previous work on mammalian genomes.[26] We further performed ChIPseq on two histone modifications: H3K4me3, a mark associated with actively-transcribed promoters (Fig 2A, S6 Fig), and H3K27ac, a mark associated with active promoters and enhancers, including superenhancers[30] (Fig 2A, S7 Fig). The H3K4me3 ChIPseq (Fig 2A and 2D, S6 Fig) exhibited broadly similar nucleotide frequencies and peak shapes as other vertebrate promoters[31] and largely overlapped, within a 2 Kb window, our annotated 5’ ends of X. laevis gene models (16,231 out of 24,384 total models; note not all of these genes will be expressed in epithelial progenitors and would thus largely lack H3K4me3 peaks). When these features were overlaid onto the TAD map, the promoter regions, as defined by both annotated TSS’s and H3K4me3 peaks, tended to lie close to the TAD boundaries (Fig 2C and 2D), while enhancers were depleted at boundaries and instead enriched in the middle of the domains themselves (Fig 2D). We localized our core MCC genes onto the TAD map using H3K4me3 binding at their promoters (Fig 2E) or the positions of the core MCC TSS’s themselves (S8 Fig), finding that, like all genes, core MCC genes were enriched at TAD boundaries, rather than located within domains.
The enrichment of MCC genes at TAD boundaries and far from the bulk of enhancers led us to examine how many enhancers were near these genes in greater detail. A single collection of all peaks with active histone marks (H3K4me3 and H3K27ac) was generated and assigned to genes with the closest TSS (S15 Table), based on the assumption that this approach estimates local enhancer density rather than attempting to predict targets of every enhancer. Using this approach, we found an average of 2.84 active peaks closer to a given gene than any other gene across the entire genome. Genes with the highest number of flanking active peaks (> 4) tended to be TFs based on GO term functional analysis (“transcription factor activity”, p value 1.02E-20 for genes with more than 4 flanking active peaks, S15 Table). To confirm this, we checked known TFs (Xenopus paralogs of 1,692 human TFs[32]) and found they had 4.04 active flanking peaks on average and were positioned inside TADs, where enhancer density was highest (S9 Fig). Conversely, genes with a single nearby active peak were enriched in GO terms such as “protein transport” (p value = 1.7E-6), and “ribosome”, (p value = 6.1E-5). Known housekeeping genes (Xenopus paralogs of 3,804 human genes[33]) contained 2.40 active flanking peaks on average, and these genes were also enriched at TAD boundaries rather than within them (S9 Fig). Finally, in a general survey, all genes with few enhancers largely resided at the boundaries, while genes with many enhancers were depleted at boundaries but instead located within domains themselves (Fig 2F). These results are consistent with the idea that housekeeping genes are located at TAD boundaries where enhancer density is low, while developmental regulated genes, such as TFs, are located in areas of high enhancer density within TADs. In this light, it is striking that the MCC-expressed genes averaged few peaks per gene (2.42) and were enriched at TAD boundaries (Fig 2E), similar to housekeeping genes. Thus, these data indicate that MCC genes, although expressed in a cell type-specific manner, resemble housekeeping genes, raising the question of how these genes become activated in this configuration.
Previous studies have found that proximal regulatory regions of genes expressed in motile ciliated cells are enriched in Rfx binding sites[34,35] and that in Xenopus, Rfx2 binding is found extensively in association with cilia genes.[22] Consistent with these observations, the binding site for the Rfx proteins was by far the most enriched motif identified when sequences (+/-100 bases) around the TSS of the core MCC genes (verified by H3K4me3, S6 Fig) were examined for overrepresented motifs using an unbiased approach (Fig 3A).[3] We also detected a similar, striking enrichment of Rfx motifs in the promoters of MCC gene orthologs in human, mouse, zebrafish, fruit fly, and sea starlet (Fig 3B), despite the fact that the core promoter sequences vary substantially across vertebrate and ecdysozoan lineages[36–38] and cnidarian promoters as well (S10 Fig). By contrast, binding motifs for other factors involved in MCC differentiation, including Forkhead, E2f, and Myb, were also enriched at MCC gene promoters but this enrichment was less impressive, especially at MCC genes in other species (Fig 3A, S11 Fig).[20,22,39–41] These data support deep conservation of direct Rfx factor binding at the promoters of MCC genes but also raise the possibility that other factors are recruited to these promoters indirectly.
To gain further insight into the TF binding required for MCC gene expression, we examined the binding targets of TFs recognizing the motifs enriched in MCC promoters: Rfx, Fox, E2f, and Myb. We interrogated recently published X. laevis ChiPseq data for Rfx2[22] and E2f4[40] and generated ChIPseq data for two additional candidate factors, Myb and Foxj1, using X. laevis epithelial progenitors manipulated to increase the number of motile cilia (see Methods, Fig 3C). We mapped reads from all datasets to the X. laevis genome (v9.1),[42] called peaks against input background, and subjected all peaks to de novo motif discovery. One of the top motifs identified in each set corresponded to the motif recognized by the factor immunoprecipitated, suggesting that these family members often bound the appropriate motif directly (Fig 3D, S12 Fig). However, this analysis also suggests possible co-binding; Foxj1 peaks at MCC genes were highly enriched for Rfx motifs, for example, whereas we saw no such enrichment across all promoters or enhancers (Fig 3D, S6 and S7 Figs).
We exploited the extensive binding data at promoters described above to assess which combination of TFs might account for the upregulation of MCC gene expression (Fig 3E). We found striking heterogeneity in the number of promoters bound by any one factor; likewise, we found heterogeneity in how often different combinations of factors bound to core MCC promoters. For example, Rfx2 bound a majority of these promoters (579 out of 693 MCC promoters bound by any of these factors), and other factors in this study rarely bound MCC promoters in its absence (114 out of 693, note extensive Rfx2 binding in Fig 3E). The largest overlap between two factors occurred between Foxj1 and Rfx2 (400), the next between E2f4 and Rfx2 (271), and finally the smallest overlap occurred between Myb and Rfx2 (230). A large fraction of core MCC promoters were not bound by any of these factors (257, or ~27%); however, we note that ~17% of MCC promoters were H3K4me3- (159/950), suggesting that their annotated TSS’s were incorrect.[43] Manual inspection of a subset of these revealed unannotated 5’ exons, and thus TSS’s, outside of our 2kb TSS window based on Mayball annotations (S13 Fig). These misannotations likely led to an underestimation of TF binding at the promoters of the core MCC gene list.
We asked whether binding of MCC transcription factors at core MCC promoters was associated with increased transcription, and examined transcriptional fold-changes of target genes between conditions repressing MCCs (Notch-icd) and conditions promoting MCCs (Notch-icd and Multicilin)(Fig 3F) (We did not examine enhancers with this approach: enhancers are more likely to influence transcription of nearby rather than distal genes[3] but often skip over the nearest TSS,[44] making target promoter prediction difficult). While Rfx2 alone bound many promoters across the genome, few of these genes exhibited increased expression in MCCs; neither did genes whose promoters were bound by E2f4 alone, Myb alone, or either co-bound with Rfx2 (Fig 3F, S14B and S15B Figs). By contrast, Foxj1 modestly increased expression of genes when bound to promoters alone and strongly increased expression when co-bound to promoters with Rfx2 (Fig 3F). While known to be involved in MCC differentiation,[41] Myb did not bind to many MCC promoters (Fig 3E), did not have a motif in MCC promoters nearly as often as Rfx or Forkhead proteins (Fig 3A), and was not nearly as associated with MCC gene transcription in promoters it did bind to (S14 Fig), suggesting Myb plays a more minor role in the differentiation of these cells. Other factors may also be involved, such as TP73; however, TP73 is not reported to strongly prefer MCC promoters (~3% of putative MCC genes in mouse),[45] nor was the TP73 consensus motif strongly enriched in X. laevis MCC promoters (Fig 3A), suggesting it may regulate more potent drivers of MCC genes such as Foxj1 rather than large numbers of MCC genes directly. Taken together, these genome-wide analyses largely support the current model that binding of both Foxj1 and a Rfx factor at a given promoter leads to potent activation of gene expression in MCCs while other combinations or individual factors were much less likely to do so.
The coincidence of Foxj1 and Rfx2 binding at MCC genes was evaluated further by examining this binding in detail across the genome. We first estimated whether co-binding might occur simply by chance by evaluating the level of random overlap within areas of open chromatin (n = 1000 iterations), as defined by H3K4me3 and H3K27ac-positive regions. While random overlap of open chromatin predicts co-binding in 2208 peaks (SD 37), actual overlap between Rfx2 and Foxj1 binding occurred significantly more often (3505, p < 0.001, see Methods). We next evaluated where Foxj1 binding occurs in the genome in relation to Rfx2 binding. Foxj1 was rarely bound at promoters where Rfx2 binding was absent (~10%, 1203/11800 peaks, Fig 4A), but Foxj1 was often at promoters where Rfx2 binding also occurred (~42%, or 1487/3505 peaks). To determine if Rfx2/Foxj1 co-binding at promoters happens more often than by chance, we found that a random collection of Foxj1 and Rfx2 peaks (equal to the number of co-binding peaks, half Foxj1, half Rfx2) bound to promoters less often (~36%, or 1274 peaks, SD 27), suggesting that when cobound with Rfx2, Foxj1 was more likely to prefer promoters (p < 0.001, see Methods). Taken together, these data suggest that Foxj1 binding is strongly influenced by the present of Rfx2 binding, particularly at promoters.
We next compared the peaks of Foxj1 and Rfx2 binding to the location of TADs within the Xenopus genome. As a factor more likely to be found at distal sites, Foxj1 peaks were enriched inside domains (much like H3K27ac, Fig 2D), as compared to Rfx2 peaks (Fig 4C), which were enriched at their boundaries (similar to promoters, Fig 2C and 2D). When we examined the numbers of sequencing tags immunoprecipitated from these experiments, however, we saw a striking enrichment for both Rfx2 and Foxj1 at domain boundaries (Fig 4C), suggesting that while there were fewer Foxj1 peaks at TAD boundaries, those present were exceptionally strong. These results further indicate that Foxj1 binding at MCC genes located at TAD boundaries is associated with the presence of Rfx2 binding.
The simplest model to explain co-binding of Foxj1 and Rfx2 at MCC promoters is based on the presence of neighboring direct binding sites with appropriate consensus motifs, as shown in a handful of Drosophila cilia genes.[15] However, when we examined the enrichment of Forkhead and Rfx motifs in the promoters and distal sites bound by Foxj1 (Fig 4B, also see Fig 3D), we found that Rfx motifs were strongly present to an equal degree in promoters and distal sites, but that Forkhead motifs were depleted in promoters relative to distal sites. Moreover, when we performed de novo motif analysis on Foxj1 peaks not co-bound with Rfx2, we again saw robust enrichment of Rfx motifs, suggesting other Rfx proteins may also co-bind with Foxj1 (S17 Fig). Finally, despite Foxj1’s known conserved role in motile cilia formation,[20,35,39,46,47] Forkhead motifs were not particularly conserved in the promoters of MCC genes across vertebrate lineages (S11 Fig). Thus these findings suggest that the while Foxj1 is required to drive this process, consensus Forkhead motifs within promoters may be dispensable.
As there was a strong overlap between Foxj1 and Rfx2 peaks, and few Fox motifs at Foxj1-bound promoters (Fig 4B), we asked if Rfx2 might stabilize Foxj1 binding at these positions. Moreover, as many Foxj1 peaks were within TADs (Fig 4C), we speculated that Foxj1 at distal sites might be recruited to promoters, possibly at the TAD boundaries, in chromatin loops and that Rfx2 might facilitate this recruitment. To directly test the possibility that Rfx2 stabilized Foxj1 binding, we knocked down Rfx2 with a well-characterized morpholino[22,48] and performed ChIPseq on Foxj1 (Fig 5). Out of 15,305 Foxj1 peaks identified in control tissue, some 6,360 were reduced at least 3-fold (~42%), and 2,084 were reduced 10-fold (~14%) in an Rfx2 knockdown. We saw strong reductions at key promoters, such as the rfx2 promoter itself (Fig 5A), while other promoters maintained Foxj1 occupancy (Fig 5B). We also saw strong reductions in regions with many peaks, such as the superenhancer surrounding tubb2b (Fig 5C); in cases where reduced Foxj1 peaks did not overlap with robust Rfx2 peaks we often saw Rfx motifs (Fig 5C) and we also found Rfx motifs enriched in Foxj1 peaks not cobound by Rfx2 (S17 Fig), further hinting that other Rfx proteins might be involved. When we looked across all regions bound by Foxj1, we saw a reduction of Foxj1 tags at those positions in the absence of Rfx2 and striking reduction of Foxj1 tags at the promoters of MCC genes (Fig 5D). By contrast, we saw a slight increase at non-MCC promoters, which may be the result of excess Foxj1 protein binding nonspecifically to regions of open chromatin.[49,50] While there is a large fraction of Foxj1/Rfx2 cobound sites in promoters that are not upstream of genes in our core MCC list (3025 out of 3505), we found the greatest Foxj1 reduction at core MCC promoters (S18 Fig).
Rfx2 might stabilize Foxj1 at sites where Foxj1 binds directly to its consensus sequence in DNA, or alternatively, it might stabilize Foxj1 at positions where Foxj1 binds indirectly. To distinguish between these possibilities, we examined changes in Foxj1 binding in peaks occupying positions containing either a strong Forkhead motif but not a Rfx motif (2,646 peaks) or a strong Rfx motif but not a Forkhead motif (2,083 peaks). We found that while tag counts in positions containing a Forkhead motif alone were modestly reduced in the absence of Rfx2, Foxj1 tag counts at positions with an Rfx motif alone were drastically reduced (S18 Fig), suggesting that Rfx2 might have a minor role in stabilizing Foxj1 directly but a much larger role in stabilizing it indirectly.
The increased likelihood of transcription of promoters bound by multiple factors is consistent with general findings by others[3,11] but does not explain the distribution of Foxj1, Rfx2, and their respective motifs relative to transcriptional start sites. Rather, the results above suggest a model where, in some cases, Rfx proteins bind at both promoters and distal enhancers, and through dimerization,[51,52] maneuver the two together to effect transcription. We further speculated that this looping enables Foxj1, bound to enhancers, to localize to promoters, where it was then crosslinked and immunoprecipitated in our assays. This proposed looping would create a protein-DNA complex wherein Foxj1 bound enhancer DNA through a motif but was held nearby its target promoter DNA via Rfx2 proteins: ChIP of a single Foxj1 protein would then pull down both its directly-bound enhancer and its indirectly-bound promoter. To evaluate this model, we first determined all significant interactions in our 3D chromosomal conformation data from wild-type epithelial progenitors using an approach published previously (see Methods, “TCC significant interactions and TADs”).[25] We then interrogated the anchor points of these interactions for co-binding of Rad21, surmising that the anchor points of significant chromosomal interactions would be enriched for known looping machinery.[28,53] To test the probability of overlap between positions bound by Rad21 and the anchor endpoints of significant interactions, we performed a hypergeometric distribution test, calculating the amount of overlap by chance alone, the amount of enrichment over this amount measured in our data (if any), and the probability of encountering this enrichment.[3,54] Pairs of Rad21 peaks were found strongly enriched at the endpoints of loop anchors (Rad21 self-interacting loops in Fig 6A), in agreement with [53]. We then applied this approach to other peaksets, finding strong enrichment for pairs of Rfx2 peaks at the endpoints of looping interactions, consistent with our model that Rfx2 could mediate such interactions. We also saw pairs of Foxj1 enriched at the ends of loops, as well as Rfx2-dependent Foxj1 peak pairs (labeled as “F3”, for peaks reduced by 3-fold or more in the absence of Rfx2). There was also a strong enrichment for pairs of MCC promoters at the endpoints of loops. Finally, we asked if the peaksets enriched at anchor endpoints changed during MCC differentiation, by comparing wild-type interaction data with data obtained with progenitors expressing Multicilin (Fig 6B). We saw an increase in enrichment of both Foxj1 peaks and Rfx2-dependent Foxj1 peaks at anchor endpoints, and an even stronger enrichment of endpoints with Rfx2-dependent Foxj1 peaks at one end and MCC promoters at the other end (connection between “F3” and “MCC”). Taken together, these data suggest a model wherein Foxj1 peaks bound to distal enhancers are recruited to MCC promoters via Rfx2 dimerization, a process that coordinates or stabilizes chromatin loops (Fig 6C and 6D).
TFs required in epithelial progenitors to drive gene expression during MCC differentiation have been identified, but how these factors cooperate at both individual positions in the genome and also within larger chromosomal topologies are largely unknown. Here, using a battery of genomic approaches, many for the first time in X. laevis, we show that genes involved in MCC differentiation typically reside at TAD boundaries and are activated by a combination of Foxj1 and Rfx2. Our data suggest a model where Foxj1, often bound directly to flanking enhancers, is recruited to the promoters of MCC genes via Rfx2, which acts as a scaffolding factor (Fig 6C–6E). We speculate that this arrangement facilitates terminal differentiation by maintaining stable activation of gene expression.
We show, using an unbiased approach,[3] that the MCC core promoters in Xenopus and other metazoans are highly enriched for motifs recognized by the Rfx factors, also known as the X-box, in line with a similar analysis of genes differentially expressed in lung tissue from patients with primary ciliary dyskinesia.[55] While the importance of Rfx binding sites is well established for cilia genes in flies and worms,[12] their exact role in promoting cilia gene expression in vertebrates is less clear, mainly because the family of Rfx factors has expanded significantly, members of this family appear to act redundantly, and the phenotype associated with a loss of any one family member may differ depending on the species (e.g. Rfx2 in mouse and Xenopus).[22,48,56] In addition, the expanded Rfx family in vertebrates clearly has major roles in enabling gene expression that serves non-cilia functions,[57] based on single family member mutants (S2 Fig) as well as compound mutants.[34] The Rfx family also appears to have broad functions based on the high occurrence of Rfx binding motifs within non-coding regions that are conserved within mammalian genomes.[58] Thus, the analysis of MCC gene promoters suggests that Rfx factors are major players in MCC gene expression but other observations suggest they act by facilitating the action of other TFs. If so, what factors are involved and by what mechanism?
To address this question, we generated ChIPseq data for Myb and Foxj1, two other regulators required for MCC differentiation whose binding preference in epithelial progenitors was unknown. These data, combined with the previously published ChIPseq data for E2f4 and Rfx2,[22,40] indicate that among the various combinations, the co-binding of both Foxj1 and Rfx2 at a promoter is the best predictor of activated expression during MCC differentiation. By examining the nature of transcription factor co-binding in our data, we uncovered several important features of this synergistic interaction at MCC promoters. First, ChIPseq data indicate that Foxj1 alone preferentially binds to distal sites, but shows a much higher preference for promoters when co-bound with Rfx2. Second, Foxj1 binding at both locations are highly enriched in Rfx motifs, but Foxj1 sites at promoters rarely have strong Forkhead motifs. This may be explained by fixation: we suspect that chromatin immunoprecipitation of individual proteins yields genomic fragments both directly and indirectly bound to that protein due to crosslinking; if so, ChIP of a single Foxj1 molecule might produce fragments from both a distal enhancer and a promoter, and we predict direct binding at fragments with motifs and indirect association at fragments without. Third, Foxj1 binding at MCC promoters and at sites lacking Forkhead motifs is severely disrupted when Rfx2 function is impaired. Taken together, these data indicate that Foxj1 transcriptional activity, especially at MCC promoters and sites with poor motifs, is directed by, and dependent on, Rfx2. These data lead us to propose that Rfx2, perhaps with other Rfx proteins, performs a scaffolding role wherein distal Foxj1-bound enhancers are maneuvered to promoters by Rfx. This scaffolding function may cause looping directly since the Rfx proteins can form homo- and heterodimers with one another,[51,52] but one can envision more complicated scenarios where Rfx functions within a larger protein complex required for looping and Foxj1 recruitment or Foxj1 binding DNA indirectly through Rfx proteins and dispensing with Fox motifs entirely. While further work will be required to determine if and how the 3D genome is perturbed when Rfx function is lost and how different Rfx proteins might cooperate to affect its conformation, this scaffolding model is consistent with extensive functional data indicating that Foxj1 is the critical limiting factor required for MCC gene expression,[39,59] explains why Rfx motifs and binding are so frequently encountered at MCC promoters, and is largely consistent with other examples where Rfx and Forkhead factors have been shown to interact to promote gene expression, not only in cells with motile cilia[14,15] but also in other contexts.[60] Also consistent with a scaffolding, and not a transcriptional, role, recent work also suggests that Rfx proteins are unable to rescue ciliogenesis in the absence of Foxj1.[45] Moreover, the scaffolding function of the Rfx proteins proposed here may be more general, enabling enhancer and promoter interactions not only during MCC cell differentiation, but in other contexts as well where this family of proteins is known to act.
Chromatin looping is thought to mediate promoter-enhancer interactions, bringing two or more elements close to one another in topological space.[5,61,62] Here, TCC data support the presence of such loops maneuvering Foxj1 to MCC genes in that there is enrichment of Foxj1 binding sites of MCC promoters at opposite anchor endpoints of chromatin loops in MCC cells. This finding, along, with the observation that MCC genes typically reside at TAD boundaries, may be significant in the context of terminal differentiation. The boundaries that define TADS have relatively stable anchor points, contain a low density of local enhancers, and are highly enriched in transcriptional machinery.[26,53,63,64] These properties may explain why TAD boundaries also contain a high density of housekeeping genes that are constitutively expressed across most cell types.[64] Thus, we speculate that positioning of certain MCC genes at TAD boundaries may facilitate stable gene expression associated with terminal differentiation, specifically that associated with the maintenance of the multiple motile cilia in the long-lived MCC.
This arrangement may also provide a modular framework for transcriptional control. While core promoter machinery may impose sequence constraints on GC-rich promoters that are incompatible with AT-rich motifs such as those favored by Foxj1, distal enhancer sequences may be more relaxed.[65] Thus, distal enhancers may be free to contain a wider range of motifs, and, as long as they also contain flanking Rfx motifs, such elements can be recruited to promoters via Rfx dimerization. As the various Rfx family members may have slightly different motif preferences, one can also speculate that such an arrangement provides for dynamic transcriptional control across tissues and conditions depending on which Rfx proteins are expressed: a distal enhancer in one cell type may get recruited to its target promoter, whereas a different enhancer might get recruited to the same promoter in another cell type.
All animal experiments reported here were carried out at the Salk Institute, supervised by a licensed veterinarian in the Animal Resource Department, under guidelines that are AAALAC certified, and using procedures reviewed and approved by the Salk IACUC committee under protocol #15–00034.
X. laevis embryos were prepared by in vitro fertilization using standard protocols.[66] Synthetic, capped mRNA was generated in vitro using DNA templates described below and injected into embryos at the two-cell or four-cell stage (typically 0.1–5.0 ng of RNA per embryo), targeting the four animal quadrants. Pluripotent epidermal progenitors (animal caps) were isolated at stage 9–10 and cultured in 0.5X MMR and harvested at 3, 6, and 9 hours at 22 deg after mid-stage 11. Progenitors injected with RNA encoding HGR-inducible constructs (Multicilin-HGR and Foxi1-HGR) were induced with 1μM dexamethasone at mid-stage 11.
DNA templates for synthesis of mRNAs have been described previously for Notch-icd, dbm, Foxi1-HGR, Multicilin-HGR, and dominant negative Multicilin.[18,19,67] The DNA template for generating an RNA encoding a tagged form of Foxj1 was obtained by subcloning a cDNA encoding Foxj1[39] into a CS2 vector with a FLAG tag, while that for a tagged form of Myb was generated by PCR amplification of a myb cDNA from a stage 17 cDNA library and subcloning into a CS2 vector with a GFP tag. The Rfx2 morpholino[22,48] was a kind gift of Meii Chung and John Wallingford.
RNAs were isolated by the proteinase K method followed by phenol-chloroform extractions, lithium precipitation, and treatment with RNase-free DNase and a second series of phenol-chloroform extractions and ethanol precipitation. RNAseq libraries were then constructed (Illumina TruSeq v2) and sequenced on an Illumina platform. Details on specific experiments are in S17 Table, and RNAseq reads are deposited at NCBI (GSE76342).
Sequenced reads from this study or obtained previously[22,40] were aligned to the X. laevis transcriptome, MayBall version[22] with RNA-STAR[68] and then counted with eXpress.[69] Effective counts from eXpress were then clustered (Cluster 3.0[70], log transformed, maxval—minval = 6) and visualized with Java Treeview (v1.1.6r, [71]) to produce heatmaps. DESeq was used to estimate dispersion and test differential expression using rounded effective counts from eXpress.[72] Changes in expression were visualized in R with beanplot (https://cran.r-project.org/web/packages/beanplot/beanplot.pdf), and to visualize RNAseq reads in a genomic context they were mapped to genome version 9.1 with bwa mem[73] and loaded as bigWig tracks into the Integrative Genomics Viewer browser.[74]
Orthologs/paralogs of MCC genes were obtained from gene assignments in Homologene or by using official gene symbols in N. vectensis gene models.[75] Promoter collections for H. sapiens, M. musculus, D. rerio, and D. melanogaster were obtained from UCSC. To determine N. vectensis promoters, 5’ ESTs were mapped to JGI genome build v1.0 with BLAT.[76] H3K4me3 ChIPseq reads from N. vectensis[77] were then mapped to v1.0, peaks called, and then overlapped with TSS’s as determined by 5’ ESTs to verify promoters, similar to our approach in confirming X. laevis promoters. 5’ ESTs were then aligned to N. vectensis gene models (v1.0) to connect transcriptional start sites to gene identity. Motif enrichment on all promoters and MCC ortholog promoters was done with HOMER.[3] Species divergence times are from Timetree.[78]
Because ChIP-grade antibodies are generally not available that recognize Xenopus proteins, we tagged Myb and Foxj1 with GFP and FLAG, respectively, and injected mRNAs encoding these constructs into embryos. Myb has functions in stem cell maintenance independent of a role in MCCs,[79] so to enrich for motile cilia targets, we performed ChIPseq on Myb in progenitors injected with Multicilin-HGR induced at mid-stage 11 (our previous work with E2f4 ChIPseq[40] also included Multicilin injections, allowing for direct comparison of MCC-enriched TF targets). Moreover, overexpression of Foxj1 promotes ectopic cilia[39], so samples injected with that construct were also enriched for motile cilia targets by virtue of expressing a tagged Foxj1 construct.
Samples were prepared for ChIP using described methods[40] with the following modifications: About 250 animal caps for TFs or 100 caps for histone modifications were fixed for 30 min in 1% formaldehyde, and chromatin was sheared on a BioRuptor (30 min; 30 sec on and 2 min off at highest power setting). Tagged proteins with associated chromatin were immunoprecipitated with antibodies directed against GFP (Invitrogen catalog no. A11122, lot no. 1296649) or FLAG (Sigma, cat #F1804). Native proteins were immunoprecipitated with antibodies directed against H3K4me3 (Active Motif, cat #39159; lot #01609004), H3K27ac (Abcam, cat #ab4729, lot #GR71158-2), or rad21 (Abcam, cat #ab992; commercially-available CTCF antibodies target regions not conserved in the X. laevis protein). DNA fragments were then polished (New England Biolabs, end repair module), adenylated (New England Biolabs, Klenow fragment 3′–5′ exo- and da-tailing buffer), ligated to standard Illumina indexed adapters (TruSeq version 2), PCR-amplified (New England Biolabs, Phusion or Q5, 16 cycles), and sequenced on an Illumina platform. Details on specific experiments are in S17 Table, and ChIPseq reads are deposited at NCBI (GSE78176).
ChIP-seq reads from this study or published previously[22,40] were mapped to X. laevis genome v9.1 with bwa mem,[73] peaks called with HOMER[3] using input as background. Peak positions were annotated relative to known exons (Mayball gene models[22]), with promoters defined as being +/- 1Kb around the TSS. Peak sequences were interrogated for de novo motif enrichment with HOMER and MCC promoters were clustered (based on if they were bound/not bound) with Cluster 3.0 and visualized with Java Treeview (v1.1.6r). Tags or motifs were counted at peak positions with HOMER and plotted with Excel or R. To determine enrichment statistics, we used the hypergeometric distribution to evaluate the likelihood of peak overlap.[3] Additionally, as regions of open chromatin occupy a fraction of the entire genome, we also assessed overlap between pairs of factors by determining the distribution of randomly picked open chromatin sites (equal to the number of sites in the first and second TFs in the comparison; open chromatin was estimated by using all H3K4me3 and H3K27ac sites[80]) and repeated this approach 1000 times to get a distribution of overlap between two sets of random regions of open chromatin. We then compared this distribution to the overlap observed in pairs of measured TFs, again restricted to open chromatin (as in Fig 4 and S14 and S15 Figs). To determine if subsets of overlapped peaks preferred promoters or other features more than by chance, we employed a similar approach: instead of selecting overlapped peaks, we sampled from all peaks, half from each set, and determined how many peaks form this null set bound to promoters. We repeated this operation 1000 times to obtain the distribution of promoter preference from these random combinations. Finally, when determining differential binding and generating bigwig files for visualization, each experiment was normalized to 10M reads in HOMER-style tag directories to account for differences in sequencing depth.
TCC was performed as described[81] with the following modifications: 100 animal caps were harvested at the 9 hour timepoint and fixed for 30 minutes in 1% formaldehyde. Tissue was thawed, rinsed in PBS, and lysed directly in 250 μl ice-cold wash buffer with SDS added (50 mM Tris.HCl pH = 8.0, 50 mM NaCl, 1 mM EDTA, 0.56% SDS) rather than using lysis buffer. A complete protocol is available on request. Details on specific experiments are in S17 Table, and TCC reads are deposited at NCBI (GSE76363).
TCC reads were cleared of PCR duplicates, trimmed down to exclude MboI restriction sites (GATC) and mapped to the X. laevis genome v9.1 with bwa mem. Topological domain peakfiles, interaction matrices, and metagene plots were generated with HOMER (additional details below). Matrices were visualized with Java TreeView.
All TADs and significant interactions were called against a background of expected interactions. This background model was constructed using the approach described in [25]. Briefly, the frequency of proximity-ligated fragments that are connected across a certain distance on a linear chromosome is expected to increase as that distance shrinks. To generate the model, the genome is broken into bins with a sliding-window approach to boost signal, and sequence depth in those regions is normalized to account for mapping errors and biases in restriction enzyme accessibility. Once depth is normalized in all regions, the expected variation in interaction between regions is calculated as a function of distance.
As the background model estimates how many interactions between regions are expected given a specific linear distance and sequencing depth, we looked at the number of interactions we actually observed between regions in the TCC data. To determine interactions with some measure of statistical confidence, we used the cumulative binomial distribution, where the number of trials was the total number of reads mapping within regions, success was the expected interaction frequency, and observed successes came directly from read pairs connecting regions from the TCC data itself. We then used the hypergeometric distribution to ask if the anchor points of these interactions overlapped with ChIPseq peaks or other genomic features more than would be expected by chance (Fig 6A) and visualized these values with Cytoscape.[82] We also determined interactions that were stronger in wild-type or Multicilin-injected epithelial progenitors: after determining significant interactions for one of these conditions, we used TCC reads from the other to score their strength (Fig 6B). TADs were determined by calculating a directionality index with HOMER; we observed similar numbers of TADs with varying resolutions (10–25 kb). All methods in this section were adopted from [25].
|
10.1371/journal.pgen.1007503 | No unexpected CRISPR-Cas9 off-target activity revealed by trio sequencing of gene-edited mice | CRISPR-Cas9 technologies have transformed genome-editing of experimental organisms and have immense therapeutic potential. Despite significant advances in our understanding of the CRISPR-Cas9 system, concerns remain over the potential for off-target effects. Recent studies have addressed these concerns using whole-genome sequencing (WGS) of gene-edited embryos or animals to search for de novo mutations (DNMs), which may represent candidate changes introduced by poor editing fidelity. Critically, these studies used strain-matched, but not pedigree-matched controls and thus were unable to reliably distinguish generational or colony-related differences from true DNMs. Here we used a trio design and whole genome sequenced 8 parents and 19 embryos, where 10 of the embryos were mutagenised with well-characterised gRNAs targeting the coat colour Tyrosinase (Tyr) locus. Detailed analyses of these whole genome data allowed us to conclude that if CRISPR mutagenesis were causing SNV or indel off-target mutations in treated embryos, then the number of these mutations is not statistically distinguishable from the background rate of DNMs occurring due to other processes.
| The ability to precisely modify the genome has immense therapeutic potential and also represents a powerful tool to understand gene function. One of the key gene editing tools is CRISPR (Clustered Regularly Interspaced Short Palindromic Repeats). Although CRISPR can be used to generate a range of genomic alterations or to correct disease causing mutations, concerns over unwanted alterations introduced by poor editing fidelity, so-called “off-targets”, have tempered the use of this technology. Here we use whole genome sequencing to show that off-target mutations are rare and not distinguishable in frequency when compared to naturally occurring mutations.
| CRISPR-Cas9 technologies have transformed genome-editing of experimental organisms and have immense therapeutic potential. Despite significant advances in our understanding of the CRISPR-Cas9 system, concerns remain over the potential for off-target effects. Recent studies have addressed these concerns using whole-genome sequencing (WGS) of gene-edited embryos or animals to search for de novo mutations (DNMs), which may represent candidate changes introduced by poor editing fidelity[1–3]. Critically, these other studies used strain-matched, but not pedigree-matched controls (i.e. the parents of sequenced offspring were themselves not sequenced) and thus were unable to reliably distinguish generational or colony-related differences from true DNMs. Here we assessed the impact of colony variation on the accuracy of detection of off-target CRISPR editing events.
In this study, we used a trio design and whole genome sequenced 8 parents and 19 embryos, where 10 of the embryos were mutagenised with well-characterised gRNAs targeting the Tyrosinase (Tyr) locus (Fig 1A). Tyr is responsible for black coat colour and eye pigmentation in C57BL/6 mice[4], so its disruption should not be detrimental to embryonic development. We chose two gRNAs targeting exon 2 of Tyr, Tyr2F and Tyr2R, with 395 and 1,502 total predicted off-targets respectively (Methods). The CRISPR-treated group was split to include five embryos treated with Tyr2F and five embryos treated with Tyr2R, while three untreated embryos from each of three control groups (“Cas9 only”, “No injection” and “Sham injection”) were also collected. Microinjections were performed into the cytoplasm of 1-cell zygotes[5], which were then briefly cultured to assess viability and then transferred into 0.5 day post coital (d.p.c) pseudopregnant females. Embryos in the “Sham injection” group were microinjected with water only, and the “Cas9 only” embryos were microinjected with Cas9 protein solution only. All embryos were harvested at 12.5 d.p.c (S1 Table) and genomic DNA from both parents and embryos extracted. Sequencing was performed on the Illumina X10 WGS platform yielding a median sequencing depth of 39.5x per genome (Fig 1B, S2 Table). In parallel, targeted Illumina MiSeq sequencing to a mean depth of 10,800 reads of the Tyr target site was performed to comprehensively profile mosaicism, with these data analyzed using the CRISPResso software[6] (S3 Table). MiSeq analysis revealed a targeting efficiency of 88% for Tyr2F and 91% for Tyr2R and mosaicism, with a median of two variants per embryo (S3 Table). Importantly, to ensure our experiment was representative of the many thousands of CRISPR experiments performed worldwide, including those of the International Mouse Phenotyping Consortium, we compared these data to MiSeq data from 324 mice mutagenized using the Tyr2R gRNA, revealing good concordance of targeting efficiency and mosaicism (S3 Table).
We next performed variant calling on the WGS data using bcftools mpileup and bcftools call[7], configured to be sensitive to low allele-fraction indels (insertions/deletions)(Fig 1C). We started with a median of 324,561 variants per sample (S4 Table). Candidate DNMs (those not inherited from either parent) in the embryos were called using the TrioDeNovo software[8], which resulted in a median of 7,460 unfiltered DNMs per embryo.
We first looked for the presence of these unfiltered DNMs within 10bp of any potential CRISPR off-target site for Tyr2F or Tyr2R, as determined by the Cas-OFFinder software[9] with up to 3 mismatches with a 1 nucleotide DNA/RNA bulge or up to 4 mismatches without a DNA/RNA bulge. Importantly, we found no such coincident sites in any embryos from the CRISPR-untreated group and only on-target variants in the CRISPR-treated group (S5 Table), suggesting that if there are recurrent CRISPR-induced off-target alterations they are exceedingly rare. When we extended the number of mismatches up to 7 nucleotides (without a DNA/RNA bulge), we did find between 4 and 14 coincident sites per animal, but there was no significant difference in the number of such sites between CRISPR-treated and untreated groups (S5 Table, Methods). In fact, given the total fraction of bases at each possible off-target site and the number of candidate DNMs, we expected to find approximately 32 intersections between our candidate DNMs and the expected off-target positions based on chance alone (Methods). This is higher than the observed intersection counts, but is of the same order of magnitude.
We then filtered the original calls to ensure adequate depth (10x) and a variant quality of 10, in all samples, resulting in a median of 225,671 variants per sample and a median of 6,852 TrioDeNovo called DNMs per embryo (489 SNVs / 6,450 indels) (S4 Table). We next applied a validated filtration strategy to refine our candidate DNM calls[10], removing false positives arising from mosaic alleles in either parent, as well as those in proximity to repeats. Alignments for all SNVs and indel variants were then inspected visually for the presence of mosaic alleles (i.e. a second alternative allele at the same locus). In the same way, all indels were visually inspected to remove further false positives. This resulted in a median of 19 SNVs and 1 indel per embryo (Table 1), which is broadly consistent with prior work that aimed to classify and estimate the de novo mutation rates in mice[10]. All variants were validated using targeted MiSeq sequencing to a depth of at least 10,000 reads per locus (S8 Table), yielding an 87% validation rate for SNVs and a 73% validation rate for indels. This is lower than comparable rates from the Mouse Genomes Project[11], but this is expected, given the lower variant allele fraction of our variants. These rates are broadly comparable to validation rates for somatic variants called from cancer genomes.
A comparison of the expected variants at the Tyr locus detected by WGS and targeted MiSeq sequencing (S6 Table) shows that of the 20 indels detected by the MiSeq pipeline, 18 were also detected by the WGS pipeline, with the missing indels having low allele frequencies (7% and 7.5%, as defined by MiSeq sequencing). We are therefore confident that our WGS pipeline will detect genome-wide off-target damage with a range of allele fractions and mosaicism similar to on-target variants. Our median on-target variant allele fraction is 0.28 (S6 Table, column I), corresponding to a heterozygous mutation occurring at the two-cell embryo stage. Given our median depth (39.5x) and our minimum required de novo allele frequency (10%, Methods), our expected power to detect a DNM occurring in the single-cell or two-cell stage of the zygote is predicted to be at least 99.5%.
Using the final counts of filtered SNVs and indels for each embryo (Table 1), we conducted a Kruskal-Wallis Rank test, detecting no significant difference in DNM counts between the “no injection”, “sham” and “cas9 only” untreated embryo groups (p = 0.30 and p = 0.37 for SNVs and indels, respectively). Similarly, a Wilcoxon Rank Sum test failed to detect significantly different SNV- or indel- DNM counts between the “Tyr2F” and “Tyr2R” CRISPR-treated groups (p = 0.25 and p = 0.43 for SNVs and indels, respectively). Based on these analyses (Fig 1D), we combined variant calls from embryos in the two CRISPR-treated groups and in the same way combined data from the three untreated groups. Notably, using these data a Wilcoxon Rank Sum test failed to detect a significant difference in SNV or indel counts between the CRISPR-treated and untreated groups; p = 0.30 and p = 0.45, respectively (S7 Table).
We also measured the impact of using unrelated parents on the false-positive DNM rate by deliberately choosing the parents of the Cas9-only embryos when analyzing all embryos in the study; the male parent (CBLT8902) was greater than five generations removed from all other male parents (S1 Fig). Performing a comparable subset of filtrations and comparing variant counts by sample to the correctly analysed embryos at the same filtration point showed a median increase of 66 false variants per embryo (S1 Fig, S9 Table), highlighting the importance of using trios of mice when studying potential off-target rates.
Finally, to investigate the robustness of this result, we reanalyzed all of our genome sequencing data using a combination of two somatic variant callers (searching for SNVs with CaVEMan[12] and indels with cgpPindel[13]). These callers were used to search for SNVs and indels present in the treated and untreated embryos that were not present in the parents. The caller output was subject to an analogous filtering pipeline (Methods). Although not designed to search for DNMs, somatic variant callers can be more sensitive to low variant allele fraction mutations. Indeed, we detected more filtered mutations with this approach than with TrioDeNovo (S10 Table, S11 Table). However, identical statistical tests on the final filtered counts of SNVs and indels (Methods, S12 Table) again failed to detect any significant differences between gRNA-treated and untreated groups.
We conclude that if CRISPR mutagenesis performed under the conditions we have described were causing SNV or indel off-target mutations in treated embryos, then the number of these mutations is not statistically distinguishable from the background rate of DNMs occurring due to other processes. This work should support further efforts to develop CRISPR-Cas9 as a therapeutic tool.
Following approval by the Sanger Institute Animal Welfare and Ethical Review Body (AWERB) all procedures were performed at Wellcome Trust Sanger Institute Research Support Facility under Home Office licensed authority, Establishment Licence Number X3A0ED725, Project licence number P96810DE8. The use of animals in this study has been carried out in accordance with the UK Home Office regulations under the Animals (Scientific Procedures) Act 1986.
We chose two gRNAs (Tyr2F and Tyr2R) within exon 2 of the Tyrosinase (Tyr) gene, as mutations at this locus do not cause lethality or influence embryogenesis. Tyr is responsible for black coat colour and eye pigmentation in wild type (WT) C57BL/6 (B6) mice and bi-allelic mutation of this gene results in complete loss of pigmentation (albinism)[4]. Therefore, detection of biallelic mutations caused by CRISPR/Cas9 activity is easily visible as a change in coat colour.
The Tyr2R gRNA was selected as previously described[14] and together with the Tyr2F gRNA represent a typical range of off-target scores. Off-target scores for the Tyr2F and Tyr2R gRNAs were determined for the default (NGG) PAM only, using the reference mouse genome (mm10) in the Cas-OFFinder tool[9] (Table 2). Off-target scores were also determined for the Schaefer sgRNA#4[15] using the FVB/NJ mouse genome, to ensure that off-target scores were similar. Whilst we are not trying to directly compare our results to the Schaeffer paper [3], we feel that it is important to demonstrate that the off-target scores for our gRNAs are not significantly different to their study. The sequences for the gRNAs used in this experiment are shown in Table 3.
To determine the expected cutting efficiency and mosaicism from a large number of experiments we chose Tyr2R, comparing different Cas9 sources (mRNA or protein) as well as gRNA sources (in vitro transcribed or synthetic). The gRNA was mixed in RNase free water (Ambion) at a concentration of 25ng/ul together with either Cas9 mRNA or protein at 50ng/ul. The CRISPR reagents were injected into the cytoplasm of zygotes and F0 pups were scored for black, mosaic and albino coat colour. In addition, genomic DNA from earclips was extracted from F0 mice as described in section 3, and MISEQ sequencing analysis performed as described in section 4. This enabled us to determine the targeting efficiency and how many different alleles were present within each founder animal, therefore making it possible to score mosaicism (more than one mutated allele detected in the animal) for each condition. The synthetic gRNA was the most efficient, with 70% of pups showing a mutant genotype. Synthetic gRNAs were therefore used for the trio experiment (S3 Table).
For the trio experiments, synthetic gRNA consisting of crRNA and tracrRNA (Sigma) were diluted and mixed in RNase free water at equimolar ratios of 0.7pmol/ul each. Cas9 protein (obtained from Marko Hyvonen, Department of Biochemistry, University of Cambridge) was added to a working concentration of 50ng/ul and the mixture was incubated at 25°C for 10 minutes before zygote injection. The concentrations of reagents injected for each of the experimental groups are shown in Table 4.
DNA was extracted as described in section 3 and MISEQ sequencing analysis performed as described in section 4. The results are presented in S3 Table labelled with the treatment condition (e.g. ‘No injection’, etc), and are used to check that the trio experiment is comparable to historical data at the target locus.
4 x 4-week old C57BL/6NTac females were super-ovulated by intraperitoneal (IP) injection of 5 IU of pregnant mare’s serum (PMSG) at 11.00hrs (12hr light/dark cycle, on at 07:30/off at 19:30) followed 48hrs later by an IP injection of 5 IU human chorionic gonadotrophin (hCG) and mated overnight with C57BL/6NTac stud males. The next morning females were checked for the presence of a vaginal copulation plug as evidence of successful mating and females housed separately, noting which males were used to plug each of the females. Oviducts were dissected one female at a time and harvests of cumulus masses kept separately in 4 different groups. Cumulus masses were released and treated with hyaluronidase as previously described[16]. Fertilized 1-cell zygotes, confirmed by the presence of 2 pronuclei, were selected and maintained in KSOM media prior to cytoplasmic injection at 37°C in 4 separate dishes. Microinjections were carried out between 23–25hrs post hCG. Although the exact cell cycle stage varied from zygote to zygote, microinjections were done prior to coalesence of the pronuclei, and therefore the completion of S-phase. Cytoplasmic Injections were carried out as in S1 Table. The tyrosinase Tyr2F and Tyr2R microinjected embryo groups both came from the same zygote pool and hence had the same parentage. All other microinjection groups had a unique set of parents.
The Cas9 ribonucleoproteins (RNP) were backfilled into a microinjection needle. Microinjections were carried out using positive balancing pressure, microinjecting into the cytoplasm of fertilized 1-cell zygotes held in FHM medium. A successful injection was indicated by visible movement in the cytoplasm after breaking the Oolemma. Microinjected 1 cell embryos were briefly cultured and viable zygotes were transferred the same day by oviducal embryo transfer into a 0.5 days post coital (d.p.c.) pseudo-pregnant female F1 (CBA/C57BL/6J) recipients[16]. After 12.5 d.p.c. recipient mice were humanely culled and embryos dissected and snap frozen. Previously, the parents from each group were humanely culled, tissue taken and labelled according to which microinjection group they contributed to (S1 Table).
All procedures performed in studies involving animals were in accordance with the ethical standards of the institution or practice at which the studies were conducted and performed with approval of the UK Home Office.
For the historical data, genomic DNA was extracted from earclips of F0 mice using the Sample-to-SNP kit lysis buffer (Life Technologies).
For the trio experiments, genomic DNA was extracted from the kidney of parent animals or from macerated whole 12.5 d.p.c. embryos using the DNeasy Blood & Tissue Kit (Qiagen) according to manufacturer's instructions. DNA was quantified using a NanoDrop spectrophotometer.
Note that for the historical data, extracting DNA from ear-clips could bias the estimate of on-target mosaicism, depending on the distribution of mutant cells. The mosaicism rate from historical data is, however, approximately the same as the current mosaicism rate, with approximately 2 on-target variants per embryo (S3 Table). Similarly, the choice to extract DNA from parental kidneys only could bias the de novo variant count for particular pedigrees. However, any bias in a single pedigree should be mitigated by the use of three control groups from different pedigrees. Additionally, the high sequencing depth used when validating all candidate de novo variants would minimise possible bias resulting from subclonal mutations in the parental kidneys.
This analysis was performed on 324 historical mice, and on the parents and embryos in the trio experiment. 1μl of genomic DNA was used for amplicon specific PCR using genome specific primers (PE_tyrex2N_F1 and PE_tyrex2N_R1, Table 5), which flanked the expected on-target sites for Tyr2R and Tyr2F. The indexed libraries were sequenced using standard protocols and Illumina MiSeq technologies (Paired End 250bp runs).
The paired-end fastq files generated were analysed using the CRISPResso sequence analysis package[6]. A typical command line used to run CRISPResso specified both gRNA sequences concurrently and allowed a window of 30bp around the gRNA sites to detect possible mutations: CRISPResso -r1 sample_read1.fastq.fastq -r2 sample_read2.fastq -a <unedited_amplicon_sequence> -w 30 -g <Tyr2R_sequence>,<Tyr2F_sequence>. The resulting sequence variants were visually inspected for indels with an allele fraction of at least 5%.
MiSeq analysis performed on the target region in these samples using CRISPResso showed that the parents, no injection, sham and Cas9 only groups showed no on-target activity, whereas the two gRNAs showed efficient cutting at their target site (S3 Table). Of the 10 mice showing targeting for the Tyr2R gRNA, 3 had one allele (30%), 5 had two alleles (50%) and 2 had three alleles (20%). Of the 7 mice showing targeting for the Tyr2F gRNA, 4 had two alleles (57%), with the remaining three mice containing either one, three or four alleles. This is in broad concordance with the historical data (S3 Table).
In order to cover a representative range of alleles at the on-target site, we selected the following mice for WGS analysis: for the Tyr2R sample 1 mouse with one allele, 3 mice with two alleles and 1 mouse with three alleles; for the Tyr2F sample 1 mouse with one allele, 3 mice with two alleles and 1 mouse with three alleles (in total five embryos Tyr2F and five embyros for Tyr2R).
Whole genome sequencing libraries were prepared using standard protocols for the Illumina X10 platform (ENA Accession ERP024425). The resulting sequence was aligned using bwa mem to the reference mouse GRCm38 assembly. The total mapped coverage varied from 34x to 47x, with a median of 39.5x (Fig 1B). The median fraction of bases with a coverage of greater than 50x was 39% and the median fraction of bases with a coverage of less than 11x was 3.4% (S2 Table).
Our median observed on-target variant allele fraction was 0.28 (median of column I, S6 Table). This allele fraction is represented by 11 variant reads out of 39.5 reads, and most likely arises from a heterozygous mutation in the two-cell embryo. We take this frequency as a representative effect size. To find our power to resolve mosaic DNMs from our WGS data with (median) 39.5x read coverage for this effect size, we modelled the actual number of mutant reads observed with a Poisson distribution with parameter lambda = 11. Since our minimum required variant allele frequency is 10% (Methods Section 7), we must observe at least 4 mutant allele reads out of 39 to call a DNM. The probability of not calling a mutation—seeing 3 or fewer reads—in this case is given by the R function ppois(3,11,lower.tail = TRUE) = 0.005. This suggests that when we sequence our 12.5dpc embryos, we should detect DNMs, which occur in a two-cell embryo (or earlier) at least 99.5% of the time.
Aligned sequence was jointly variant called for all parents and offspring using bcftools mpileup, bcftools call, bcftools norm and bcftools filter[7]. The bcftools version and command options used are as follows: bcftools-1.6 mpileup -a AD -C50 -pm2 -F0.05 -d10000, bcftools-1.6 call -vm, bcftools-1.6 norm -m -any, and bcftools-1.6 filter -m+ -sLowQual -e"%QUAL< = 10" -g3 -G10.
Note: a joint (multiallelic) variant call was performed on all parents and offspring at the same time, using bcftools mpileup configured for sensitivity (requiring only 2 or more indel reads and a minimum indel allelic fraction of 0.05). bcftools called between 317,230 and 332,048 variants per sample with a genotyping quality > = 10, with a median of 324,561 variants (S4 Table). All variant loci were required to have a total depth of at least 10 reads for further analysis.
De novo mutation calling on all variants was performed by running TrioDeNovo software[8], using default settings, independently on each parent/offspring trio. After filtering for depth and quality as above, TrioDeNovo produced between 6,437 and 7,732 candidate mutations per offspring; with a median of 6,852 mutations (S4 Table).
Based on the read depths and variant allele fractions seen in the expected on-target mutations called by TrioDeNovo, we filtered all candidate de novo variants with the following criteria, to remove likely false positives: (1) We required a minimum variant allele fraction of 10% to allow for mosaic alleles. (2) To avoid false positives arising from low-allele-fraction mosaic variants in either parent, we removed variants with any alternate-allele reads present in either parent (“parental noise”). (3) We removed any variant coincident with an allele reported in the C57BL/6NJ strain of the Mouse Genomes Project[11] or specifically in the C57BL/6NTyr strain[17]. (4) To avoid further false positives arising from low-allele-fraction mosaic variants present in the parents, we allowed a maximum contribution of only 2% alternate reads from all other (non-parental) samples (“cross noise”). (5) As repeat regions can cause mis-alignment of reads resulting in false positive calls, we merged all individual UCSC repeat tracks with the UCSC RepeatMasker track (about 1.2Gb of sequence) and removed any variant inside or 1bp adjacent to any merged repeat region. (6) We removed any de novo variant shared by two or more samples, as this would be extremely unlikely, and such mutations are more likely mosaic in the parents. Although it is possible this could remove a preferred off-target mutation, we note this removed only 3 variants. (7) Every variant locus was visually inspected to check whether any position was actually mosaic (i.e. contained two or more alternative variants). These extra alternate variants were not consistently called and had to be manually re-inserted. (8) We found indel variants to be especially susceptible to false positive calls arising from un-annotated microsatellites or repeats, so we visually inspected all indel variants to remove any variant still in or adjacent to microsatellites/homopolymers, which were not annotated by UCSC. These filters resulted in 11 to 29 SNVs per sample (median 19) and 0 to 5 indels per sample (median 1). Every SNV and indel has been listed in S8 Table.
The CAS-OFFinder tool[9] was used to find all potential off-targets sites based on sequence homology to either the Tyr2F or Tyr2R gRNAs, allowing up to 3 mismatches with 1 bp of inserted or deleted sequence and up to 4 mismatches with no inserted or deleted sequence. This resulted in 395 and 1,502 potential off-target sites being detected for Tyr2F and Tyr2R, respectively (S5 Table; RGEN_Tyr2*_OffTargetSites). These 23bp gRNA positions were intersected (allowing for a window size of 10bp) with all candidate DNMs before filtering, using bedtools-2.23 window. The results are presented in (S5 Table, worksheet “RGEN_DNM_10bp_overlap”).
We investigated the coincidence of DNMs with expected off-target sites containing up to 7 mismatches, including both–AG and–GG PAM sequences. With 276,254 potential off-target sites for the combined Tyr2R and Tyr2F gRNA (Ty2F, 119,754; Tyr2R, 156,500), we expect approximately 32 intersections between these candidate off-target sites and our candidate DNM sites. This is based on an expected 7,460 unfiltered DNMs per animal and a total number of 43 * 276,524 = 6,360,052 bases covered by candidate off-target sites and a mouse genome size of 2.8*10^9 bases: expected intersections = number of attempts * probability of success of intersection = 7,460 * 43 * 276,524 / (2.8*10^9) = 31.6. Using bedtools-2.23 window–w10, we found between 4 and 14 overlaps per sample (S5 Table, worksheet “RGEN_7mm_overlap_counts”), which is lower, but the same order of magnitude as the estimate. We assessed the null hypothesis that the intersection counts of Tyr-treated and Tyr-untreated animals were drawn from the same population using a Wilcoxon Rank-Sum test, with the R function wilcox.test, and found no significant difference in intersection counts between the population of treated animals and the population of untreated animals.
Due to the small number of samples and the low variant counts in each group, we chose to perform either a Kruskal-Wallis Rank Sum test (in the case of more than two groups) or a Wilcoxon Rank Sum test (for only two groups) to assess the null hypothesis: namely that the counts in each group were drawn from the same population. Four tests were performed using the R function kruskal.test or wilcox.test and a Bonferroni-adjusted critical p-value of 0.0125, comparing within treated or untreated groups for SNVs or indels. These tests were not significant. Finally, a Wilcoxon Rank Sum test compared all untreated to all treated embryos for SNVs or indels. This test was also not significant. The groups compared, the values tested, the tests used and test results are presented in S7 Table.
1μl of genomic DNA was used for amplicon specific PCR using genome specific primers (S8 Table). The indexed libraries were sequenced using standard protocols and Illumina MiSeq technologies (Paired End 250bp runs).
Each candidate position was sequenced to an average depth of at least 10,000 reads in embryos and (pooled) parent samples. The sequences were directly aligned to the GRCm38 assembly using bwa-mem, and the resulting alignments directly inspected at each variant location using samtools-1.3.1 mpileup -d50000 -Q0 -q0 to check for the presence of the alternate allele in the parents’ sample, and to confirm the alternate allele in the embryo sample. Locations in the pooled parent sample with more than 100 reads showing the alternate allele and greater than 1% alternate allele fraction were classed as not validated, as were locations in the embryo samples showing less than 1% alternate allele fraction. We found that 87% of SNVs and 73% of indels were validated (S8 Table).
To estimate the effect of using a distantly related parent on the measured de novo mutation rate, we re-analyzed our embryos for DNMs using a genetically distant parent. Specifically, it can be seen from the relationships of matings contributing to our experiment’s parents and embryos (S1 Fig) that the male parent of the “Cas9 only” embryos (mating CBLT8902) was distant by more than five generations from the equivalent male parents of the other treatment groups (matings CBLT8762, CBLT9125, CLBT8712), whereas the female parents for all treatment groups were drawn from mating CBLT9125. We therefore chose to fix the parents for the “Cas9 only” group as the parents for all treatment groups as input to TrioDeNovo and reran our pipeline, including filtration of variants up to the removal of C57BL/6NJ and C57BL/6N variants as well as repeats (filtration steps 1–3 and step 5, section 7). It was not possible to run further filtration, e.g. removal of cross-animal noise, as the “correct” parents were present in the other animals.
The results show a median of 98.5 DNMs per animal for the treatment groups with distantly related controls (“No Injection”, “Sham injection”, “Tyr2F” and “Tyr2R”), whereas the treatment group with the correct control (“Cas9 Only”) has a median number of 28 DNMs per animal (S1 Fig, S9 Table). The equivalent median of all treatment groups is 32 (S4 Table, column L; ‘not on-target, vaf, parent noise, repeat and BL6NJ/N filtered’). This demonstrates the effect of using unrelated or distantly related parents as controls, when searching for CRISPR off-targets, is an inflation of 66.5 false positive variants and reinforces the need for studies to use trios in these experiments.
We investigated the robustness of our results by re-analysing our aligned reads using somatic variant callers developed for the analysis of cancer genomes. We labelled each embryo as a “tumour” and each parent sample as a “normal”. We used CaVEMan[12] to call SNVs and cgpPindel[13] (a modified Pindel version 2.0) to call indels.
CaVEMan called between 1,854 and 3,700 passing variants per sample with a median of 2,845 variants (S10 Table, worksheet “Caveman”). In order to remove false positive variants, we applied the following filters to the CaVEMan calls: (1) ASMD (the median alignment score of reads showing the variant allele) > = 140. (2) Read depth > = 10 for both the embryos and the parents, and alternative read depth > = 4 in the embryos. (3) Variant allele fraction > = 10%. (4) Alternative read depth = 0 in either parent, to avoid false positives arising from low-allele-fraction mosaic variants. (5) Somatic calls had to be present in the comparisons with both parents. (6) Variants could not be shared between two or more samples. (7) Variants could not coincide with an allele reported in any strain of the Mouse Genomes Project[11]. (8) We excluded variants inside or 1bp adjacent to any repeat region (we used UCSC repeat tracks as described above).
Using the default WGS panel of filtering rules [12], CgpPindel produced between 44 and 116 candidate indels per offspring, with a median of 71.5 indels (S10 Table). To this output, we applied the following additional filters. (1) We selected the variants with allele fraction > = 10%. (2) We selected for each embryo only the somatic calls present in the comparisons with both parents. (3) We removed any variant shared by two or more samples. (We note that only one indel is shared between two samples, which are treated with the same gRNA–namely, between MD5639a and MD5642a. However, this shared indel is in a microsatellite region, and it is not predicted to be a CRISPR off-target location). (4) We filtered variants coincident with an allele reported in any strain of the Mouse Genomes Project[11]. (5) We excluded variants inside or 1bp adjacent to any repeat region (we used UCSC repeat tracks as described in the previous section).
After filtering SNVs and indels in this way, we obtained 14 to 36 SNVs per sample (median 21) and 0 to 6 indels per sample (median 2). All 471 SNVs and indels are listed in S11 Table. The somatic calling pipelines identified 80% (313 out of 389) of the variants identified by the de novo mutation calling pipeline. The mutations identified by the somatic pipelines, which were not identified by the de novo pipeline, had a lower variant allele frequency (median of 12%) compared to those identified by both methods (median of 34%). We also noticed that cgpPindel identified an additional mosaic 337 bp deletion at the Tyr2F on-target location for MD5638a (S6 Table).
As with the previous variant analyses, we performed a Kruskal-Wallis Rank Sum or Wilcoxon Rank Sum tests to assess the null hypothesis that the SNV or indel counts from each of the control groups or treatment groups were drawn from the same population (S12 Table). We performed four such tests and they failed to produce any significant result at the Bonferroni-adjusted critical p-value of 0.0125. Finally, a Wilcoxon Rank Sum test comparing the combined untreated group to the combined treated group, also failed to produce a significant result (S12 Table).
|
10.1371/journal.ppat.1007174 | Alcohol enhances type 1 interferon-α production and mortality in young mice infected with Mycobacterium tuberculosis | In the current study, we used a mouse model and human blood samples to determine the effects of chronic alcohol consumption on immune responses during Mycobacterium tuberculosis (Mtb) infection. Alcohol increased the mortality of young mice but not old mice with Mtb infection. CD11b+Ly6G+ cells are the major source of IFN-α in the lungs of Mtb-infected alcohol-fed young mice, and IFN-α enhances macrophage necroptosis in the lungs. Treatment with an anti-IFNAR-1 antibody enhanced the survival of Mtb-infected alcohol-fed young mice. In response to Mtb, peripheral blood mononuclear cells (PBMCs) from alcoholic young healthy individuals with latent tuberculosis infection (LTBI) produced significantly higher amounts of IFN-α than those from non-alcoholic young healthy LTBI+ individuals and alcoholic and non-alcoholic old healthy LTBI+ individuals. Our study demonstrates that alcohol enhances IFN-α production by CD11b+Ly6G+ cells in the lungs of young Mtb-infected mice, which leads to macrophage necroptosis and increased mortality. Our findings also suggest that young alcoholic LTBI+ individuals have a higher risk of developing active TB infection.
| Chronic alcohol consumption modulates the host immune defense mechanism(s) and makes the host susceptible to various fungal, viral and bacterial infections, including Mycobacterium tuberculosis (Mtb). However, limited information is available about the mechanisms involved in alcohol-mediated host susceptibility to Mtb and other intracellular bacterial infections. In the current study, we fed control and alcohol diets to young and old mice and determined the mortality rates and the immune mechanisms involved in host susceptibility to Mtb infection. We found that alcohol increases the mortality of young mice but not old mice infected with Mtb. The increased mortality in alcohol-fed Mtb-infected young mice was due to IFN-α production by CD11b+Ly6G+ cells. We also found that PBMCs from young alcoholic individuals with latent tuberculosis infection (LTBI) produced significantly higher amounts of IFN-α than those from young non-alcoholic, old alcoholic and old non-alcoholic LTBI+ individuals. Our findings suggest that young alcoholic LTBI+ individuals have a higher risk of developing active TB infection. Our studies provide the first evidence that chronic alcohol consumption induces IFN-α production in young Mtb-infected mice and increases their mortality rates. Further characterization of CD11b+Ly6G+ cells and delineation of the mechanisms through which alcohol enhances IFN-α production in Ly6G+ cells during Mtb infection will facilitate the development of therapies for alcoholic individuals with latent and active Mtb.
| It is estimated that more than two billion people worldwide are infected with Mycobacterium tuberculosis (Mtb), but only 5–10% of these individuals develop TB during their lifetime [1,2]. The geriatric population represents a large reservoir of latent tuberculosis infection (LTBI) [3]. It is difficult to diagnose and treat tuberculosis in aged individuals [3,4]. Approximately 57% of tuberculosis deaths occur in the aged population (above 50), and this burden is high in developed countries [5]. Immunosuppressive conditions, such as HIV infection, diabetes mellitus and drug and alcohol abuse, are risk factors that increase the chances of tuberculosis (TB) reactivation in people with LTBI [6–9]. In addition, individuals with alcoholism show higher relapse rates and a higher probability of having multidrug-resistant TB [10].
Alcoholism leads to the development of liver cirrhosis, cancer, insulin resistance, epilepsy, hypertension, psoriasis, preterm birth complications, cardiovascular diseases and stroke [11,12]. Chronic alcohol consumption impairs the host immune response to cancer and infections [13]. Alcohol impairs monocyte phagocytic and antigen-presenting capacities and suppresses the alveolar macrophage production of monokines, such as IL-23, in response to infection [14–16]. Alcohol-exposed dendritic cells produce more IL-10 and less IL-12, suggesting an inhibitory effect on dendritic cell function [17,18]. In humans and experimental mice, chronic alcohol consumption makes neutrophils hypo-responsive to bacterial infections[19]. Prolonged alcohol consumption induces type I interferon (IFN) and tumor necrosis factor alpha (TNF alpha) production [20]. Alcohol impairs NK cell trafficking and inhibits NK cell cytotoxicity [21]. Chronic alcohol consumption impairs adaptive immune responses mediated by B and T-cells [22]. These immunosuppressive effects of alcohol are more severe in elderly individuals than in young individuals [23].
Chronic alcohol consumption makes the host susceptible to various bacterial infections, including TB [24]. Epidemiological and immunological evidence strongly suggest a link between alcoholism and the worsening of TB disease [25]. Chronic alcohol consumption impairs the immune responses of Mtb-infected mice [19]. Alcohol feeding before BCG vaccination reduces T cell responses, but there are no effects when BCG vaccination is delivered prior to alcohol feeding [26]. These studies were performed after the short term-feeding of mice with alcohol, and the mechanism(s) involved in host susceptibility remain unknown. The long-term effects of alcohol consumption on host defense mechanisms against Mtb infection are also unknown, particularly in old individuals.
In this study, we determined the survival of alcohol-fed young and old mice infected with Mtb. We also determined the immune mechanisms responsible for the early death of alcohol diet-fed young mice infected with Mtb.
Young and old mice were fed alcohol or control diets for one month and then infected with Mtb H37Rv as detailed in the methods section. Alcohol or control diet feeding was continued until the death of the mice or the termination of the experiment. As shown in Fig 1, eighty percent of Mtb-infected alcohol-fed young mice died within 6 months (p<0.01, Fig 1A); there was a twenty-five percent death rate in Mtb-infected alcohol-fed old mice, a twenty-five percent death rate in Mtb-infected control diet-fed old mice and no deaths in the control diet-fed young mice. In these groups of mice, most of the deaths occurred after three months. The bacterial burden in the lungs of these mice was measured three months after Mtb infection. As shown in Fig 1B, there was a marginal increase in the bacterial burden in Mtb-infected control diet-fed old mice compared to that in the other groups, and there was a marginal but significant decrease in the bacterial burden in the lungs of alcohol-fed old mice. The above results demonstrate that there is no correlation between the bacterial burden and increased mortality in alcoholic mice infected with Mtb. Serum alcohol levels and liver alanine transaminase activity were similar among all groups of mice (Figs 1A, 1B and S1).
We determined whether alcohol had any effect on the pro- and anti-inflammatory responses of young mice following Mtb infection. Young mice were fed control and alcohol diets and infected with Mtb as in Fig 1. After three months, the levels of various cytokines and chemokines were measured in the lung homogenates by multiplex (23-plex) ELISA. As shown in Fig 2A, at three months p.i., various cytokines and chemokines were measured, but only IFN-α levels were increased significantly in Mtb-infected alcohol-fed young mice compared to those in uninfected alcohol-fed and Mtb-infected control diet-fed mice (Fig 2A). There was a marginal but significant decrease in IL-1α levels in Mtb-infected alcohol-fed young mice compared to those in Mtb-infected control diet-fed mice (Fig 2A).
Histological analyses indicated that the number of lesions throughout the lungs was significantly higher in Mtb-infected alcohol diet-fed young mice than in Mtb-infected young control mice and uninfected young mice (Fig 2B, 2C and 2D).
To determine the cellular source of IFN-α in Mtb-infected alcohol diet-fed young mice, we first quantified the leukocyte populations by flow cytometry. As shown in Fig 3 and S2A and S2B Fig, at three months after Mtb infection, the number of CD11b+Ly6G+ cells in the lungs was significantly higher in Mtb-infected alcohol diet-fed young mice than in Mtb-infected control mice and uninfected alcohol diet-fed mice. We next determined the phenotype of IFN-α-producing pulmonary cells three months p.i.; there were no significant differences in the absolute numbers of IFN-α-producing CD11c+ and F4/80 cells (Fig 4A and 4B). However, the absolute number of IFN-α-producing CD11b+Ly6G+ cells in the lungs was significantly higher in Mtb-infected young alcoholic mice than in uninfected alcohol diet-fed mice and Mtb-infected control mice (Fig 4C and 4D). To confirm our findings that IFN-α levels were increased in the lungs of Mtb-infected young alcohol diet-fed mice, mice were euthanized three months p.i., and lung sections were examined for IFN-α+ cells by immunofluorescence staining. As shown in Fig 4F, the mean immunofluorescence intensity for IFN-α was significantly higher in Mtb-infected alcohol diet-fed young mice than in Mtb-infected control and uninfected alcoholic mice. We also found that Ly6G+ cells are the major source of IFN-α (Fig 4E). We further characterized this cell population in the lung tissues of Mtb-infected young alcohol diet-fed mice. As shown in S6 Fig, Ly6G+IFN-α+ cells were positive for CD11b, CD200 and CD163 but negative for F4/80, CD68, CD115, CD11c, and Ly6C.
Type 1 interferons have a protective role in viral infections [27]. However, in Mtb infection, type 1 interferon signaling causes immunopathology and early mortality in the infected mice [28]. We investigated whether the increased mortality of Mtb-infected alcohol-fed young mice was due to enhanced IFN-α production. Young mice were fed control and alcohol diets and infected with Mtb as in Fig 1. After three months, the mice were treated with either a neutralizing anti-IFNAR-1 mAb or an isotype-matched IgG1 control mAb. As shown in Fig 5A, 100% percent (p = 0.001) of the Mtb-infected alcohol diet-fed young mice that received the isotype-matched control mAb died within 2 months. In contrast, all mice that received the anti-IFNAR-1 mAb survived. Histological analyses indicated that there were significantly fewer necrotic lesions throughout the lungs of the anti-IFNAR-1 mAb-treated Mtb-infected alcohol diet-fed young mice than in those of the isotype antibody-treated Mtb-infected alcohol diet-fed young mice (Fig 5B, 5C and 5D).
We further examined the lung lesions of Mtb-infected (three months after infection) alcohol diet-fed young mice using confocal microscopy. As shown in S3 Fig, cleaved caspase 3 expression was similar in Mtb-infected young alcoholic mice, Mtb-infected control and uninfected alcoholic mice, suggesting that there is no significant difference in lung cell apoptosis in these groups of mice. We next examined the expression of receptor-interacting serine/threonine-protein kinase (RIP)-1 and RIP-3, which are known to be expressed by cells undergoing programmed necrotic cell death (necroptosis) [29]. Six months after Mtb infection, lungs were isolated from control and alcohol diet-fed young and old mice, and the gene expression levels of RIP-1 and RIP-3 were determined by real-time PCR. As shown in Fig 6A, the expression levels of RIP-1 and RIP-3 in the lungs were significantly higher in Mtb-infected alcohol diet-fed young mice than in Mtb-infected alcohol diet-fed old mice and control diet-fed young mice. Confocal microscopy examinations of the lung sections also indicated significantly higher RIP-1 and RIP-3 expression in the lung lesions of Mtb-infected alcohol diet-fed young mice than in the lungs of Mtb-infected control and uninfected alcoholic mice (Fig 6B and S3B Fig). RIP-1 and RIP-3 are expressed by F4/80 macrophages but were not expressed by IFN-α-producing Ly6G+ cells (Fig 6C and 6D). To determine whether IFN-α-producing Ly6G+ cells are involved in the enhanced RIP-1 and RIP-3 expression in lung macrophages, we first used confocal microscopy to examine the lung sections of Mtb-infected young alcoholic mice for Ly6G+ and RIP-1 and RIP-3-expressing F4/80+ cell interactions. At three months p.i., the imaging results indicated that RIP-1 and RIP-3-expressing F4/80+ cells from Mtb-infected young alcoholic mice were in closer proximity to IFN-α-producing Ly6G+ cells than those from Mtb-infected control mice (Fig 6D and S4 Fig). More importantly, our results indicate that the marked increase in IFN-α production was spatially defined at the region where both Ly6G+ and RIP-1 and RIP-3-expressing F4/80+ cells interact with each other (Fig 6D). RIP-1 and RIP-3 expression levels in the lungs were significantly lower in anti-IFNAR-1 mAb-treated Mtb-infected young alcoholic mice than in isotype antibody-treated Mtb-infected young alcoholic mice (Fig 6E).
We compared IFN-α production in the lungs of alcohol-fed young and old mice following Mtb infection. Young and old mice were fed control and alcohol diets and infected with Mtb as in Fig 1. After three months, the mice were euthanized, and the lung sections were examined for IFN-α by confocal microscopy. As shown in S5 Fig, the immunofluorescence intensity for IFN-α was significantly lower in Mtb-infected old alcoholic mice than in Mtb-infected young alcoholic mice. We also found that RIP-1 and RIP-3 expression levels were significantly lower in the Mtb-infected old alcoholic mice than in the Mtb-infected young alcoholic mice (S5B, S5C and S5D Fig).
To determine the relevance of the above findings to the clinical manifestation of human Mtb infection, we obtained blood samples from alcoholic and non-alcoholic LTBI+ individuals. First, on the basis of age, we characterized the LTBI+ individuals by age group: <45 years (young) and >50 years (old). We cultured peripheral blood mononuclear cells (PBMCs) in the presence of 10 μg/ml γ-irradiated Mtb. After 72 hours, IFN-α levels were determined by ELISA as detailed in the methods section. As shown in Fig 7, γ-irradiated Mtb significantly enhanced the IFN-α levels by 2-fold in PBMCs from young alcoholic LTBI+ individuals and compared with those from non-alcoholic young LTBI+ individuals and by 2.9-fold compared to those from old alcoholic LTBI+ individuals. The baseline IFN-α levels were also high in young alcoholic LTBI+ individuals compared with those of other groups (Fig 7).
Chronic alcohol consumption modulates host immune defense mechanism(s) and makes the host susceptible to various fungal, viral and bacterial infections, including Mtb [13,15,19]. However, limited information is available regarding the mechanisms involved in alcohol-mediated host susceptibility to Mtb and other intracellular bacterial infections. In the current study, we fed young and old mice control and alcohol diets and determined the mortality rates and the immune mechanisms involved in host susceptibility to Mtb infection. Approximately 80% of the Mtb-infected alcohol-fed young mice died within 5 months; however, only 25% of Mtb-infected alcohol-fed old mice and 25% of alcohol-fed uninfected young mice died during the same period. There were no significant differences in the bacterial lung burdens of control and alcohol diet-fed young mice and alcohol diet-fed old and young mice. IFN-α levels were significantly higher in the lungs of Mtb-infected alcohol-fed young mice, and treatment with an anti-IFNAR-1 antibody increased their survival. In the lungs of Mtb-infected alcohol-fed young mice, IFN-α enhanced the expression of RIP-1 and RIP-3, which are known to be involved in necroptosis. Mtb-infected alcohol-fed old mice and Mtb-infected control diet-fed old and young mice did not express IFN-α, RIP-1 or RIP-3 in their lungs. In response to Mtb, PBMCs from alcoholic LTBI+ healthy individuals produced significantly higher amounts of IFN-α than PBMCs from non-alcoholic young LTBI+ individuals and alcoholic and non-alcoholic aged LTBI+ individuals. Our findings demonstrate that alcohol enhances Ly6G+ cell infiltration and IFN-α production and increases necroptosis in the lung macrophages of young mice infected with Mtb, which leads to enhanced mortality.
Chronic alcohol consumption inhibits host protective immune responses to infections, including Mtb infection, and increases the mortality rates of young and aged individuals [30,31]. According to the Centers for Disease Control’s (CDC) estimations, one-third of binge drinkers are old individuals, and human studies have found that compared young individuals, old individuals are more susceptible to various diseases [32–34]. Old individuals are likely to take prescribed medications, and in some cases, malnourishment and alcohol may have different effects on these individuals [34–36]. No experimental animal studies have been performed to determine the effects of chronic alcohol feeding in aging and Mtb infection. In the current study, we found that alcohol diet-fed young mice (1–2 months) are more susceptible to Mtb infection and have a higher mortality rate than alcohol diet-fed old mice (17–22 months) (Fig 1A). Our findings suggest that alcohol worsens the TB pathology in the early stage of life and leads to increased mortality.
We found that IFN-α is responsible for the early death of alcoholic Mtb-infected young mice. IFN-α is a type 1 interferon that belongs to the interferon family, which regulates the immune responses to infection, cancer and autoimmune diseases [37,38]. Type 1 interferons have a protective role during viral infections, but during Mtb infection, they enhance the pathogenicity [39,40]. Mtb proteins induce the production of type 1 interferons by host myeloid cells [41]. Type 1 interferons inhibit IL-1β production and enhance Mtb growth in myeloid cells [42]. In the current study, we found that IFN-α produced by Ly6G+ cells was associated with macrophage necroptosis and fatal immunopathology in the lungs of young alcohol diet-fed mice IFNAR1 signaling is detrimental during Mtb infection and promotes excess inflammation [28]. Furthermore, in the current study, we found that Mtb-infected alcohol diet-fed mice survived for a longer period of time and had less bacterial burden in their lungs when type 1 IFN signaling pathways were blocked.
Old mice express transient early resistance to pulmonary tuberculosis, and type 1 cytokines have no influence on this early resistance [43]. It is known that several signaling pathways are defective in old mice [44]. The transient resistance in Mtb-infected old mice is due to a population of memory CD8+ T cells that express several receptors for Th1 cytokines; in addition, in aged mice, lung macrophages secret more proinflammatory cytokines in response to Mtb [43]. We have not determined the CD8+ cell and macrophage responses in alcoholic old mice; however, our results demonstrate that alcohol-fed mice were unable to enhance IFN-α production in Mtb-infected old mice, and there were no effects on mortality compared to the non-alcoholic Mtb-infected old mice. IFN-α production in young alcoholic Mtb-infected mice significantly reduced their survival. Our current findings suggest that defective signaling pathways that are involved in the production of IFN-α in Mtb-infected old mice may be protecting them from alcohol-mediated lung cell necroptosis.
In various experimental models, it was shown that immune cells, such as macrophages, dendritic cells, T cells and Ly6G+ cells, produce type 1 interferons, and plasmacytoid dendritic cells are the major source [45]. However, only less than 1% of leucocytes are dendritic cells, and fifty percent of blood leucocytes are Ly6G+ cells [46]. Necrosis and Ly6G+ cell infiltration in the lung granuloma are characteristic features of tuberculosis granulomas, and these properties are associated with increased mycobacterial load and exacerbated lung pathology in human and experimental animals [47]. Netting Ly6G+ cells are the major inducers of type I IFN production [48]. Whole blood transcript signatures for active TB patients and pathway analyses revealed that the TB signature is dominated by a neutrophil-driven interferon (IFN)-inducible gene profile that consists of both IFN-γ and type I IFNαβ signaling [49]. Alcohol consumption can reduce the recruitment of Ly6G+ cells to the site of infection [19,50]. We have further investigated whether IFN-α-producing CD11b+ Ly6G+ cells are neutrophils. We found that these cells express a unique phenotype, including some neutrophils markers (positive for CD11b, CD200 and CD163 but negative for F4/80, CD68, CD115, CD11c, and Ly6C), suggesting these cells are not neutrophils. In the current study, we found fewer CD11b+Ly6G+ cells in the lungs of alcohol-fed Mtb-infected old mice than in those of alcohol-fed Mtb-infected young mice. We have not determined the factors underlying the reduced CD11b+Ly6G+ infiltration in the lungs of alcohol-fed Mtb-infected old mice, but it is known that in old individuals, the neutrophil lifespan is decreased, neutrophil precursors in the bone marrow proliferate less, neutrophil recruitment at the site of inflammation is reduced, and CD11b+ Ly6G+ cells are less functional due to alterations in signaling pathways [51,52]. Our results suggest that alcohol consumption can enhance these defects in old mice and that Ly6G+ neutrophil- like cells are unable to migrate to the lungs of Mtb-infected mice, resulting in less IFN-α production and necroptosis and enhanced survival.
We found that IFN-α produced by Ly6G+ cells in alcohol-fed Mtb-infected young mice induces necroptosis in lung macrophages. Necroptosis is programmed necrosis that differs from other death pathways (apoptosis, autophagy and pyroptosis) due to the requirement of a unique signaling pathway associated with the activation of receptor-interacting protein (RIP) kinases 1 and 3 [53,54]. Caspase 1 expression was similar among all groups of infected mice, but RIP-1 and RIP-3 expression was significantly higher in the lungs of alcohol-fed Mtb-infected young mice than in those of control diet-fed Mtb-infected young mice and alcohol diet-fed Mtb-infected old mice. We also found that RIP-1 and RIP-3 expression was restricted to macrophages and that IFN-α-producing Ly6G+ cells were colocalized around macrophages. These findings suggest that IFN-α-producing Ly6G+ cells in alcohol-fed Mtb-infected young mice enhance necroptosis in lung macrophages.
Necroptosis exacerbates inflammatory responses to infection which contributes to tissue damage and pathology [55,56]. Our current findings demonstrate that alcohol enhances IFN-α mediated necroptotic death of lung macrophages in young Mtb-infected mice (Fig 6C). This leads to tissue damage and mortality in young alcoholic Mtb-infected mice.
To determine the clinical relevance of our mouse studies, we compared IFN-α levels in the culture supernatants of γ-irradiated Mtb-cultured PBMCs from young and old alcoholic and non-alcoholic healthy LTBI+ individuals. We found that PBMCs from young alcoholic LTBI+ individuals produced significantly higher amounts of IFN-α after culture with γ-irradiated Mtb than those from young non-alcoholic, old alcoholic and old non-alcoholic healthy LTBI+ individuals (Fig 7). Our findings suggest that young alcoholic LTBI+ individuals have a higher risk of developing active TB infection.
In conclusion, our studies demonstrate that alcohol increases the mortality of young but not old mice infected with Mtb. The increased mortality of alcohol-fed Mtb-infected young mice is due to IFN-α production by Ly6G+ cells. Further characterization of the exact phenotype of CD11b+ Ly6G+ cells and the delineation of the mechanisms through which alcohol enhances IFN-α production by Ly6G+ cells during Mtb infection will facilitate the development of therapies for alcoholic individuals with latent and active Mtb. Our findings may also be applicable to other intracellular pathogen infections.
All animal studies were performed with specific pathogen-free, 6- to 8- week -old and 17- to 22-month-old male and female C57BL/6 mice (Jackson Laboratory and National Cancer Institute). The Institutional Animal Care and Use Committee of the University of Texas Health Science Center at Tyler approved the studies. The animal procedures involving the care and use of mice were conducted in accordance with the guidelines of the NIH/OLAW (Office of Laboratory Animal Welfare).
Blood was obtained from 17 non-alcoholic and 20 alcoholic healthy LTBI+ individuals who were 18–75 years of age. PBMCs were isolated from freshly collected blood samples. All subjects were HIV seronegative. The alcoholic LTBI+ individuals had a history of drinking at least 10–12 drinks per week.
All human studies were approved by the Institutional Review Board of the Bhagwan Mahavir Medical Research Centre, and informed written consent was obtained from all participants. All human subjects involved in our study were adults. All animal studies were approved by the Institutional Animal Care and Use Committee of the University of Texas Health Science Center at Tyler (Protocol #554). All animal procedures involving the care and use of mice were undertaken in accordance with the guidelines of the NIH/OLAW (Office of Laboratory Animal Welfare).
All mice were maintained on a standard rodent chow diet (LabDiet, catalog number 5053, St. Louis, MO, 4.07 kcal/gm) until the beginning of the experiment, when they were randomized into control or alcohol-containing liquid diet groups. The mice were fed alcohol using the Lieber-DeCarli liquid diet formulation (Dyets Inc., catalog number 710260; Bethlehem, Pa.4.5 kcal/gm), which supplies 36% of the caloric intake as ethanol, or were fed an isocaloric liquid control diet (LCD) (Dyets Inc., catalog number 710027; Bethlehem, Pa, 4.5 kcal/gm) as previously described [57]. The animals were fed the respective liquid diets for 5 of 7 days and the chow diet for 2 of 7 days. Animals in the liquid ethanol diet (LED) group were given water containing 20% (wt/vol) ethanol on the two chow diet days. The weights of the mice were recorded weekly.
Mice were fed the alcohol and control diets, and after three months, they were infected with Mtb H37Rv using an aerosol exposure chamber as described previously [58]. Briefly, Mtb H37Rv was grown to the mid-log phase in liquid medium and then frozen in aliquots at -70°C. Bacterial counts were determined by plating on 7H10 agar supplemented with oleic albumin dextrose catalase (OADC). For infection, the bacterial stocks were diluted in 10 ml of normal saline (to 0.5 ×106 CFU [colony forming units]/ml, 1 ×106 CFU/ml, 2 ×106 CFU/ml, and 4 × 106 CFU/ml) and placed in a nebulizer within an aerosol exposure chamber custom made by the University of Wisconsin. In the preliminary studies, groups of three mice were exposed to the aerosol at each concentration for 15 min. After 24 h, the mice were euthanized, and homogenized lung samples were plated on 7H10 agar plates supplemented with OADC. CFUs were counted after 14–22 days of incubation at 37°C. The aerosol concentration that resulted in ~50–100 bacteria in the lungs was used for the subsequent studies.
For some experiments, mice were treated with anti-IFNAR-1 antibodies. One month after control and alcohol diet feeding, the mice were challenged with aerosolized Mtb. After 3 months, the mice received 0.3 mg of anti-IFNAR-1 (BioXcell, Clone: MAR1-5A3, Catalog number: BP0241) or isotype-matched control Ab (rat IgG1, Clone: MOPC-21, Catalog number: BE0083) intravenously every 4 days for up to 2 months.
Lungs were harvested from the alcohol and control diet-fed mice at the indicated time points after Mtb challenge and were placed into 60-mm dishes containing 2 ml of Hank's balanced salt solution (HBSS). The tissues were minced with scissors into pieces no larger than 2–3 mm, and the fluid was discharged onto a 70-μm filter (BD Biosciences, San Jose, CA) that had been pre-wetted with 1 ml of PBS containing 0.5% bovine serum albumin (BSA, Sigma-Aldrich) suspended over a 50-ml conical tube. The syringe plunger was then used to gently disrupt the lung tissue before washing the filter with 2 ml of cold PBS/0.5% BSA. The total number of viable cells in the lungs was determined with the trypan blue exclusion method. For flow cytometry experiments, we gated based on the total lung CD45+ cells (leukocytes) and measured various cell populations.
For flow cytometry, we used FITC anti-CD3, PE anti-CD8, APC anti-CD4, APC anti-NK1.1, APC anti-CD11b, FITC anti-Ly6G, PE anti-IFN-α, APC CD11-C, and FITC anti-F4/80 antibodies (all from BioLegend). The antibodies used for the in vivo neutralization experiments were purchased from BioXcell (anti-mouse IFNAR-1, Clone: MAR1-5A3, Catalog number: BP0241, and mouse IgG1 isotype control, Clone: MOPC-21, Catalog number: BE0083). Anti-Ly6G (Sigma-Aldrich; MABF474), anti-F4/80 (Abcam; ab6640), anti-IFN-α-FITC conjugated (R&D Systems; 22100–3), anti-cleaved caspase-3 (Cell Signaling Technology; 9661S), anti-CD163 (Santa Cruz Biotechnology, INC; sc-58965), anti-Ly6C (Santa Cruz Biotechnology, INC; sc-52650), anti-CD115 (Santa Cruz Biotechnology, INC; sc-46662), anti-CD200 (Santa Cruz Biotechnology, INC; sc-53100), anti-CD11c (Abcam; ab33483), anti-CD68 (Abcam; ab53444), anti-RIP-1/3 (Santa Cruz Biotechnology, INC; sc-133102/sc-374639) and secondary antibodies (goat anti-rat IgG (H+L) -Alexa 647, goat anti-rabbit IgG (H+L), Alexa Fluor 488, and goat anti-mouse IgG (H+L), Alexa Fluor 594) were obtained from Life Technologies, and fluoroshield mounting medium with DAPI (Abcam, ab104139) was used for the confocal microscopy analyses.
For surface staining, 106 cells were resuspended in 100 μl of staining buffer (PBS containing 2% heat-inactivated FBS) and Abs. The cells were then incubated at 4° C for 30 min, washed twice and fixed in 1% paraformaldehyde before acquisition using a FACS Calibur flow cytometer (BD Biosciences). In some experiments, intracellular staining for IFN-α was performed. Controls for each experiment included cells that were unstained, cells to which PE-conjugated rat IgG had been added and cells that were single stained, either for a surface marker or for intracellular molecules. For IFN-α analysis, we gated based on CD11c, F4/80, CD11b or Ly6G-positive cells and determined the percentages or the number of IFN-α expressing cells.
In the lung homogenates, the following 27 cytokines and chemokines were measured using a multiplex ELISA kit (M60009RDPD, Bio-Rad). The cytokines and chemokines analyzed were IL-1b, IL-1ra, IL-2, IL-4, IL-5, IL-6, IL-7, IL-8, IL-9, IL-10, IL-12 (p70), IL-13, IL-15, IL-17, basic FGF, eotaxin, G-CSF, GM-CSF, IFN-γ, IP-10, MCP-1 (MCAF), MIP-1a, MIP-1b, PDGF-BB, RANTES, TNF-α and VEGF.
RNA was isolated from lungs using TRIzol (Invitrogen) according to the manufacturer's instructions. Complementary DNA (cDNA) was generated from 0.5 mg of RNA and random hexamer primers using a Maxima First Strand cDNA Synthesis Kit for RT-qPCR (BIO-RAD) according to the manufacturer's instructions, and real-time PCR was performed. Gene expression for RIP-1 and RIP-3 was determined using Sybr green master mix (Qiagen), gene-specific primers (Sigma-Aldrich) and an ABI Prism 7600. All gene expression levels were normalized to β-actin internal controls, and the fold changes were calculated using the 2-ΔΔCT method.
IFN-α levels were measured using ELISA kits (Abcam, USA, catalog number: ab213479) according to the manufacturer’s instructions.
Serum was collected without anti-coagulant by cardiac puncture from control and alcohol diet-fed mice. Serum alcohol levels were determined by using an ethanol assay kit as per the manufacturer’s guidelines (Abcam, USA, catalog number: ab65343).
At the specified time points, mice were euthanized, and the harvested lungs were placed in 10% neutral buffered formalin (Statlab, McKinney, TX, USA) for 48 hours to inactivate the infectious agent. Paraffin-embedded blocks were cut into 5 μm-thick sections. For morphometric lesion analyses, the lung sections were stained with hematoxylin and eosin (H&E) and examined in a blinded manner to assess the necrotic lesions as previously described by Sibila et al. [59]. Briefly, each lung lobe was quantified for the lesion area and percentage of the lung lesions by using digital software (NIH ImageJ; developed at the U.S. National Institutes of Health and available on the Internet at https://imagej.nih.gov/ij/). Two investigators, DT and SC, independently assessed the immunohistochemical readouts using morphometric analyses.
Confocal microscopy was performed to colocalize IFN-α-producing Ly6G and RIP1/3-expressing F4/80 cells in the lung sections. The lung tissues were stored in 10% neutral buffered formalin; then, the samples were paraffin embedded and cut into 5 μM thick sections that were deparaffinized and rehydrated. The tissue sections were subjected to heat-induced antigen retrieval in 10 mM sodium citrate buffer (pH 6.0). Then, the lung tissue sections were incubated in 0.025% Triton X-100 in PBST for 10 min and washed 3 × 5 min using PBS. Nonspecific binding was blocked with 5% goat serum in PBST for 1 hour, and the slides were washed 2 × 5 min with PBS. The slides were then incubated at 4°C overnight in PBST with the appropriate dilutions of the following primary antibodies: anti-Ly6G (1:200), anti-F4/80 (1:50), anti-IFN-α-FITC-conjugated (1:50), anti-cleaved caspase-3 (1:400), anti-CD68 (1:100), anti-CD115 (1:50), anti-CD200 (1:50), anti-CD163 (1:50), anti-CD11c (1:100), anti-Ly6C (1:50) and anti-RIP-1/RIP-3 (1:50); subsequently, the slides were washed thoroughly 3 × 5 min with PBS. Then, the tissue sections were stained with the respective secondary antibodies at 1:1000 dilutions (v/v), washed again with PBS for 3 × 5 min, and mounted with fluoroshield mounting medium with DAPI. The slides were then examined and analyzed under a laser-scanning confocal microscope (Zeiss LSM 510 Meta). An IgG isotype secondary control was used for all the confocal microscopy studies, and Zen 2009 software (Carl Zeiss) was used for image acquisition; then, the images were processed/quantified uniformly for each experiment using ImageJ NIH software. Representative images from three different independent experiments are shown.
Data analyses were performed using GraphPad Prism (GraphPad Software, Inc., La Jolla, CA). The results are expressed as the mean ± SE. For normally distributed data, comparisons between groups were performed using a paired or unpaired t-test and ANOVA as appropriate. Mouse survival was compared using the Kaplan- Meier log-rank test.
|
10.1371/journal.pntd.0003231 | Hepatotoxicity in Mice of a Novel Anti-parasite Drug Candidate Hydroxymethylnitrofurazone: A Comparison with Benznidazole | Treatment of Chagas disease, caused by Trypanosoma cruzi, relies on nifurtimox and benznidazole (BZL), which present side effects in adult patients, and natural resistance in some parasite strains. Hydroxymethylnitrofurazone (NFOH) is a new drug candidate with demonstrated trypanocidal activity; however, its safety is not known.
HepG2 cells dose response to NFOH and BZL (5–100 µM) was assessed by measurement of ROS, DNA damage and survival. Swiss mice were treated with NFOH or BZL for short-term (ST, 21 d) or long-term (LT, 60 d) periods. Sera levels of cellular injury markers, liver inflammatory and oxidative stress, and fibrotic remodeling were monitored.
HepG2 cells exhibited mild stress, evidenced by increased ROS and DNA damage, in response to NFOH, while BZL at 100 µM concentration induced >33% cell death in 24 h. In mice, NFOH ST treatment resulted in mild-to-no increase in the liver injury biomarkers (GOT, GPT), and liver levels of inflammatory (myeloperoxidase, TNF-α), oxidative (lipid peroxides) and nitrosative (3-nitrotyrosine) stress. These stress responses in NFOH LT treated mice were normalized to control levels. BZL-treated mice exhibited a >5-fold increase in GOT, GPT and TNF-α (LT) and a 20–40% increase in liver levels of MPO activity (ST and LT) in comparison with NFOH-treated mice. The liver inflammatory infiltrate was noted in the order of BZL>vehicle≥NFOH and BZL>NFOH≥vehicle, respectively, after ST and LT treatments. Liver fibrotic remodeling, identified after ST treatment, was in the order of BZL>vehicle>NFOH; lipid deposits, indicative of mitochondrial dysfunction and in the order of NFOH>vehicle>BZL were evidenced after LT treatment.
NFOH induces mild ST hepatotoxicity that is normalized during LT treatment in mice. Our results suggest that additional studies to determine the efficacy and toxicity of NFOH are warranted.
| Hydroxymethylnitrofurazone (NFOH) is a promising drug candidate with demonstrated trypanocidal activity in experimental models of Trypanosoma cruzi infection and chronic disease development. In this study, we monitored the safety of NFOH in established in vitro and in vivo models. Our data show that NFOH did not induce hepatocyte cell death. Short-term or long-term treatment of mice with NFOH did not induce hepatic stress measured by cellular injury, inflammation or fibrosis. Benznidazole, the currently used treatment against acute infection in humans, was more toxic and induced chronic inflammation and liver injury in mice. We conclude that NFOH should be studied further to determine its potential safety for human use as an anti-parasite therapy.
| Chagas disease is endemic in 21 countries, and the World Health Organization estimates that approximately 9-million people are infected by Trypanosoma cruzi [1]. The acute phase lasts for ∼2-months, and is characterized by high parasitemia and fever. Chronic chagasic cardiomyopathy is the most severe clinical consequence of T. cruzi infection, which is detectable in 30% of the patients several years after the primary infection [2].
A considerable advancement in the knowledge about the biology of the T. cruzi parasite has been made in the last few decades (reviewed in [3]). Yet, treatment of Chagas disease still relies on two drugs, namely nifurtimox (NFX, Bayer) and benznidazole (BZL), that were developed during the 1960s and are currently manufactured and distributed by Roche (Rochagan, in Brazil), and Maprimed and ELEA laboratories (Abarax) in Argentina [4]. Since the T. cruzi life cycle involves intracellular division and subsequent parasite release to peripheral blood, a long-term treatment period (usually an oral dose for 60 days) is required. Both drugs are well tolerated by infants and used as standard treatment to control acute infection by T. cruzi in children [5]. However, chronic chagasic cardiomyopathy is a complex malady that involves immune mechanisms as well as parasite persistence. Several studies have noted that treatment with BZL or NFX is not always effective in controlling the chronic disease in chagasic patients. Moreover, these drugs trigger multiple side effects in adults, resulting in noncompliance with long-term treatment [6]. Further, BZL and NFX are not effective against all naturally occurring T. cruzi strains, and some strains have been documented to acquire resistance to these drugs [7]–[9]. Thus, alternative therapeutic drugs that are well-tolerated, safe, and effective against T. cruzi are urgently needed.
Due to their limited commercial potential, development of new drugs for Chagas disease is perhaps not a viable option. Instead, drug repurposing–finding a new indication for an existing drug—has enormous potential in developing a new therapy against parasitic diseases. Nitrofurazone (NF), commercialized as a topical medicine for bactericidal activity against gram-positive and gram-negative bacteria, has recently been shown to inhibit trypanothione reductase, the main enzyme responsible for xenobiotic metabolism in T. cruzi [10]. Subsequently, NF was found to have significant anti-T. cruzi activity [11]. Unfortunately, NF showed toxicity against mammalian cells [12] and long-term treatment with NF induced ovarian cancer in mice and rats [13], [14], spurring the identification of alternative chemicals with specific activity against T. cruzi only. Among the latter, hydroxymethylnitrofurazone (NFOH) was identified as a derivative of NF. The reduction of nitrofurazone is pH-dependent and in acidic medium the hydroxylamine derivative, involving four electrons, is the principal product formed. In aqueous-alkaline medium, the reduction of nitrofurazone occurs in two steps, the first involving one electron to form the nitro-radical anion and the second corresponding to the hydroxylamine derivative formation. NFOH presented the same voltammetric behavior and electroactivity, indicating that the molecular modification performed in NF did not change its capacity to be reduced [15]. We and others have shown NFOH has 2-fold more cytotoxic activity than the parental compound NF against T. cruzi [15], [16]. The mechanism of action of NFOH against T. cruzi is not completely clear; however, like all nitroheterocyclic compounds, it is enzymatically reduced at the nitro group resulting in the generation of nitroanion (RNO2•−) and hydronitroxide (RNHO•−) free radicals [17]. NFOH has also been shown to at least partially interfere with mRNA trans-splicing [16] and cruzipain activity [18] that are essential for parasite invasion as well as differentiation to replicative form.
Before NFOH can be tested and promoted further as an anti-T. cruzi drug for human use, it is essential that we evaluate its safety profile. In a therapeutic regimen administered to T. cruzi infected mice, NFOH and BZL provided comparable control of T. cruzi and survival from infection (84% and 67%, respectively) [11], while NF caused 75% mortality in infected mice. Accordingly, in this study, we have evaluated the liver toxicity of NFOH in comparison with BZL by using in vitro and in vivo models. We treated HepG2 liver cells and mice with the two drugs and assessed inflammation, oxidative stress, and cell survival or tissue remodeling. We chose to treat mice with NFOH for short-term (ST) and long-term (LT) periods that were similar to the recommended regimen for the BZL treatment of children and adults exposed to T. cruzi infection.
All animal experiments were performed according to the National Institutes of Health Guide for Care and Use of Experimental Animals and approved by the Ethical Committee of the National University of Salta and the Animal Care and Use Committee at the UTMB (protocol # 08-05-029).
A hepG2 human hepatocyte cell line was obtained from the American Tissue Culture Collection (Maryland, USA). The cells were cultured in complete DMEM high glucose media (Gibco) supplemented with 10% FBS at 37°C, 5% CO2.
Female Swiss mice were bred at the Instituto de Patologia Experimental mouse facility. Swiss mice have been widely employed as animal models for experimental chemotherapy in Chagas disease [19]. NFOH was kindly provided by Dr. Man Chin Chung (Faculdade de Ciencias Farmaceuticas, Universidade Estadual Paulista, Brazil). BZL (Roche Pharmaceuticals, Brazil) was obtained from the Ministry of Public Health, Province of Salta, Argentina. Mice (30-day old) received NFOH (150 mg/kg/day) or BZL (150 mg/kg/day), suspended in 9% NaCl/5% Tween-80 (vehicle solution), as an oral dose of 100 µl, once a day (six days per week). The selected doses of the BZL and NFOH mimicked the dose per kg and dose regime (two months of daily treatment) of humans. All experiments were carried out in female mice because they develop disease symptoms similar to those seen in human infection. Additionally, we have noted that the female mice exhibited a higher tolerance to infection than did male mice [11]. One set of mice (n = 9/group) received the treatment for 21 days to allow us to determine the acute (short-term, ST) liver toxicity of the drugs. Another set of mice (n = 9/group) received the treatment for 60 days to facilitate our knowledge of the chronic (long-term, LT) liver toxicity of the drugs. Mice given the vehicle solution were used as controls. After treatment, mice were sacrificed, and sera samples and liver tissues stored at −80°C until use.
HepG2 cells were cultured as above, seeded at 7.5×104 cells/well in 96-well microplate, and incubated overnight in complete media at 37°C, 5% CO2. Cells were treated in serum-free medium with NFOH or BZL (5, 50 and 100 µM) for 24–48 h. To examine the drug-induced changes in cell viability and proliferation, after the drug treatment, we incubated the cells for 30 min in the presence of AlamarBlue reagent (Life Technologies/Invitrogen, 10% final concentration). Resazurin, the active ingredient of alamarBlue reagent, is a non-toxic, cell-permeable compound that is blue in color and virtually non-fluorescent. Upon entering cells, resazurin is reduced to resorufin, a compound that is red in color and highly fluorescent. Viable cells continuously convert resazurin to resorufin, increasing the overall fluorescence and color of the media surrounding cells. Fluorescence was measured at Ex540/Em590 nm on a SpectraMax Microplate Reader. Results were analyzed as per the manufacturer's instructions.
HepG2 cells were incubated in the presence or absence of NFOH or BZL for 24 h, as above. For the quantitation of reactive oxygen species (ROS), 5 µM CellROX Green reagent was added to each well in complete media, and cells were incubated at 37°C for 30 minutes. CellROX Green is cell-permeant and non-fluorescent, or very weakly fluorescent, in the reduced state. Upon oxidation, the reagents exhibit strong fluorescence and remain localized within the cell. Cells were washed with PBS and analyzed by flow cytometry.
For assessing the effect of NFOH and BZL in inducing DNA damage, HepG2 cells were treated with NFOH or BZL for 24 h, as above. Cells were harvested, washed with PBS, fixed with 3.7% paraformaldehyde for 15 min at 4°C, and permeabilized with 90% methanol. Cells were then incubated at room temperature for 2 h with mouse anti-8-oxo-dG antibody (250-fold dilution, EMD Milipore, Billerica, MA) and for 30 min with PE-conjugated, anti-mouse IgG (eBioscience, San Diego, CA). Cells stained with isotype-matched IgGs were used as controls. Samples were visualized on an LSRII Fortessa Cell Analyzer, acquiring 30–50,000 events in a live cell gate, and further analysis performed by using FlowJo software (version 7.6.5, Tree-Star, San Carlo, CA).
Frozen liver tissues (25 mg) were homogenized in 0.5 ml of ice-cold lysis buffer (25 mM Tris pH 7.6, 150 mM NaCl, 1% sodium deoxycholate, 1% Igepal CA-630, 0.1% SDS, 10 µl/ml sodium orthovanadate, 10 mM PMSF and 10 µl/ml Sigma protease inhibitor cocktail), and centrifuged at 3000 g for 10 min at 4°C. Supernatants were stored at −80°C, and protein concentration determined by the Bradford method.
The activities of glutamate oxaloacetate transaminase (GOT) and glutamate pyruvate transaminase (GPT), alternatively called aspartate transaminase (AST) and alanine transaminase (ALT), respectively, were determined by using commercially available assay kits (Wiener Lab, Rosario, Argentina). Briefly, for the GOT/AST assay, 50 µl of sera sample or liver homogenate (∼100-µg protein) was added to reagent A containing 12 mM 2-oxoglutarate, 0.18 mM NADH, 420 U/l malate dehydrogenase (MDH), and 600 U/l lactate dehydrogenase (LDH). The reaction was started by adding 80 mM Tris HCl buffer, pH 7.8, containing 240 mM L-aspartate, and the resultant oxaloacetate formation coupled with NADH oxidation by MDH monitored at 340 nm. For the GPT/ALT assay, 50 µl of sample was added to reagent A containing NADH, 2-oxoglutarate (as above) and 1200 U/l LDH. The reaction was started by adding 80 mM Tris HCl buffer, pH 7.8, containing 500 mM L-alanine, and resultant reduction of pyruvate coupled with NADH oxidation by LDH monitored at 340 nm (ε = 6,220 M−1cm−1).
We measured lipid peroxides, a biomarker of oxidative stress [20], by using a LPO Assay Kit (Cayman). Briefly, liver homogenate LPOs were extracted into chloroform, mixed with methanol (1∶1, v/v), and added in triplicate (55 µl/well) to 96-well plates. The reaction was started with addition of 50 µl/well of 4.5 mM FeSO4/0.2 M HCl, 3% ammonium thiocyanate (chromogen) solution. The redox reaction with ferrous ions was stopped after 5 min, and absorbance monitored at 500 nm (standard curve: 0–500 µM 13-hydroperoxy octadecadienoic acid).
The level of protein nitrosylation, an indicator of nitrosative stress, was determined by Western blotting [21]. Tissue homogenates (10-µg protein) were resolved on 10% SDS polyacrylamide gels, and transferred to PVDF membranes by using a vertical Criterion Blotter (Bio-Rad). Membranes were washed in TBS (20.4 mM Tris, 150 mM NaCl, pH 7.6), blocked for 1 h in 5% nonfat milk (NFM), and incubated overnight at 4°C with anti-3-nitrotirosine antibody (clone 2A8.2, 1∶2000, Millipore). After washing with TBS-T (TBS/0.1% Tween-80), membranes were incubated for 1 h at room temperature with HRP-conjugated secondary antibody (1∶10,000, Southern Biotech), and signal was developed with an enhanced chemiluminiscence detection system (GE-Healthcare). Membranes were incubated in stripping buffer (Thermo Scientific) and probed with anti-β-actin antibody (1∶10,000, Sigma) to confirm an equal loading of samples. All antibody dilutions were made in NFM. Spot densitometry for protein bands was carried out using a FluorChem HD2 Image Analyzer (Alpha Innotech).
MPO activity was determined as a biomarker of macrophage/neutrophil activation [20]. Liver homogenates (10 µg protein) were added in triplicate to 0.53 mM o-dianisidine dihydrochloride and 0.15 mM H2O2 in 50 mM KH2PO4/K2HPO4 buffer (pH 6.0). After incubation for 5 min at room temperature, the reaction was stopped with 30% sodium azide, and the change in absorbance was measured at 460 nm. Sample protein content was measured by the Bradford Method, and 1 unit MPO was defined as that degrading 1 n mol H2O2/min at 25°C (ε = 11300 M−1.cm−1).
TNF-α levels were measured as a molecular marker of inflammation. Total RNA was isolated from frozen tissue sections by using the RNeasy plus Kit (Qiagen), and analyzed for quality and quantity on a SpectraMax UV microplate reader. After reverse transcription of 2 µg RNA with poly(dT)18, first-strand cDNA was used as a template in a real-time PCR on an iCycler Thermal Cycler with SYBR-Green Supermix (Bio-Rad) and specific oligonucleotides for TNF-α (5′-GTT CTA TGG CCC AGA CCC TCA CA-3′ and 5′-TAC CAG GGT TTG ACC TCA GC-3′) and GAPDH (5′-TGG CAA AGT GGA GAT TGT TG-3′ and 5′-TTC AGC TCT GGG ATG ACC TT-3′). The PCR Base Line Subtracted Curve Fit mode was applied for Threshold Cycle (Ct) and mRNA level measured by iCycler iQ Real-Time Detection Software (Bio-Rad). The threshold cycle (Ct) values for target mRNA were normalized to GAPDH mRNA, and the relative expression level of TNF-α gene was calculated with the formula n-fold change = 2−ΔCt, where ΔCt represents Ct (TNF-α)−Ct (GAPDH) [22].
Tissue homogenates were also subjected to measurement of TNF-α cytokine by using an optEIA ELISA kit (Pharmingen, San Diego, CA).
Liver tissues were fixed in formalin, embedded in paraffin, and 5-µm sections were stained with hematoxylin and eosin (H&E) and Masson's Trichrome to examine inflammatory infiltrates and collagen deposition, respectively. Cryostat tissue-sections (fixed in OCT cryostat-embedding medium, TissueTek) were stained with Oil red O to examine lipid/fat deposition.
In general, we analyzed each tissue section for >10-microscopic fields (100× magnification), and examined three different tissue sections/mouse (n = 3–4 mice/group) to obtain a semi-quantitative score. Presence of inflammatory cells was scored as I (absent), II (focal or mild, 0–1 foci), III (moderate, ≥2 foci), IV (extensive inflammatory foci, minimal necrosis, and retention of tissue integrity), and V (diffused inflammation with severe tissue necrosis, interstitial edema, and loss of integrity). Inflammatory infiltrates were characterized as diffused or focal depending upon how closely the inflammatory cells were associated. Fibrosis and lipid deposition were assessed by measuring the Masson's Trichrome-stained collagen area (blue) and Oil Red O-stained intrahepatocyte lipid area (red), respectively, as a percentage of the total area by using Simple PCI software (version 6.0; Compix, Sewickley, PA) connected to an Olympus polarizing microscope system (Center Valley, PA). All pixels with blue stain in Masson's trichrome-stained sections and red stain in Oil Red O were selected to build a binary image, and utilized for calculating the percentage of the area occupied by collagen and lipid droplets, respectively. The fibrotic area was further scored as I (<10% of total area), II–III (10–30% of total area), III–IV (30–60% of total area) and V (>60% of total area). Oil red O (intrahepatocyte lipid deposition) was scored as I (absent), II (<10% of total area or patchy distribution of tiny red droplets), III (10–30% of total area or scattered tiny red droplets), and IV (>30% of total area or intense red staining of variable size droplets) [23].
Data (mean ± SD) were derived from at least triplicate observations per sample (n = 9–12 animals/group), confirmed to be normally distributed by a Q-Q test and histogram plot, and analyzed by Student's t test (comparison of two-groups) and 1-way analysis of variance (ANOVA) with a Holm-Sidak test (comparison of multiple groups). Non-parametric Kruskal-Wallis Dunn's test was used to analyze the statistical significance for each cytokine's gene expression. The level of significance is presented by * (normal versus treated; *p<0.05, **p<0.01).
We first evaluated the dose response of the HepG2 cell line to NFOH and BZL. Hepatocytes express cytochrome P450 isoforms, including Cyp2E1 that elicit ROS generation under stress conditions. Additionally, impairment of mitochondrial permeability transition, fatty acid β-oxidation, and inhibition of mitochondrial respiration are all potential mechanisms of ROS generation under stress conditions. We noted a 40–75% increase in CellRox fluorescence (detects intracellular ROS, (Fig. 1A.a) in 27–72% (Fig. 1A.b) of the HepG2 cells treated with 5–100 µM NFOH. The maximal increase in ROS generation was observed when HepG2 cells were treated with 50 µM NFOH treatment. The 8-oxo-2′-deoxyguanosine(8-oxo-dG) is the major product of DNA damage and concentrations of 8-oxo-dG within a cell are used as a measurement of oxidative stress. We noted an up to 33% increase in 8-oxo-dG levels (Fig. 1B.a) in 23% (Fig. 1B.b) of the HepG2 cells treated with increasing concentrations of NFOH. Despite the increase in ROS and DNA damage biomarkers, cell viability, measured by AlamarBlue assay, was not significantly altered by NFOH treatment for 24 h (Fig. 1C.a) or 48 h (Fig. 1C.b). In comparison, HepG2 cells treated with increasing concentrations of BZL showed no statistically significant increase in ROS generation and DNA damage (Fig. 1A&B); however, a cytotoxic response to increasing concentrations of BZL was noted (Fig. 1C). The maximal cytotoxicity (33% cell death) was observed when cells were treated with 100 µM BZL for 24 h (Fig. 1C.a) or 48 h (Fig. 1C.b). Together, these data suggest that NFOH (5–100 µM) induces a mild stress response in hepatopcytes, while BZL at 100 µM concentrations is cytotoxic and causes cell death.
We used a well-established experimental model of Swiss mice for assessing the in vivo cytotoxicity properties of NFOH. We have previously demonstrated NFOH activity against T. cruzi in Swiss mice [11]. These outbred mice display a broader response to drugs than is observed in in-bred C3H/HeN, C57BL/6 and Balb/c mice [19].
We measured elevation in GOT and GPT activities as a general biochemical marker of liver injury after 21 d (ST) and 60 d (LT) treatment with NFOH (controls: vehicle solution). We observed no significant difference in GOT and GPT activities in the sera of mice treated with NFOH or vehicle for the ST and LT period (Fig. 2B&C). Likewise, liver levels of GOT and GPT activities in mice treated with NFOH or vehicle for ST and LT were not statistically different, and comparable to those noted in normal controls (Fig. 2D&E). In comparison, mice treated with BZL for LT exhibited a >5-fold (p<0.05) increase in the sera levels of GPT activity and liver levels of GOT activity (Fig. 2C&D). These data suggest that NFOH is not hepatotoxic, and its treatment for ST or LT is safe. In comparison, LT treatment with BZL was hepatotoxic and caused chronic cellular injury.
The host defense response to NFOH or BZL can result in activation of macrophages and neutrophils that produce oxidative burst, nitric oxide (•NO), and HOCl supported by induction of NADPH oxidase, inducible nitric oxide synthase (iNOS) [23]–[25], and MPO [26], respectively. The cytotoxicity of reactive oxygen species (ROS) results in oxidation of cell constituents, including proteins, lipids, and DNA, which lead to deterioration of cellular structure and function. Additionally, •NO reacts with O2•− and forms peroxynitrite (ONOO−) and peroxynitrous acid (ONOOH) that cause increased protein 3-nirotyrosine (3NT) formation [20], [23], [27].
We investigated host defense responses to NFOH and BZL by measurement of MPO activity and oxidative/nitrosative stress. The level of MPO activity in liver homogenates of mice treated with NFOH for ST or LT was not statistically different when compared to that noted in mice given vehicle only, and was within the basal-level range (100–150 milli-units/mg protein). The BZL-treated mice exhibited a 20–40% increase in liver level of MPO activity when compared to that noted in NFOH-treated mice (Fig. 3A). LPO refers to highly reactive hydroperoxides of saturated and unsaturated lipids, formed by oxidation [28]. NFOH and BZL treatment for ST or LT resulted in no significant increase in the liver levels of LPO formation in comparison to those in control mice treated with vehicle solution (Fig. 3B). The basal level of LPO (<0.25 n mol/mg protein) in all mice, irrespective of ST or LT treatment with NFOH, BZL or vehicle was within the lowest detection range. The polypeptide-bound 3-NT residues, formed by peroxynitrite attack, were monitored by Western blotting. These data showed that the 3-NT level in liver homogenates of mice given ST NFOH or BZL treatment was increased by ∼2.5-fold when compared to those in normal controls, and were similar to those noted in mice given vehicle only (Fig. 3C&D). The 3-NT contents normalized to β-actin were unchanged after LT exposure to NFOH, BZL, or vehicle in treated mice (Fig. 3C&D). Overall, the data presented in Fig. 3 suggested that treatment of mice with NFOH or BZL caused a short-term increase in nitrosative stress that was likely a placebo effect, and, in general, both anti-parasitic drugs did not elicit long-term liver injury by phagocyte activation and oxidative damage in mice.
Next, we determined the effects of NFOH and BZL treatment on liver inflammation. Mice treated with NFOH, BZL or vehicle for ST exhibited a 13–15-fold increase in TNF-α mRNA expression when compared to that noted in normal (untreated) controls (Fig. 4A). The increase in TNF-α expression was also reflected by increased levels of TNF-α protein in liver homogenates of mice treated with NFOH, BZL or vehicle only for ST (range: 58–74-pg/mg protein, Fig. 4B). When given for LT, the NFOH-induced increase in TNF-α mRNA and protein level was decreased by >4-fold, when compared to that noted after ST NFOH treatment and similar to that noted in controls (Fig. 4A&B). Mice given BZL treatment for LT exhibited a 2-fold decline in TNF-α mRNA and a 30% increase in TNF-α protein level when compared to that noted after ST BZL treatment (Fig. 4A&B).
Histological studies showed that ST treatment with NFOH and BZL resulted in a mild increase in liver inflammatory infiltrate, extensive inflammatory lesions (histological score: II–III) being detected in BZL-treated mice followed by vehicle- and NFOH-treated mice (Fig. 5A.a,c,e & 4B). Upon LT treatment, liver inflammatory lesions were noted to be in the order of BZL>NFOH ≥ vehicle. Focal lesions with 0–2 inflammatory foci per microscopic field (histological score: II–IV) were primarily noted in the livers of mice after LT treatment with NFOH or vehicle solution (Fig. 5A.b,d & 5B). The LT treatment with BZL resulted in widespread inflammation in liver, evidenced by finding of >4-inflammatory foci/mf, extensive and diffused inflammation associated with severe tissue necrosis, interstitial edema, and loss of integrity (histological score: III–V, Fig. 5A.f & 4B). Overall, the data presented in Figs. 4&5 suggested to us that NFOH, BZL and vehicle caused a short-term increase in liver levels of TNF-α and inflammatory infiltrate that was likely a placebo effect. When used for LT, NFOH was well-tolerated, while LT treatment with BZL resulted in extensive tissue inflammation in mice.
ROS and inflammatory mediators have been suggested to promote tissue remodeling and dysfunction through diverse mechanisms [23], [29]. We performed histological staining of the liver tissue sections with Masson's Trichrome and oil red O, respectively, for the detection of collagen (Fig. 6) and lipid droplets (Fig. 7). Our data showed the ST with NFOH resulted in a mild degree of collagen deposition in <10% of the tissue area (histological score: II, Fig. 6A.b & 6B, p<0.01) that was significantly lower than that noted in mice given vehicle or BZL treatment. The BZL-treated mice exhibited an up to 30% fibrotic area (histological score: II–III) in liver tissue (Fig. 6A.c & 6B). Very few collagen lesions (<10% fibrosis, histological score: 0–I) indicative of liver remodeling were noted after LT treatment with NFOH, BZL or vehicle solution (Fig. 6A.d–f & 6B).
The extent of lipid deposition, an indicator of mitochondrial dysfunction, was not significantly different after ST treatment with NFOH, BZL or vehicle (Fig. 7A.a–c & 7B). LT treatment with NFOH and BZL resulted in extensive and uniformly scattered lipid droplets of variable size in the liver tissue of mice (20–35% of total area; histological score III–IV) than was noted in vehicle-treated mice (Fig. 7A.d–f & 7B, p<0.01). The extent of lipid deposition appeared to be high in NFOH-treated mice. Together, the results presented in Figs. 6&7 suggest that BZL resulted in mild-moderate acute remodeling of the liver that was replaced by extensive lipid deposition after LT treatment. In comparison, ST treatment with NFOH was not pro-fibrotic, and long-term treatment with NFOH resulted in minimal remodeling and a degree of metabolic dysfunction in the liver of treated mice.
BZL and NFX, drugs that provide the only line of therapy against acute T. cruzi infection, were released to the market without extensive testing for possible adverse effects, which have been reported over the four decades that these drugs have been in use. It was confirmed that NFX has more severe secondary effects than does BZL; these range from alterations of the cellular immune responses [30], peripheral nervous system toxicity, and testicular and ovarian damage to mutagenic effects [17], [31]–[34]. In clinical practice, it is recommended to interrupt anti-T. cruzi treatment when the adverse effects of BZL are detected in adult patients [35]. In contrast, children and newborns show a better tolerance to BZL [36], [37].
Different strategies with the overall aim of finding a cure for Chagas disease are currently under investigation. The major challenges to testing and implementation of therapeutic use of the currently available drugs (BZL and NFX) include the resistance of many of the naturally occurring parasite isolates (e.g. Colombiana) [9] and low efficacy of treatment during indeterminate and chronic phase of disease [6]–[9]. Moreover, additional challenges are faced in immuno-suppressed chagasic patients that become recipients of transplanted organs or are HIV co-infected. The immuno-suppressed patients present a short window of time when they should be treated with anti-parasite drugs. Otherwise, parasite recurrence results in severe acute infection and organ failure. Due to high toxicity concerns, BZL and NFX are not always recommended for treatment of immuno-suppressed patients [38]. In this scenario, NFOH has emerged as a promising compound for its anti-T. cruzi activity, both in vitro and in vivo, and its favorable pharmacological properties [39], [40]. NFOH, derived from hydroxymethyl substitution at the primary amide of nitrofurazone [15], also has a higher solubility in water than does NF and BZL, which likely would facilitate its oral administration. In a murine model of acute T. cruzi infection, NFOH was highly effective in controlling parasitemia evidenced by the observation that infected mice, after NFOH treatment, exhibited no signal for parasite DNA by a highly sensitive PCR approach, and sero-converted with depletion of anti-parasite antibodies [11].
NFOH is a derivative of nitrofurazone (NF). NF is highly toxic and shown to result in single strand DNA breaks [41] and oxidative DNA damage [42], and is considered to be potentially carcinogenic [13]. Considering the high toxicity of NF, it is important to evaluate the toxicity of NFOH in vitro and in vivo before it is recommended for treatment of T. cruzi infection in humans. Accordingly, the present study was designed to examine the adverse effects of NFOH treatment in HepG2 cells and ST and LT treatment of NFOH in mice. We focused on examining the effects of NFOH on hepatocytes and liver because the liver is the main detoxifying organ in mammals. Metabolism of xenobiotics in the liver involves phase I and phase II reactions that add hydroxyl and methyl groups, respectively, to a given compound. NFOH is a nitrofurazone with an N-hydroxymethylation at the primary amide and was anticipated to cause significantly reduced toxicity [11].
Our in vitro studies evaluating the dose response of HepG2 cells clearly demonstrated that NFOH at higher concentrations (50–100 µM) induced mild stress as was evidenced by the observation of ROS production and DNA damage (Fig. 1A&B). However, the NFOH-induced stress was controlled as we observed no cell death in NFOH-treated HepG2 cells. Under similar conditions, cytotoxicity of BZL was evidenced by induction of cell death in 33% of the cells (Fig. 1C). Others have shown the nifurtimox and BZL inhibited DNA and protein synthesis in hepatocytes [43]. In murine studies, the selected doses for toxicity evaluation were the same as those that we have previously tested in mouse models of acute and chronic T. cruzi infection. We included mice treated with BZL and vehicle (NaCl/Tween-80) as controls. Our data showed a moderate increase in sera levels of GOT and GPT (Fig. 1) after ST and LT treatment with NFOH that was similar to that noted in mice treated with vehicle only. Further, MPO activity and LPO production, measured as markers of neutrophil activation and macrophage oxidative burst, were present at the lowest limits of detection in all mice given NFOH or vehicle throughout the treatment schedule (Fig. 3A&B). These data, along with the observation of no significant change in TNFα mRNA and protein level (Fig. 4) suggested to us that NFOH did not induce hepatic stress associated with cellular injury, oxidative stress, and innate immune cell activation in vitro or in vivo after ST or LT treatment. Our observation of a >5-fold increase in hepatic levels of MPO activity and TNF-α expression in the livers of mice within 3 d after a single dose treatment with 2,3,7,8-tetrachlorodibenzo-p-dioxin (TCDD, 20-µg/Kg/100 µl peanut oil) in other studies suggest that when compared with TCDD effects, NFOH has no acute liver toxicity. The notion of NFOH being liver-safe is also supported by the observation of a higher degree of GOT release (5-fold), MPO activity (20–40%), and TNF-α protein levels (up to 20%) in BZL-treated mice in this study (Figs. 2–4). Others have also shown the BZL toxicity by alterations in mitochondrial function in liver of treated rats [44].
We corroborated the biochemical findings (Figs. 1–4) of NFOH safety by histological observations of tissue inflammatory infiltrate, fibrosis, and lipid deposition in the livers of mice treated for ST or LT with NFOH, and compared the findings with BZL-treated mice (Figs. 5–7). NFOH-treated mice consistently exhibited none-to-low levels of inflammatory infiltrate and fibrotic lesions in liver tissue sections that were similar to or lower than was noted in vehicle-treated mice (Figs. 5–7). The extent of liver inflammatory infiltrate was significantly higher in BZL-treated mice, especially after LT treatment (Figs. 5). Likewise, BZL-treated mice exhibited an acute liver fibrosis (Fig. 6) similar to what we have noted in TCDD-treated mice (unpublished data). The extent of lipid deposition in liver tissue after LT treatment with BZL and NFOH was comparable to lipid deposits provoked within three days after treatment with a single dose of TCDD. Though not withdrawn from market, BZL has been demonstrated to be more toxic than NFOH with regard to elicitation of liver inflammatory responses, fibrosis, and a cellular dysfunction at mitochondrial level, both in our data presented in this study and other published reports [17], [33]. We surmise that NFOH only causes a mild transient hepatic injury similar to that caused by vehicle treatment only in mice. Our results encourage further research on carcinogenicity, and mechanism of action of NFOH to address its potential safety for human use as an anti-parasite therapy.
In conclusion, our results showed that ST and LT treatment with NFOH elicited similar or lower levels of liver oxidative stress, inflammation and tissue remodeling responses, when compared to that noted by a similar regimen of treatment with vehicle only. In comparison, BZL that has been used for the treatment of human chagasic patients was more toxic and induced chronic inflammation and liver injury. Our data provide the impetus for future studies focusing on further characterization of anti-parasite efficacy, toxicity, and carcinogenicity of NFOH, aiming to determine its potential safety to be considered as a drug candidate for the treatment of Chagas disease.
|
10.1371/journal.ppat.1005230 | Antigenic Characterization of the HCMV gH/gL/gO and Pentamer Cell Entry Complexes Reveals Binding Sites for Potently Neutralizing Human Antibodies | Human Cytomegalovirus (HCMV) is a major cause of morbidity and mortality in transplant patients and in fetuses following congenital infection. The glycoprotein complexes gH/gL/gO and gH/gL/UL128/UL130/UL131A (Pentamer) are required for HCMV entry in fibroblasts and endothelial/epithelial cells, respectively, and are targeted by potently neutralizing antibodies in the infected host. Using purified soluble forms of gH/gL/gO and Pentamer as well as a panel of naturally elicited human monoclonal antibodies, we determined the location of key neutralizing epitopes on the gH/gL/gO and Pentamer surfaces. Mass Spectrometry (MS) coupled to Chemical Crosslinking or to Hydrogen Deuterium Exchange was used to define residues that are either in proximity or part of neutralizing epitopes on the glycoprotein complexes. We also determined the molecular architecture of the gH/gL/gO- and Pentamer-antibody complexes by Electron Microscopy (EM) and 3D reconstructions. The EM analysis revealed that the Pentamer specific neutralizing antibodies bind to two opposite surfaces of the complex, suggesting that they may neutralize infection by different mechanisms. Together, our data identify the location of neutralizing antibodies binding sites on the gH/gL/gO and Pentamer complexes and provide a framework for the development of antibodies and vaccines against HCMV.
| Human Cytomegalovirus (HCMV) is a double stranded DNA, enveloped virus infecting >60% of the population worldwide. Typically asymptomatic in healthy adults, HCMV infection causes morbidity and mortality in immunocompromised patients and is the most common viral cause of birth defects in industrialized countries. Despite more than 30 years of research, however, no vaccine against HCMV is available. HCMV utilizes two distinct glycoprotein complexes, gH/gL/gO and gH/gL/UL128/UL130/UL131A (Pentamer), to enter fibroblast and endothelial/epithelial cells, respectively and both are neutralizing antibodies targets. We used orthogonal techniques to study the interaction between gH/gL/gO or Pentamer and a panel of naturally occurring human neutralizing antibodies. The results of this analysis identify three neutralizing epitopes in gH, which are conserved in both glycoproteins complexes, and a different subset of five neutralizing sites in the UL128/Ul130/Ul131A (ULs) portion of the Pentamer. Moreover, EM analysis defines two distinct surfaces targeted by neutralizing antibodies on the ULs suggesting different neutralization mechanisms. Our results reveal regions of the gH/gL/gO and Pentamer complexes important for eliciting strong neutralizing responses in humans and for function in viral entry. Together our data will guide the development of therapeutic monoclonal antibodies and vaccines against HCMV.
| Human Cytomegalovirus (HCMV), a member of the Betaherpesvirinae sub-family of Herpesviridae, infects 40–60% of the human adult population [1]. Similar to other herpesviruses, after primary infection, HCMV becomes latent and persists for the host’s life span. Reactivation is generally asymptomatic in immune-competent individuals. However, primary infection or reactivation can cause severe disease or death in immuno-suppressed hosts such as solid organ and hematopoietic stem cell (HSC) transplant patients and individuals with HIV infection [2–5]. HCMV is also the most common cause of viral induced birth defects affecting 0.2% of the newborns in industrialized countries [6–8]. For this reason the development of an effective HCMV vaccine able to prevent congenital infection was identified as a top priority by the Institute of Medicine [9–13]. Nevertheless, despite more than 20 years of vaccine research, there is no vaccine available against HCMV.
HCMV can infect a broad spectrum of cell types including epithelial and endothelial cells, fibroblasts, dendritic cells, neurons, and leukocytes [14,15]. The virus uses several envelope glycoprotein complexes to enter cells. Like other herpesviruses, glycoprotein B (gB), the viral fusion protein, and a gH/gL containing complex are required for HCMV cell entry [16,17]. Specifically, HCMV entry into epithelial and endothelial cells requires a pentameric glycoprotein complex (Pentamer) comprised of the gH, gL, UL128, UL130, and UL131A subunits (from now on referred to as ULs) [18,19] and is dependent on low pH [19]. Instead, entry into fibroblasts requires the gH/gL/gO complex and viral membrane fusion is thought to occur at the plasma membrane [20–23].
Recent data indicate that all HCMV strains contain gH/gL/gO and Pentamer complexes on the viral envelope and little, if any, unbound gH/gL [24]. Mutations in the UL131A-UL128 gene locus occur spontaneously within just a few passages of wild-type (WT) HCMV in fibroblasts and are sufficient to eliminate epithelial/endothelial cell tropism [25]. Conversely, deletion of gO from the HCMV genome compromises virion assembly and replication in fibroblasts [26]. Of note, cell surface Pentamer over-expression prevents HCMV entry into epithelial cells, but not into fibroblasts, presumably through host protein sequestration, indicating the presence of a cell-type specific Pentamer-receptor [27].
We have recently described the biochemical characterization of HCMV gH/gL, gH/gL/gO and Pentamer and defined the overall architecture of each complex on its own or bound to a Fab fragment from the neutralizing antibody MSL-109 [28]. Electron microscopy (EM) data showed that like HSV-2 gH/gL, HCMV gH/gL adopts a boot shaped structure. A similar structure was seen when HCMV gH/gL is complexed with gO or with the ULs. The EM analysis also revealed that gO, in gH/gL/gO, and the ULs, in Pentamer, bind to the same site at the N-terminal end of the gH/gL heterodimer, thus forming mutually exclusive cell entry complexes. Consistent with these observations mass spectrometry (MS) studies demonstrated that the same cysteine in gL, C144, forms disulfide bridges with UL128-C162 in Pentamer, gO-C351 in gH/gL/gO or with the same cysteine in homodimers of gH/gL heterodimers. Notably, mutation of gL-C144S was sufficient to prevent formation of covalent complexes between gH/gL and either gO or UL128 in gH/gL/gO and Pentamer, respectively. The same mutation resulted in formation of monomeric gH/gL heterodimers [28].
Highly potent monoclonal antibodies targeting conformational epitopes of the Pentamer were initially isolated from the memory B-cell repertoire of HCMV immune donors and later from rabbits and mice immunized with an experimental vaccine virus in which the expression of the Pentamer was restored or an adjuvanted Pentamer protein, respectively [29–31]. These antibodies were a thousand-fold more potent than antibodies against gB or the gH/gL complex and were extraordinarily effective in neutralizing HCMV infection of epithelial and endothelial cells. Recently, different groups have demonstrated that immunization with adjuvanted Pentamer protein or vectors expressing the Pentamer elicit a strong neutralizing response in small animals and rhesus macaques [31–34].
Despite the fact that these data indicate that Pentamer represents a key antigenic target for HCMV vaccine development, limited mapping data are available to describe the sites on the Pentamer that are recognized by neutralizing antibodies and responsible for eliciting its potent neutralizing response. In this study we describe the interaction between HCMV gH/gL/gO and Pentamer with naturally-elicited potently neutralizing human monoclonal antibodies [29] using a combination of EM and MS. Our data identify HCMV gH/gL/gO and Pentamer epitopes important for generating strong neutralizing responses providing a framework for the development of effective HCMV vaccines and antibody therapeutics.
Two neutralizing monoclonal antibodies, 13H11 and 3G16, isolated from immortalized memory B-cells of HCMV-immune donors have been shown to bind to HCMV glycoprotein H (gH) and prevent infection [29,35]. We initially investigated binding of gH/gL to these two antibodies as well as to MSL-109, an antibody isolated from the spleen of an HCMV seropositive individual [36,37], whose gH binding site we previously characterized [36–38]. We also generated Fab fragments for the three antibodies for additional binding and structural studies.
An ELISA assay was initially used to study the interaction between the three monoclonal antibodies and the gH/gL homodimer (Fig 1). For each of the antibodies we were able to confirm gH/gL binding. In addition, we observed binding competition between MSL-109 and 3G16. 13H11, however, was able to form a ternary complex with either gH/gL/MSL-109 or gH/gL/3G16 (Fig 1). Equivalent results were obtained with the corresponding Fabs by gel shift assay using either gH/gL homodimer or the gH/gL-C144S mutant [28], which is mostly monomeric in solution, and the Fab fragments of the three antibodies (S1 Fig).
To characterize the binding of these antibodies to gH, we reconstituted gH/gL/Fab and gH/gL/gO/Fab complexes and used EM and single particle analysis to identify the position of the specific antibody as compared to the unbound complex. Initial EM analysis revealed the same binding sites for the Fabs in both gH/gL-C144S monomer mutant and gH/gL/gO complexes. Only the latter was studied further due to the higher image quality (Fig 2A). EM analysis of the gH/gL/gO/3G16 complex demonstrated that the epitope of this antibody is localized in the C-terminal domain of gH. EM analysis of the previously characterized MSL-109 bound complex [28], showed that MSL-109 binds the external part of the gH/gL/gO complex ‘heel’ (Fig 2A). Comparison of the 3G16 and MSL-109 complexes suggest that the two Fabs cannot bind simultaneously to gH due to the spatial overlap of their constant regions. This result is consistent with the competition effect observed in ELISA and gel shift assay (Fig 1 and S1 Fig). Finally, analysis of 13H11-bound gH/gL/gO showed that the epitope of this antibody is localized internally at the gH/gL kinked region, opposite to the epitopes recognized by 3G16 and MSL-109. Analysis of gH/gL/gO bound to 3G16/13H11 or MSL-109/13H11 further confirms these conclusions (Fig 2A).
To support the results of our 2D EM analysis of the gH/gL/gO/Fab complexes, we determined a 3D-reconstruction of gH/gL/gO bound to 3G16 Fab and to both 3G16 and 13H11 Fabs (Fig 2B and 2C). Since these complexes adopted a strong preferential view we calculated the 3D structure using the Random Conical Tilt (RCT) method to ~19 Å and 29 Å resolution, respectively [39]. The reconstructed density maps of the gH/gL/gO-Fab complexes allowed fitting of an HCMV gH/gL model based on the HSV gH/gL crystal structure [40]. Models obtained for the 3G16 and 13H11 Fab fitted well into the density map and were located respectively proximal to the gH C-terminal domain and close to the gH-kinked region opposite to 3G16 and MSL-109. Additional density emerging from the N-terminal region of gH/gL describes the overall shape of the gO subunit as we previously reported [28].
Hydrogen-deuterium exchange coupled to MS (HDX-MS) was used to identify residues that are part of the gH/gL epitopes targeted by 3G16 and 13H11 Fabs. To simplify the MS analysis, we expressed the monomeric gH/gL-C144S mutant in HEK293S GnTI-/- cells as previously described [28]. The purified complex was deuterated either alone or in the presence of each Fab and the averaged deuterium exchange behaviors of 118 overlapping peptides of gH and gL were investigated. No gL peptides were observed having a difference in deuterium incorporation when gH/gL was bound to any of the Fabs confirming that these Fabs bind the gH subunit.
3G16 Fab binding induced significant reduction of deuterium uptake in two gH peptides, 677–684 and 705–725 (Fig 3A), demonstrating that 3G16 binds to a composite, not linear, epitope. In addition, we were able to narrow down the second peptide to the unique amino acid sequence 705–708 since two additional gH overlapping peptides, 709–725 and 711–725, presented the same deuterium incorporation in the presence or absence of 3G16 Fab (Fig 3A).
Only the gH peptide 238–247 showed a reduction in deuterium uptake in presence of the 13H11 Fab (Fig 3B). Although the difference is of low amplitude, it was highly reproducible in three independent experiments. A similar analysis was previously performed with the gH/gL/MSL-109 complex, where the peptides 403–419 and 442–446 showed a reduction in deuterium uptake and were identified to be part of the MSL-109 epitope ([28]; Fig 3C).
Overall the HDX-MS results were consistent with the EM analysis in identifying peptides in the C-terminal domain of gH for gH/gL/3G16 and the central region of gH for gH/gL/13H11, respectively (Fig 3C). In addition, the HDX-MS data together with the competition and EM data suggest that the two Fabs compete via their constant domains but not via binding to the same epitope (Fig 3).
In summary, the combination of EM and HDX-MS identified two new neutralizing epitopes in gH/gL.
We then focused our attention on Pentamer-specific neutralizing antibodies previously isolated from the memory B-cell repertoire of HCMV immune donors [29]. These antibodies were previously assigned to seven distinct sites based on binding and cross-competition experiments on cells transfected with different combinations of the gH, gL, UL128, UL130, UL131A genes [29] (Fig 4A). Importantly, this study identified subunits of the Pentamer that are required for binding to antibodies belonging to each of these groups. To identify more precisely the location of the epitopes of these antibodies on the intact complex we used negative stain EM analysis. We reconstituted a Pentamer/3G16 complex bound to additional individual Fabs representing each site and analyzed them by EM and single particle analysis (Fig 4B). In preliminary experiments, 3G16 was found to improve the overall quality of the EM images of the Pentamer/Fabs complexes likely by decreasing flexibly of the gH C-terminal domain.
We started our analysis with 15D8 (Site 1), which has been shown to cross-react specifically with UL128 [29] (Fig 4A). Reference free 2D analysis of Pentamer/3G16/15D8 showed that the 15D8 Fab binds to a region contiguous to the N-terminal portion of the gH/gL complex (Fig 4B), consistent with the proposed proximity between UL128 and gL [28]. Analysis of Pentamer/3G16 bound to 10F7 (Site 2), specific for UL130/UL131A, indicated that this Fab recognizes a distal region of the Pentamer forming a separated domain protruding from gH/gL (Fig 4A and 4B). We used 4N10, also reported by Macagno and colleagues [29] to cross-react with UL130/UL131A, as a representative Fab for Site 3 (Fig 4A). EM analysis showed that 4N10 binding occurs on a position very close to Site 2 but the Fab is oriented at a very different angle (Fig 4B). Binding of 10P3 (Site 4), specific for UL130/UL131A, occurs on the side opposite to Sites 2 and 3 and close to Site 1 (Fig 4A and 4B). 2C12 (Site 5), specific for UL128/UL130/UL131A, was bound to the most distal portion of the Pentamer from the gH/gL region whereas 7I13 (Site 6), an antibody specific for UL128/UL130/UL131A, bound to a similar position as Site 4 (Fig 4A and 4B). Finally, analysis of 8I21 (Site 7), specific for gH/gL/UL128/UL130, showed a very similar localization described for Sites 2 and 3 (Fig 4A and 4B).
Together, our EM analysis identifies two major surfaces on UL128/UL130/UL131A as targets for neutralizing antibody binding. One surface on one side of the Pentamer includes Sites 1, 4 and 6, whereas a second one, on the opposite side, includes Sites 2, 3, 5 and 7. Of note with the exception of Site 5 most of the antibodies seem to cluster in a similar area close to the bottom of the V-shaped part of the UL extension (Fig 4B).
To gain additional insights into the location of the different epitopes, we established a Multiplex assay to assess competition for Pentamer binding among the different monoclonal antibodies (Fig 5 and S2 Fig). As suggested by the EM 2D class averages, 10P3 (Site 4) and 7I13 (Site 6), which appear to bind to the same location on the complex, competed for Pentamer binding in Multiplex (Fig 5A). However, Multiplex analysis showed that neither of the Sites 4 nor 6 antibodies cross-compete with 15D8 (Site 1) despite EM showing that their binding sites are in proximity to each other on the Pentamer surface (Fig 4B). Similarly, 10F7 (Site 2) and 4N10 (Site 3) did not show cross-competition in Multiplex (Fig 5A) demonstrating that, despite close spatial proximity (Fig 4B), these epitopes are not actually overlapping. Finally, 8I21 (Site 7), which occupies a very similar localization described for Sites 2 and 3 (Fig 4B), cross-competed with 4N10 (Site 3) but not 10F7 (Site 2) (Fig 5A), indicating that Site 2 and Sites 3–7 are structurally distinct. In summary, the combination of EM and binding studies demonstrates that this panel of monoclonal antibodies binds to five distinct non-competing sites on the Pentamer surface (i.e. Sites 1, 2, 3–7, 4–6, 5; Fig 4C).
Finally we tested if the antibodies that did not compete as pairs could bind to Pentamer at the same time. Pentamer was mixed with antibodies targeting four of the non-competing Pentamer specific sites and competition with an antibody binding to the fifth site was assessed with the Multiplex assay (S3 Fig). Interestingly, this analysis revealed that simultaneous binding of four antibodies interfered with binding of the fifth one, though the magnitude of the effect differed across the five antibodies tested (S3 Fig). The antibodies that bind to these five sites do not compete among each other when tested as pairs (S3 Fig) suggesting that the competition effect observed when multiple antibodies are mixed together with Pentamer may be caused by allosteric effects though steric hindrance effects cannot be excluded completely.
To gain additional structural insights into the interaction between Pentamer and this set of human neutralizing antibodies we determined RCT 3D-reconstructions of Pentamer bound to the corresponding Fabs. For these reconstructions we selected only Fabs that do not compete among each other for Pentamer binding to define structurally distinct epitopes. Therefore RCTs of Pentamer/3G16/10P3/8I21, Pentamer/3G16/10F7 and Pentamer/3G16/15D8/2C12 complexes at a resolution of respectively ~31 Å, ~30 Å and ~39 Å were obtained. These reconstructions were sufficient to describe the spatial organization of the Pentamer components and define the region of neutralizing antibodies interaction with the ULs (Fig 6A–6C). Also in this case, an HSV-based model of HCMV gH/gL was first fitted into the density maps. Extra densities emerging from the gH C-terminal region and from the UL subunits were consistent with the size of Fabs and each well accommodated a Fab model (Fig 6A–6C). The 2D reconstructions were used to assign the densities protruding from the Pentamer to each of the Fabs. A comparison of the RCT structures of gH/gL/gO and Pentamer revealed a similar architecture of the gH/gL portion bound to 3G16 and confirmed that the gO and UL subunits interact with a common surface on gH/gL (S4 Fig).
Finally, we superimposed the EM 3D reconstructions by manually aligning the surfaces corresponding to their gH/gL/3G16 regions. The Pentamer/3G16/10F7 and Pentamer/3G16/15D8/2C12 reconstructions gave the most reliable superposition. This analysis revealed a good structural conservation of the gH/gL portion of the complex and of its interaction with 3G16. However, this superposition also showed that the ULs are rotated to a different angle relative to gH/gL in the two complexes (Fig 6D). Therefore, antibody binding seems to lock the Pentamer in similar though distinct conformations. The differences in the orientation of the ULs among some of the Pentamer/antibody complexes may in part explain the competition observed when multiple antibodies are mixed simultaneously with Pentamer. Together the EM analysis and the competition data define the location of Pentamer-specific neutralizing antibody binding sites and suggest that the ULs portion of the Pentamer is conformationally flexible.
To identify regions of the Pentamer recognized by neutralizing antibodies at the amino acid level, we carried out chemical cross-linking coupled to MS analysis on purified Pentamer and Pentamer/Fab complexes. Disulfosuccinimidyl glutarate (DSSG), a homo-bifunctional cross-linker that reacts with the primary amines of lysines spaced up to 25 Å apart, was used as a cross-linking agent. In preliminary experiments, we observed that UL131A-K27 cross-linked to a large number of lysines on the Pentamer suggesting that this lysine is in a highly flexible part of the molecule. We therefore introduced the UL131A-K27R mutation to simplify the MS analysis and confirmed that this mutation does not affect antibody binding.
Cross-linking of unbound Pentamer showed an intricate network of interactions within all the components (S1 Table; Fig 7A). In particular, a region of gH between lysine 130 and 452 contains many sites of inter-molecular cross-linking. Lysine 283 at the N-terminal end of gH cross-linked UL130-K131-145/154-157 and UL128-K117. In turn, several lysines in UL128 cross-linked to a region of UL130 extending from lysine 108 to lysine 131. This 20 amino acid fragment in UL130 is involved in interactions with lysines in gH as well as with UL131A-K103 suggesting a central hub of interaction in the Pentamer. We also observed a large number of lysines in UL128 being internally cross-linked, indicating that this component is exposed on the complex and could potentially represent a target for antibody neutralization. Finally, no lysines were cross-linked in the N-terminus of UL130 suggesting a more buried location in the Pentamer. Taken together the cross-linking data indicate a strong interconnection between the ULs and the N-terminal region of gH/gL confirming our biochemical and EM results.
A similar approach was utilized on Pentamer bound to neutralizing Fabs (S1 Table; Fig 7B). Despite numerous attempts, no cross-linking was observed for Sites 3 and 7 antibodies, suggesting lack of exposed lysines in their proximity. All of the other Fabs cross-linked lysines in UL128 and UL130 (S1 Table) and some also cross-linked a lysine on UL131A, (UL131A-K103). Of note, peptides containing or adjacent to crosslinked lysines in the complexes with Fabs 10P3 (Site 4) (UL128-K165 and UL130-K48), 2C12 and 7I13 (Site 5 and 6 respectively) (UL131A-K103) were previously reported to raise neutralizing antibodies in immunization experiments in mice [41] and could therefore be part of epitopes for additional neutralizing antibodies [31].
Overall the cross-linking data identifies regions of the complex that are exposed and available for antibody binding and suggest that most antibodies examined here bind close to UL128 and UL130. The analysis also identifies a new set of UL subunit peptides that may be potentially used to raise and isolate new HCMV neutralizing monoclonal antibodies or as a component of a HCMV vaccine.
The gH/gL- and Pentamer-specific antibodies were tested for binding to purified complexes by SPR. Most of the antibodies bound very tightly to the recombinant proteins with KDs (M) between 1.0e-9 and 6.0e-11 (Table 1). gH/gL specific antibodies bound similarly to recombinant gH/gL/gO and Pentamer confirming that these epitopes are structurally conserved in the two complexes. We also noted that the gH/gL-specific antibodies 3G16 and 13H11 bind to the recombinant complexes equally well although they neutralize infection of epithelial cells at titers 10–100 fold less than Pentamer-specific antibodies [29]. This discrepancy likely reflects differences in the neutralization mechanism between gH/gL- and the UL-specific antibodies.
Here we have characterized the interaction between gH/gL/gO and Pentamer with a panel of naturally-elicited human neutralizing monoclonal antibodies [29]. Our data confirm the presence of two major regions for neutralizing antibody binding. The first region involves the gH part of the complexes similar to that described for EBV and HSV [42,43]. We show that this region, and its interaction with neutralizing antibodies, is structurally conserved in gH/gL/gO and Pentamer. It is important to note that the gH/gL region is targeted by antibodies neutralizing HCMV entry in all cell types albeit less potently than the Pentamer-specific antibodies in epithelial and endothelial cells [29]. This region may play a role in steps of the viral fusion process, such as gB binding. In this respect we observed that the binding site for 13H11 is in proximity of a site proposed to mediate gB binding in HSV-2 gH/gL [40].
The second region comprises different neutralizing sites on UL128/UL130/UL131A and it likely includes the binding site for the entry receptor for epithelial/endothelial cells. EM analysis and competition studies using purified Pentamer show that the neutralizing sites can be subdivided into two groups defining two distinct surfaces in the UL region of the Pentamer. One surface includes Sites 1, 4, and 6 whereas the second surface includes Sites 2, 3, 5 and 7. These two surfaces are on opposite sides of the elbow formed by the V-shaped ULs component of the Pentamer. It seems unlikely that both surfaces are involved in receptor binding and we speculate that some of the antibodies may instead prevent conformational changes required for activation of membrane fusion. Consistent with this hypothesis comparison of 3D EM analysis of Pentamer/Fab complexes suggest that the ULs portion of the Pentamer is flexible and trapped in different conformations by different Fab combinations. Alternatively, one of the two areas targeted by the neutralizing antibodies may interfere with the interaction of Pentamer with additional cellular and/or viral proteins following receptor binding.
Crosslinking coupled to MS analysis of Pentamer bound to representative neutralizing antibodies for each of the sites on the Pentamer showed that most of them bind in proximity of UL128 and UL130 suggesting that UL131 is partially buried in the complex. Indeed western blot analysis using Cytotect, human IgGs from HCMV positive individuals that contain a high fraction of Pentamer specific neutralizing antibodies, or a polyclonal sera from rabbits immunized with MF59 adjuvanted Pentamer protein fail to detect UL131 in the purified complex (S5 Fig).
Immunization of mice with a subunit Pentamer vaccine adjuvanted with MF59 was shown to raise high titers of antibodies that competed with representative antibodies binding each of the neutralizing sites described above [33,34]. The subunit vaccine raised neutralizing responses in mice that were comparable to or higher than those observed in infected human subjects [33,34]. Therefore, the Pentamer subunit appears to expose all the functionally relevant neutralizing sites including those that are hidden in the context of the native virus particle (e.g. MSL-109 binding site; [37]).
In conclusion, using EM, HDX and chemical cross-linking with MS analysis, we have identified sites on the Pentamer that are targeted by neutralizing antibodies. Together these findings will facilitate the analysis of the human antibody response to HCMV infection and vaccination. Finally, these data will support the development of a new generation of Pentamer-based HCMV vaccines with the potential to elicit potent and clinically relevant protective neutralizing antibody responses.
Human HCMV Merlin strain gH, gL, gO, UL128, UL130 and UL131A genes synthesized by GeneArt (Regensburg, Germany) codon-modified for expression in Homo sapiens, and carrying an optimal Kozak sequence immediately 5’ of each gene, were sub-cloned using NheI/KpnI restriction sites into the pcDNA3.1(-)A plasmid expression vector (Invitrogen, Life Technologies; Carlsbad, CA). The gH gene was terminated at amino acid 715 and therefore was missing the transmembrane domain and cytoplasmic tail. UL130, gO and all the Fab VH chains indicated in the text, and characterized elsewhere [29], were fused to a C-terminal TEV-cleavable Strep-tag II for purification purposes.
293EBNA cells (Invitrogen, Life Technologies; Carlsbad, CA) were transfected with individual plasmids encoding each subunit. Pentamer, gH/gL/gO as well as all the Fabs were purified using the double Strep-tag at the C-terminus of UL130, gO and of the VH chain, respectively. Affinity purified complexes were eluted from the resin using 5 mM desthiobiotin containing elution buffer (25 mM Tris pH 7.5, 300 mM NaCl). The final step of the purification of both HCMV complexes and Fabs consisted of size exclusion chromatography (SEC) on a Superose 6 PC 3.2/30 or Superdex 200/30 column, equilibrated in 25 mM Tris pH 7.5, 300 mM NaCl. Complexes between Fabs and either gH/gL/gO or Pentamer were generated by incubation of these proteins on ice for 2 h using a 1.5-fold molar excess of Fab and purified by SEC to remove Fab excess.
Purified gH/gL monomer or dimer protein (previously described in [28]) were incubated with 3G16, MSL-109 or 13H11 Fabs in various combinations at a ratio of 10 μg gH/gL to 5 μg each Fab for 1 h at room temperature (RT) before being resolved on NativePAGE Novex 3–12% Bis-Tris Protein Gels (Invitrogen Inc.) along with NativeMark Unstained Protein Standard (Invitrogen Inc.). The gels were subsequently stained with Coomassie Blue to visualize the bands.
HCMV Pentamer or Pentamer/Fab complexes (36 pmol) were mixed with a 600-fold excess of isotope-labeled cross-linker di-(sulfosuccinimidyl)-glutarate (1:1 mixture of light DSSG-d0 and heavy DSSG-d6) (Creative Molecules, Victoria, BC, Canada) in a final volume of 50 μl of 10 mM HEPES, pH 7.5, 300 mM NaCl at RT. The reaction was stopped after 30 min by adding 5 μl of 1 M ammonium bicarbonate.
Cross-linked proteins were reduced with 5 mM TCEP (tris (2-carboxyethyl) phosphine; Thermo Scientific, Rockford, IL, USA) at 37°C for 30 min and alkylated with 10 mM iodoacetamide (Sigma-Aldrich, St. Louis, MO, USA) for 30 min in the dark. Proteins were first digested with endoproteinase Lys-C (Wako, Neuss, Germany) at an enzyme-to-substrate ratio of 1:100 (w/w) for 3 h at 37°C and subsequently with sequencing grade trypsin (Promega, Madison, WI, USA) at an enzyme-to-substrate ratio of 1:50 (w/w) at 37°C overnight. Peptides were acidified with 2% formic acid (Sigma-Aldrich) and purified by solid-phase extraction (SPE) using C18 cartridges (Sep-Pak; Waters, Milford, MA, USA). The SPE eluate was evaporated to dryness and reconstituted in 20 μl of SEC mobile phase (water/acetonitrile/TFA, 70:30:0.1). 15 μl were injected on a GE Healthcare (Uppsala, Sweden) Äkta micro system. Peptides were separated on a Superdex Peptide PC 3.2/30 column (300×3.2 mm) at a flow rate of 50 μl min−1 using the SEC mobile phase [44]. Two-minute fractions (100 μl) were collected into 96-well plates.
LC-MS/MS analysis was carried out on an Eksigent 1D-NanoLC-Ultra system connected to a Thermo LTQ Orbitrap XL mass spectrometer equipped with a standard nanoelectrospray source. SEC fractions were reconstituted in mobile phase A (water/acetonitrile/formic acid, 97:3:0.1). A fraction corresponding to an estimated 1 μg of peptides was injected onto a 11 cm × 0.075 mm I.D. column packed in house with Michrom Magic C18 material (3 μm particle size, 200 Å pore size). Peptides were separated at a flow rate of 300 nl min−1 ramping a gradient from 5% to 35% mobile phase B (water/acetonitrile/formic acid, 3:97:0.1).
Cross-linked peptides were identified using an in-house version of the dedicated search engine, xQuest [45]. Tandem mass spectra of precursors differing in mass by 6.037660 Da (difference between DSSG-d0 and DSS-d6) were paired if they had a charge state of 3+ to 8+ and were triggered within 2.5 min of each other. These spectra were then searched against a pre-processed fasta database containing the target sequences. A valid identification of the cross-linked peptides required an xQuest score of at least 16 (corresponding to a false discovery rate of > 5%) and at least four bond cleavages in total or three in a series for each peptide and a minimum peptide length of six amino acids.
EM grids were prepared by placing five microliters of purified sample on a freshly glow discharged 400-mesh copper grid covered with a thin layer of continuous carbon. After 30 sec of incubation, the grid was stained with 5 drops of a freshly prepared 2% (w/v) uranyl formate solution. Samples were imaged on a Tecnai Spirit T12 transmission electron microscope operating at 120 keV with a magnification of 49,000× (1.57 Å/pixel at the detector level) using a defocus range of −0.8 to −1.2 μm. Images were recorded on a Gatan 4096 × 4096 pixel CCD camera under low-dose conditions. Random Conical Tilt (RCT) dataset images were collected at −56° and 0°. Particle picking for all datasets was executed using the Eman2 e2boxer software [46] and a 224 × 224-pixel particle box size window. All datasets were band-pass filtered with a 20-Å low-pass—200-Å high-pass cutoff. Reference Free 2D class averaging of individual complexes was generated using iterative Multivariate Statistical Analysis (MSA) and Multi-Reference Alignment (MRA) in IMAGIC [47] including, on average, ~20 particles per class average. Epitope localization was performed by alignment and cross correlation between reference free 2D classes of Fab bound and unbound samples and by subtracting the unbound class averages from the Fab-bound classes. The obtained difference maps, considered meaningful only if the signal was at least three standard deviations above the mean, indicated the location of each Fab on HCMV complexes. RCT three-dimensional models of the negatively-stained gH/gL/gO and Pentamer complexes were generated by collecting 50 tilt-pair images (0° and 56°) using the same conditions described above. Pair tilts were manually selected for a total of approximately 5000 particle pairs for each sample. The ML2D in the XMIPP package was used to generate reference-free 2D averages from the 0° micrographs [48]. 3D RCT reconstructions were finally generated using SPIDER routines integrated into Appion starting from the more populated RCT classes.
gH/gL/Fab complexes were formed by incubating for 30 min at RT ~300 pmoles of gH/gL-C144S to 1:2 molar excess of 3G16 and 13H11 Fabs. To perform sample labeling, deuterated buffer (Tris-HCl 25mM, NaCl 150 mM, pH 7.1) was added at RT, reaching a deuterium excess of 78%. At different time points (between 30 sec and 30 min), 30 μL of sample were removed and mixed with an equal volume of ice-cold 200 mM sodium phosphate, 4 M guanidinium chloride, 200 mM TCEP, pH 2.1 buffer to quench the deuterium exchange reaction and promote Fab dissociation. Quenched aliquots were flash frozen in liquid nitrogen and stored at -80°C before analysis. Unbound gH/gL-C144S was used as control. Samples were analyzed using a Waters nano-ACQUITY UPLC with HDX Technology coupled to a Waters SynaptG2 mass spectrometer equipped with a standard ESI source (Waters). The data generated with this equipment was analyzed and interpreted using a previously reported method [28]. Only peptides present in at least three repeated digestions were considered for the analysis.
SPR single cycle kinetics experiments were carried out on a Biacore T100 instrument using a human IgG binder kit as previously described [28]. Both ligand and analyte samples were diluted in HBS-EP buffer (GE healthcare BR100669). One channel of a CM5 chip was used as reference while the second was used to capture HCMV neutralizing IgG. Ligand levels were maintained between 45–85 RU. Concentrations of Pentamer or gH/gL/gO from 0, 3.125, 6.25, 12.5 and 25 nM were injected over the two channels for 120 s at 50 μl/min followed by 600 s of dissociation time. The single cycle kinetic curves were fitted using a 1:1 binding stoichiometry for ka, kd, KD.
A capture ELISA assay was performed to determine competing binding of the gH-specific monoclonal antibodies. Individual antibodies were biotinylated using a commercial kit according to manufacturer instructions (NHS-PEG4-Biotin, No-Weigh Format, Thermo Scientific, cat# 21329). 96-well ELISA plates (Immuno F96 MaxiSorp, Nunc cat# 439454) were coated with 100 μl/well of individual monoclonal antibodies diluted to 1 μg/ml in PBS. Following overnight incubation at 4°C, wells were washed 3 times with 300 μl/well of PBS containing 0.05% (w/v) Tween 20 (wash buffer). The wells were blocked with 100 μl/well of 1% (w/v) BSA in PBS (blocking buffer) for 1 h at RT. The blocking buffer was removed by aspiration and purified gH/gL complex, 200 ng/well, was added to the plates in 100 μl/well of PBS containing 1% (w/v) BSA and 0.1% (w/v) Triton X-100 (sample buffer) and incubated for 1 h at RT. Wells were washed 3 times with 300 μl/well of wash buffer and incubated for 1 h at RT with biotinylated detection monoclonal antibodies in sample buffer, 100 μl/well. After washing, HRP conjugated avidin (Vector cat# A-2004), was added at 100 μl/well of sample buffer and incubated for 1 h at RT. Wells were washed and incubated for 30 min with 100 μl/well of TMB substrate (Rockland cat# TMBE-1000). Following incubation, the reaction was stopped by adding 100 μl/well of 2.0 N Sulfuric Acid (BDH cat# BDH3500). The optical density was determined spectrophotometrically at 450 nm wavelength using a microplate reader (Infinite M200 NanoQuant, Tecan).
A capture multiplex assay was performed to determine competing binding of the Pentamer-specific monoclonal antibodies. For direct competition between pairs of monoclonal antibodies Luminex microspheres (MagPlex microspheres, Luminex Corp. cat# MC100XX), of different classification, were coupled with individual monoclonal antibodies by chemical coupling according to manufacturer instructions. Individual antibodies were also biotinylated using a commercial kit (NHS-PEG4-Biotin, No-Weig Format, Thermo Scientific, cat# 21329). In 96 well white plates, 1000 microspheres/well for each classification/monoclonal were mixed in 50 μl/well of DPBS + 1% BSA + 0.05% sodium azide (assay buffer) with purified Pentamer, starting at 100 ng/well with three-fold dilutions down to 0.05 ng/well. After washing three times with 200 μl/well of PBS containing 0.05% (w/v) Tween 20 (wash buffer) to remove excess antigen, individual biotinylated monoclonal antibodies (at 0.75–1.5 μg/ml) were added in separate wells in 50 μl/well of assay buffer and incubated for 1 h at RT with orbital shaking in the dark. After washing, R-Phycoerythrin conjugated Streptavidin (Jackson ImmunoResearch, cat# 016-110-084) was added, 50 μl/well in assay buffer, and incubated for 1 h at RT. After a final wash, Fluorescence intensity was measured using a Luminex FlexMap 3D (Life Technologies model FM3D000).
For evaluation of whether antibodies that did not compete as pairs could bind to Pentamer at the same time, Luminex microspheres were coupled with individual monoclonal antibodies for sites 1 to 5 of the Pentamer (15D8, 10F7, 4N10, 10P3, 2C212). In 96 well white plates, serial dilutions of monoclonal antibody pools containing 1000 ng/well of combinations of four of the five antibodies were made. To these dilutions of antibody pools, 1000 microspheres/well for each classification/monoclonal were added in 50 μl/well of DPBS + 1% BSA + 0.05% sodium azide (assay buffer) together with purified Pentamer, 33 ng/well, and incubated for 1 h at RT with orbital shaking in the dark. After washing three times with 200 μl/well of PBS containing 0.05% Tween 20 (wash buffer) to remove excess antigen, a commercially available biotin conjugated anti His-tag mAb (Rockland, Cat # 200-306-382) for direct detection of the Pentamer, was added with 50 μl/well of assay buffer and incubated for 1 h at RT with orbital shaking in the dark. After washing, R-Phycoerythrin conjugated Streptavidin (Jackson ImmunoResearch, cat# 016-110-084) was added, 50 μl/well in assay buffer, and incubated for 1 h at RT. After a final wash, Fluorescence intensity was measured using a Luminex FlexMap 3D (Life Technologies model FM3D000).
|
10.1371/journal.pgen.1000399 | Destabilizing Protein Polymorphisms in the Genetic Background Direct Phenotypic Expression of Mutant SOD1 Toxicity | Genetic background exerts a strong modulatory effect on the toxicity of aggregation-prone proteins in conformational diseases. In addition to influencing the misfolding and aggregation behavior of the mutant proteins, polymorphisms in putative modifier genes may affect the molecular processes leading to the disease phenotype. Mutations in SOD1 in a subset of familial amyotrophic lateral sclerosis (ALS) cases confer dominant but clinically variable toxicity, thought to be mediated by misfolding and aggregation of mutant SOD1 protein. While the mechanism of toxicity remains unknown, both the nature of the SOD1 mutation and the genetic background in which it is expressed appear important. To address this, we established a Caenorhabditis elegans model to systematically examine the aggregation behavior and genetic interactions of mutant forms of SOD1. Expression of three structurally distinct SOD1 mutants in C. elegans muscle cells resulted in the appearance of heterogeneous populations of aggregates and was associated with only mild cellular dysfunction. However, introduction of destabilizing temperature-sensitive mutations into the genetic background strongly enhanced the toxicity of SOD1 mutants, resulting in exposure of several deleterious phenotypes at permissive conditions in a manner dependent on the specific SOD1 mutation. The nature of the observed phenotype was dependent on the temperature-sensitive mutation present, while its penetrance reflected the specific combination of temperature-sensitive and SOD1 mutations. Thus, the specific toxic phenotypes of conformational disease may not be simply due to misfolding/aggregation toxicity of the causative mutant proteins, but may be defined by their genetic interactions with cellular pathways harboring mildly destabilizing missense alleles.
| Correct folding and stability are essential for protein function. In cells, a network of molecular chaperones and degradative enzymes facilitate folding, prevent aggregation and ensure degradation of the misfolded proteins, thus maintaining protein homeostasis. In many diseases, including Amyotrophic Lateral Sclerosis (ALS), expression of a single mutant protein that misfolds and aggregates causes cellular toxicity that is strongly dependent on the genetic background. To address the influence of genetic background on the toxicity of aggregation-prone proteins, we established a C. elegans model of misfolding and aggregation of several distinct ALS-related mutants of superoxide dismutase 1 (SOD1). In one wild type genetic background (N2), these proteins exhibited only mild cellular toxicity despite strong, mutant-specific aggregation phenotypes. However, when SOD1 mutants were expressed in the background of mildly destabilized protein polymorphisms, their toxicity was enhanced and a number of distinct phenotypes were exposed. These synthetic phenotypes reflected the loss-of-function of the destabilized polymorphic proteins. Furthermore, the degree to which each of these phenotypes was exposed depended on the nature of the SOD1 mutation. These data suggest that the presence of mildly destabilizing polymorphisms in the genetic background may modulate and direct the specific toxic phenotypes in protein aggregation diseases.
| ALS (OMIM #105400 http://www.ncbi.nlm.nih.gov/entrez/dispomim.cgi?cmd=entry&id=105400) is a progressive degenerative disorder affecting motor neurons in the brain stem and spinal cord. Up to 10% of cases have a dominant familial inheritance pattern with mutations in SOD1 (OMIM *14750 http://www.ncbi.nlm.nih.gov/entrez/dispomim.cgi?cmd=entry&id=147450) contributing about 20% of those [1],[2]. While it is accepted that disease results from toxic gain of function by the mutant protein [3]–[5], the mechanisms contributing to toxicity remain unknown. Two main hypotheses have been proposed; the first invokes abnormal chemistry of mutant SOD1 proteins, resulting in nitration of tyrosine residues on cellular proteins [6] and increased production of hydroxyl radicals [7],[8]. However, mutant SOD1 retains its toxic properties even when abnormal chemical reactions are greatly reduced [9] suggesting that abnormal chemistry alone may not be the basis of toxicity. Furthermore, the role of the dismutase genes in preventing the long-term protein damage have recently being questioned [10]. The second hypothesis suggests that, as for many other neurodegenerative diseases, the toxicity is mediated by misfolding and aggregation of mutant proteins [9], [11]–[13]. Accumulation of proteinaceous inclusions in conformational disease indicates an inability of the protein folding quality control machinery to efficiently recognize, fold, and degrade abnormal proteins [14], including the mutant forms of SOD1. The role of damaged proteins is further supported by observations that elevated levels of molecular chaperones decrease mutant SOD1 toxicity [15],[16]. However, it is still unclear how misfolding or aggregation of SOD1 mutant protein leads to cellular toxicity.
ALS patients harboring different or even the same SOD1 mutations exhibit a high degree of clinicopathologic variation, including clinical severity, age at onset, and the types of motor neurons involved [17]–[20]. Both different biophysical properties of mutant proteins and variation in the genetic background may independently modulate the toxicity, providing a range of phenotypes [21],[22]. The importance of genetic interactions in modulating disease is further underscored by findings that ALS phenotypes in transgenic mice vary greatly depending on the strain in which the mutant protein is expressed [23],[24]. Understanding the differences between SOD1 mutants in misfolding/aggregation behavior and in their interactions with cellular proteins and pathways may thus provide insights into the toxic mechanisms and the nature of modifier genes.
To systematically examine the aggregation behavior and genetic interactions of mutant forms of SOD1, we established a C. elegans model expressing human SOD1-YFP fusion proteins in the body-wall muscle cells. The ability to employ dynamic imaging in live animals throughout their lifespan and availability of both forward and reverse genetic approaches makes C. elegans an attractive model to study aggregation toxicity. Similar models in C. elegans have been used to investigate the aggregation toxicity and genetic modifiers of polyglutamine expansions and α-synuclein [25]–[27]. Here, we show that three biophysically distinct [28] mutants of SOD1 form strikingly polymorphic aggregates in C. elegans. Expression of mutant SOD1 alone was associated with mild toxicity. However, when mutant SOD1 was introduced into genetic backgrounds harboring destabilizing temperature-sensitive mutations, the toxicity was enhanced significantly and a variety of toxic phenotypes was observed. These phenotypes reflected both the specific SOD1 mutant and the loss-of-function of each of the destabilized temperature-sensitive proteins. Thus, we propose that specific phenotypes in conformational disease may be influenced by the mildly destabilizing missense mutations present in the genetic background.
We established a C. elegans model to study SOD1 aggregation toxicity by expressing wild type and mutant SOD1 in body wall muscle cells, employing a tissue-specific promoter (pUnc-54) and C-terminal YFP-tagging scheme (Figure S1A) [25]. The YFP-tagged wild type SOD1 retained its enzymatic activity (Figure S2, lane 1), indicating that the tag does not interfere with SOD1 folding. Because various mutations in the SOD1 protein exhibit different biophysical and biochemical properties [28], we chose three distinct mutant SOD1 proteins associated with ALS. G85R is representative of inactive “metal-binding” mutants [29], deficient in copper and zinc binding and significantly destabilized [30]. G93A represents “wild type-like” mutants that bind copper and zinc, exhibit mild loss of thermal stability when fully metallated, and retain enzymatic activity [31]. 127X (G127insTGGGstop) is a frameshift mutation resulting in a C-terminal truncation of the last 21 amino acids and a highly unstable protein [32]. These mutants form protein aggregates with toxic phenotypes when expressed in mammalian cultured cells and transgenic mice [3],[5],[32],[33].
Transgenic lines expressing pUnc-54::WT SOD1::YFP ( WT SOD1), pUnc-54::SOD1-G85R::YFP (G85R), pUnc-54::SOD1-G93A::YFP (G93A) and pUnc-54::SOD1- G127insTGGGstop::YFP (127X) were established. We verified that SOD1 proteins expressed in the muscle cells were of the expected molecular sizes (Figure S1B). Only transgenic lines expressing steady-state levels of mutant proteins similar to or lower than WT SOD1 were selected for further study since high expression levels could influence aggregation and toxicity (Figure S1B).
Wild type SOD1 exhibited diffuse fluorescence in body wall muscle cells throughout development and during adulthood (Figure 1A,I) with broad distribution in the muscle belly (the cytoplasmic space of a muscle cell below the myofilaments) and the muscle arms (the projections from muscle cells toward the neural ring). Although WT SOD1 had patchy appearance in some of the cells, the brighter areas were diffuse upon examination at higher magnification (insert in Figure 1I), corresponding to soluble protein (Figure 2A). In contrast, all three mutant SOD1 proteins presented a punctate fluorescent pattern that appeared in embryonic stages (Figure 1B–D) and persisted throughout larval development and adulthood (Figure 1 and Figure S3). In all three mutant forms of SOD1, we observed both diffuse and punctate fluorescence corresponding to two populations of protein. Thus, SOD1 proteins in C. elegans exhibit properties similar to those observed in other model systems, where only the mutant SOD1 protein forms inclusions when expressed ectopically [3],[5],[32],[33].
Accumulation of mutant SOD1 proteins into visible foci, although consistent with their in vitro aggregation propensity, does not necessarily indicate the formation of aggregates. We used dynamic imaging and FRAP analysis to establish the aggregation state of SOD1 proteins in live animals. As shown in Figure 2A, the diffusion of WT SOD1-YFP fusion protein in body wall muscle cells (light blue) was indistinguishable from that of YFP alone (purple), with nearly complete recovery within the dead-time of measurement post bleaching. In contrast, the fluorescent foci of all three mutant proteins exhibited reduced recovery indicative of immobile aggregate species (Figure 2A,B). G85R and G93A proteins had 35 and 50% recovery over 275 seconds, respectively, higher than that observed for foci of well-characterized aggregation-prone polyQ40 (30%), which contain only an immobile protein [34]. The recovery of fluorescence in 127X foci continued beyond 100 sec and reached nearly 60% over the course of the experiment, which suggests either partially mobile species, or the presence of multiple populations of protein.
Since these SOD1 mutants have different structural and biophysical properties in vitro [28], the observed differences in the fluorescence recovery of aggregates could reflect the presence of different molecular species or interactions in vivo. We analyzed the oligomeric state of SOD1 proteins by native gel electrophoresis. Extracts from G85R, G93A and 127X lines contained soluble SOD1 protein in addition to large aggregate species that did not enter the gel, while extracts of WT SOD1 lines contained mainly soluble protein (Figure 2C). The distribution of the soluble G93A protein appeared similar to the WT SOD1, with one major band containing enzymatically active protein (Figures 2C, star and S2, arrow). In contrast, G85R and 127X were resolved as multiple species of different intensities lacking enzymatic activity (Figures 2C and S2B). This is in agreement with the known native-like properties of G93A [31] and suggests that human SOD1 proteins can preserve their characteristics when expressed in C. elegans. 127X extracts also contained a heterogeneous population of electrophoretic states (Figure 2C, bracket), which could indicate conformational instability and continual dissociation of larger molecular species, in agreement with the continual recovery of fluorescence by FRAP assay (Figure 2A, dark blue).
We further characterized the SOD1 aggregates using detergent solubility. The large molecular weight material was resistant to non-ionic detergent (0.5% Triton X-100), but was readily dissociated by 5% SDS at room temperature (Figure 2D).
Thus, mutant SOD1 in C. elegans appears to form a molecularly heterogeneous mixture of SDS-labile aggregates and soluble protein, which for G85R and 127X does not attain a stably folded, native conformation.
The observed biochemical heterogeneity paralleled a striking heterogeneity in aggregate morphology and distribution in SOD1-transgenic strains. While all three mutant SOD1 strains had some cells devoid of visible aggregates, most contained aggregates that exhibited a wide range of shapes, sizes, and cellular distribution (Figure 3). The majority of cells (up to 75%) in G85R animals contained 1–5 aggregates with the apparent size of the fluorescent foci of 5–7 µm, while some cells contained more than 20 smaller (less than 1 µm) dispersed aggregates (Figure 3B). Both types of aggregates exhibited similar biophysical properties by FRAP analysis (not shown). The G93A strain had a more uniform distribution of morphological types (Figure 3C), with aggregates often seen in close apposition to each other (Figure S4A,B). 127X animals differed from both G85R and G93A strains in that they contained up to 40% of cells with irregular, non-spherical aggregates (Figure 3D). The presence of distinct morphological classes did not depend on the expression level, as we observed a similar distribution in heterozygous SOD1 animals, despite lower extent of aggregation (not shown). These data show that the wild type and three different mutants of SOD1 form morphologically, structurally and enzymatically different molecular species in vivo, supporting the possibility of distinct interactions with cellular components.
We next asked whether expression of these aggregation-prone proteins caused toxicity. We assessed several phenotypes as indicators of dysfunction of muscle cells expressing the transgenes, such as decrease in motility of animals, disturbance of ultrastructural organization of myofilaments, developmental defects, and egg-laying defects. The motility of WT SOD1 animals grown at 15°C was similar to that of wild type (N2) strain on the second day of adulthood, as measured by number of body bends per minute (Figure 4A). In contrast, animals expressing G85R, G93A or 127X SOD1 mutant proteins had 25–30% reduction in motility relative to that of N2 animals. This decrease in motility was only minimally enhanced by the sixth day of adulthood (Figure 4A, light grey bars).
To assess the ultrastructural organization of myofilaments, we visualized actin filaments with Rhodamine-labeled phalloidin. We found no major disruptions of the organization of actin filaments in cells containing aggregates in either of three mutant strains (Figures 4B, S4A). In fact, SOD1 aggregates were localized to a different focal plane than the myofilaments (Figure S4B), unlike polyQ aggregates, which were found intercalating into filaments and disrupting their structure (Figure S4AVII and AVIII).
Dysfunction of muscle cells during embryonic development leads to defective elongation of the body shape of C. elegans and to embryonic lethality (emb) or hatching of deformed, growth arrested larvae (lva). Although SOD1 mutant proteins aggregate already in embryos, neither of the mutant SOD1 strains exhibited substantially elevated emb+lva phenotype relative to the WT SOD1 strain (Table 1). The highest toxicity was in G85R strain (8.2% phenotype, compared to 1.8% in WT SOD1). Likewise, despite the presence of aggregates in vulva muscle cells, no increase in egg laying defect was found in mutant SOD1 strains (Table 1) compared with WT SOD1 strain. Thus, under given experimental conditions, expression of mutant SOD1 proteins seems to exert limited toxicity in the muscle cells of C. elegans.
The relatively mild toxicity, despite misfolding and aggregation of mutant SOD1, indicates that the putative toxic species are either transient or suppressed by the cellular folding/quality control machinery. We had previously found that metastable temperature-sensitive (ts) mutations in various unrelated genes, such as unc-15, unc-45 and let-60, coding for paramyosin(ts), UNC-45(ts) and Ras(ts) proteins, respectively, destabilized the cellular folding environment and modulated the toxicity of polyQ expansions [35]. To examine whether protein polymorphisms in genetic background could influence SOD1 toxicity, we introduced WT and mutant SOD1 proteins into ts mutant strains and assayed ts phenotypes at the permissive temperature. WT SOD1 showed no synthetic toxicity with ts mutants. In contrast, G85R, G93A and 127X mutants of SOD1 caused exposure of each specific ts phenotypes at permissive temperature (Table 1).
Loss of function of paramyosin during embryonic development leads to defects in muscle structure and thus to the emb+lva phenotype [36]. Expression of mutant SOD1 in the paramyosin(ts) strain caused differential exposure of these phenotypes at the permissive temperature (Table 1), depending on the identity of the SOD1 mutant: expression of G85R resulted in 55% of emb+lva at 15°C, 127X had 33% of emb+lva phenotype, and G93A had intermediate toxicity. The surviving animals had very few progeny. To ask whether this toxic interaction is specific to paramyosin, we crossed SOD1 strains to a strain harboring a ts mutation in unc-45 gene. Expression of SOD1 mutant proteins in unc-45(ts) genetic background resulted in exposure of egg laying and reproductive phenotype (egl+rep) at the permissive temperature (Table 1). This phenotype is characteristic of dysfunction of UNC-45-expressing embryonic muscle cells, vulva muscle cells and gonad sheath cells, and is present in a 100% of unc-45(ts) mutant animals at the restrictive temperature. Here, G93A exhibited over 80% toxicity at 15°C, compared with less than 5% for either G93A or UNC-45(ts) expressed alone. Furthermore, the surviving animals expressing both G93A and UNC-45(ts) developed into severely uncoordinated (Figure S5), sick adults.
Paramyosin and UNC-45 both affect the formation of myofilaments: paramyosin is a structural component and UNC-45 regulates myosin assembly. To assess whether SOD1 mutants were toxic towards a metastable mutant in a different cellular pathway, we crossed SOD1 strains to a strain expressing a temperature-sensitive Ras variant. Expression of G85R and G93A in ras(ts) background did not have strong effects on embryonic lethality (Table 1), while 127X, which was the least toxic with paramyosin(ts) and UNC-45(ts), caused lethality in 23% of embryos. SOD1 mutant proteins did cause Ras(ts) animals to exhibit defects in egg laying (Table 1), with 72% egl+rep phenotype exposed in 127X and 14% in G85R animals. The ts strain by itself, however, did not present the same egg laying defect (egl) at 25°C. Instead, Ras(ts) animals raised at the restrictive temperature had a fluid-filled appearance with degenerated gonads, and produced few embryos (100% reproductive phenotype, rep). Note that Table 1 shows a combined egl+rep phenotype. Ras is ubiquitously expressed, and pleiotropic phenotypes that are exposed at the restrictive temperature reflect its dysfunction in different cell types. The egl phenotype exposed by SOD1mutants in ras(ts) background is likely due to the genetic interaction between the two mutations specifically in the in vulva muscles or gonad sheath cells, where the Unc-54 promoter driving SOD1 expression is active.
These data show that while expression of the three distinct SOD1 mutants in the wild-type N2 strain of C. elegans leads to mild toxicity, their expression in the genetic backgrounds harboring diverse temperature-sensitive mutations uncovers toxic phenotypes, with each SOD1 mutant affecting the activity of a given metastable protein to different extents.
We show here that introduction of mildly destabilized protein polymorphisms into the defined genetic background of C. elegans modulates the toxicity of three mutant SOD1 proteins, leading to the development of specific toxic phenotypes. While either SOD1 aggregation or loss-of-function of metastable ts proteins can each be viewed as a separate consequence of failure of protein folding homeostasis, the toxic phenotypes observed here resulted from their genetic interaction and thus were directed by the nature of the ts mutation present. Similar to ts mutations in C. elegans, the phenotypic expression of mildly destabilizing protein polymorphisms in higher organisms is thought to depend on the robustness of the protein folding environment [37]. Thus, the demand on the folding resources as a consequence of aging, proteotoxic conditions, and genetic background, may alter the threshold for the toxicity of an aggregation-prone protein, while specific pathways and protein complexes containing such polymorphisms may direct cell-type specific phenotypes.
We established a C. elegans model in which aggregation, toxicity, and cellular interactions can be directly compared between different SOD1 mutants. Furthermore, as models of other aggregation-prone proteins, such as polyglutamine expansions [25], α-synuclein [27], and Aβ [38] use similar expression schemes in the same N2 genetic background, these C. elegans models could be instrumental in deciphering both common and protein-specific regulation of aggregation or toxicity. Expression of three different ALS-related mutants of SOD1 in body-wall muscle cells of C. elegans lead to mild cellular disfunction and appearance of protein aggregates with distinct morphological characteristics. We also observed an unexpected variability of aggregate morphology in neighboring cells of the same animal, which could indicate that factors other than genetically encoded interactions also affect the fate of SOD1 in the cell. Similar stochastic variability between muscle cells of the same animal was previously reported with respect to onset of sarcopenia in wild type C. elegans [39].
We find that ALS-related SOD1 variants exert a potent destabilizing influence on the functionality of metastable temperature-sensitive proteins at permissive conditions, exposing a range of phenotypes that are not present in strains expressing SOD1 mutants alone. Moreover, the most toxic of the SOD1 mutants (127X) in the Ras(ts) background was the least toxic in the paramyosin(ts) background, whereas the G85R mutation was most toxic with paramyosin(ts). Strain-dependent differences in SOD1 toxicity were previously observed in a mouse model carrying G86R mutation in murine SOD1 (corresponding to G85R in human SOD1), with complete suppression of toxicity up to 2.5 years in one genetic background, but a rapid onset of paralysis by 90–120 days in a different genetic background [23]. Our data suggests that mildly destabilizing missense mutations, present in the genetic background, could effect the exposure of specific phenotypes.
The nature of the toxicity of aggregation-prone proteins remains one of the central questions for diseases of protein conformation. We have previously showed that unrelated ts mutations caused premature aggregation of polyQ-expanded proteins [35]. Furthermore, metastable proteins encoded by these ts mutations were found to misfold and lose function in polyQ strains, indicating that protein folding homeostasis was disrupted by chronic protein misfolding. Unlike polyQ, mutant SOD1 proteins, though highly aggregation-prone, exhibited much lower toxicity on their own. The demonstration that toxicity of both mutant SOD1 and polyQ expansions can be modulated by metastable proteins supports our contention that the proteostasis network [40] is sensitive to cumulative protein damage, and that the disruption of protein folding may be a common mechanism that underlies the toxicity of different aggregation-prone proteins.
Each of the SOD1 mutant proteins used in this study exhibits distinct biophysical properties in vitro [28], and forms morphologically, structurally and enzymatically distinct molecular species and aggregates in C. elegans. It is thus possible that SOD1 mutant proteins form different intermediate folding states in vivo depending on the nature of the mutation, and as such may possess different functional interactions with the folding machinery of the cell. Indeed, G85R and G93A proteins were recently shown to have different interactions with HSP70 in cultured cells [41]. On the other hand, the structure and functions of paramyosin, UNC-45, and Ras are diverse (a structural coiled-coil protein, a soluble TPR domain-containing protein and a small GTPase, respectively) and these proteins are not overexpressed as they are expressed from their endogenous chromosomal loci. Thus, it is unlikely that the synergistic effects on toxicity are because of direct and specific molecular interactions between these protein polymorphisms and mutant SOD1. This is in agreement with our previous observation that polyQs cause misfolding of metastable proteins in the absence of direct molecular interactions [35], and with a recent report that many of the modifiers of toxicity of polyQ-expanded ataxin-3 in Drosophila also rescue the generic toxicity of protein misfolding due to the reduced function of HSP70 [42]. Furthermore, both the functionality of metastable proteins and polyglutamine aggregation can be compromised by neuronally-mediated overexcitation of the muscle cells in C. elegans [43]. These findings parallel recent computational evidence that the selection against the toxicity of misfolding due to mistranslation exerts strong evolutionary pressure specifically on the highly expressed proteins [44], indicating that the flux of destabilized proteins in a cell bears a significant fitness cost, and that folding resources are indeed limiting. In support of this, we show that overexpression of the heat-shock transcription factor HSF-1 rescues the toxic phenotypes in a strain co-expressing an SOD1 mutant G93A and a metastable ts mutant of UNC-45 (Figure S5B). Thus, we propose that the genetic interactions between disease-causing mutations and mildly destabilizing protein polymorphisms are mediated at the cellular level by competition of their respective gene products for folding resources.
This hypothesis could offer an explanation for the apparent paradox of cell-type-specific toxicity caused by ubiquitously expressed toxic proteins in conformational diseases. Indeed, in SOD1-related ALS, Huntington's disease, and Alzheimer's disease, specific neuronal subtypes are affected despite ubiquitous expression of SOD1, huntingtin and APP, respectively. The differential modulation of mutant SOD1 toxicity in C. elegans by specific ts mutations suggest that the presence of mildly destabilizing protein polymorphisms in the genetically diverse human population could direct such specific phenotypes: because each cell type contains characteristic complement of expressed proteins, the genetic interactions of aggregation-prone proteins with destabilizing polymorphisms are expected to manifest in a cell type-specific manner. The disease variability across the population suggests that such protein polymorphisms may be specific to individuals or families, and missed in the population-based linkage analyses. A recent study found that up to 70% of rare missense alleles are mildly deleterious in humans [45]; some of these polymorphisms may result in the production of metastable or folding-deficient proteins [46],[47]. Identification of cell-specific pathways or protein complexes, which may disfunction when folding or stability of their components is challenged by co-expression with an aggregation-prone protein, may thus provide specific toxic mechanisms for conformational diseases and help focus the search for disease-modifying polymorphisms.
Nematodes were grown on NGM plates seeded with E. coli OP50 strain. Animals were synchronized by picking L4 larva or pre-comma stage embryos onto fresh plates. Assays were performed with young adult animals, at the second day of reproductive adulthood at either 15°C (3.5 days after L4 stage) or 25°C (2 day after L4 stage). C. elegans strains were obtained from the Caenorhabditis Genetic Center. Ts mutants were: paramyosin(ts) - CB1402[unc-159(e1402)], UNC-45(ts) - CB286[unc-45(e286)] and Ras(ts) - SD551[let-60(ga89)].
The SOD1 transgenic strains were created by injection and integration of complex arrays, allowing for uniform expression of transgenes. Human SOD1 sequences were obtained by PCR amplification from plQL01 or plQL03 (gift from Dr. Q. Liu, Harvard Medical School), and cloned into a Fire Lab pPD30.38 plasmid. DNA mixture for injection contained 1 ng of linearized plasmid DNA and 100 ng of worm genomic DNA digested with PvuII (NEB).
Crosses between SOD1 transgenic and ts strains were performed by first mating N2 males with SOD1 hermaphrodites, and subsequently mating SOD1 heterozygous fluorescent males with ts hermaphrodites. 3–5 fluorescent F1 hermaphrodite progeny from these crosses were allowed to self, and 15–20 F2 fluorescent progeny were singled onto individual plates. Plates containing 100% temperature-sensitive progeny were used for generation of double-homozygous strains. To generate a strain double homozygous for G85R and Ras(ts), a singe fluorescent F2 hermaphrodite exhibiting a strong muv phenotype was picked directly from a pool of F2 progeny. We could not generate a WT SOD+Ras(ts) animals, presumably due to close genetic location of respective loci.
We noted that strains co-expressing SOD1 and ts mutant proteins tended to accumulate suppressors, similar to what was observed with polyQ expansions [35]. Strains with low progeny number were particularly unstable, resulting in segregation of strains revertant for the toxic phenotype either during selection of homozygotes or in the first few generations thereafter, or in gradual improvement of fitness of the entire population (data from such strains were discarded). All double homozygous lines were periodically re-built, and assays were performed within two weeks of obtaining double homozygous animals.
Fluorescence recovery after photobleaching (FRAP) analysis was performed as previously described [48]. Imaging was performed on a Zeiss LSM 510 Meta confocal microscope. YFP alone (polyQ0) and polyQ40 strains, used as controls for soluble and stably aggregated protein species, are described in [34].
For native extracts, nematode pellets were mechanically disrupted, lysed in native lysis buffer (50 mM Tris pH7.4, 5 mM MgCl2, 0.5% Triton-X 100, 0.2 mM PMSF, 1 ug/ml Leupeptin, 1 ug/ml Pepstatin A, Complete protease inhibitor (Roche)) and centrifuged for 1 min at 30×g (Eppendorf 5417C centrifuge). All reagents were from Sigma, unless indicated otherwise. This protocol is optimized for the removal of debris and large fragments of cuticle while preserving the majority of aggregates in the supernatant, verified by examination of supernatant under fluorescent microscope. For detergent solubility, native extracts were incubated in the indicated detergent for 15 min at room temperature prior to resolving by native PAGE. 20 or 30 micrograms total protein was run on a 5% or 7.5% continuous native gels. Gels were imaged on Storm 860 scanner (Molecular Dynamics) with ImageQuant software to detect YFP fluorescence, or processed for the in-gel enzymatic assay.
For epifluorescence, nematodes were mounted on 1% agarose pads with 1 mM levamisole and imaged using Zeiss Axiovert 200 microscope. For immunofluorescence and confocal imaging, synchronized adults were fixed, permeabilized and stained with Rhodamine-phalloidin (Molecular Probes), as described previously [35], and imaged with Zeiss LSM 510 Meta confocal microscope through a 40×1.0 numerical aperture objective with either a 514-nm or 543-nm line for excitation.
To measure motility, nematodes at indicated age were placed individually in a drop of M9 buffer and acclimated for 1 min; the completed body bends were counted for 1 min. At least15 animals were used per experiment. Similar decrease in motility was found by this method and by measuring rate of movement of animals raised at 20°C on a plate seeded with OP50 bacteria (not shown).
For emb+lva at 15°C, freshly laid pre-comma stage embryos were picked onto new plates. Unhatched embryos and larvae that hatched but did not crawl or were severely deformed were scored after 2 days. Alternatively, young adults were acclimated to 25°C for 1 day prior to egg laying, transferred onto new plates and allowed to lay embryos. Embryos were picked and scored one day later at 25°C. About one hundred embryos was used per experiment and experiments were repeated at least three times.
To score egl+rep and severe uncoordination, 30 L4 larvae grown at 15°C were picked to a fresh plate and incubated for 3 days at 15°C or 1.5 days at 25°C. Animals retaining eggs or containing three-fold embryos (detected with Nomarski optics) were scored as egg laying defective (egl). Animals with degenerated gonads, sterile, and those accumulating oocytes were scored as having a reproductive defect (rep). Animals that did not move on their own or did not exhibit sinusoidal movement pattern after being prodded were scored as severely uncoordinated. Experiments were repeated at least three times.
|
10.1371/journal.ppat.1004281 | Human APOBEC3 Induced Mutation of Human Immunodeficiency Virus Type-1 Contributes to Adaptation and Evolution in Natural Infection | Human APOBEC3 proteins are cytidine deaminases that contribute broadly to innate immunity through the control of exogenous retrovirus replication and endogenous retroelement retrotransposition. As an intrinsic antiretroviral defense mechanism, APOBEC3 proteins induce extensive guanosine-to-adenosine (G-to-A) mutagenesis and inhibit synthesis of nascent human immunodeficiency virus-type 1 (HIV-1) cDNA. Human APOBEC3 proteins have additionally been proposed to induce infrequent, potentially non-lethal G-to-A mutations that make subtle contributions to sequence diversification of the viral genome and adaptation though acquisition of beneficial mutations. Using single-cycle HIV-1 infections in culture and highly parallel DNA sequencing, we defined trinucleotide contexts of the edited sites for APOBEC3D, APOBEC3F, APOBEC3G, and APOBEC3H. We then compared these APOBEC3 editing contexts with the patterns of G-to-A mutations in HIV-1 DNA in cells obtained sequentially from ten patients with primary HIV-1 infection. Viral substitutions were highest in the preferred trinucleotide contexts of the edited sites for the APOBEC3 deaminases. Consistent with the effects of immune selection, amino acid changes accumulated at the APOBEC3 editing contexts located within human leukocyte antigen (HLA)-appropriate epitopes that are known or predicted to enable peptide binding. Thus, APOBEC3 activity may induce mutations that influence the genetic diversity and adaptation of the HIV-1 population in natural infection.
| Cytidine deaminases of the human APOBEC3 gene family act as an intrinsic defense mechanism against infection with HIV-1 and other viruses. The APOBEC3 proteins introduce mutations into the viral genome by inducing enzymatic modification of nucleotide sequences and inhibiting synthesis of cDNA strands from the viral RNA. Viral Vif counters this impediment to the fidelity of HIV-1 replication by targeting the APOBEC3 proteins for degradation. Low-level APOBEC3 activity that outlasts blockade by viral Vif may foster infrequent mutations that provide a source of genetic variation upon which natural selection acts. Here, we defined the APOBEC3 nucleotide contexts of the edited sites by titration of the wild type and non-editing APOBEC3 mutant in cultured cells. We then followed the patterns of G-to-A mutations we identified in viral DNA in cells obtained from ten patients with acute infection. Our deep sequencing analyses demonstrate an association between sub-lethal APOBEC3 editing and HIV-1 diversification. Mutations at APOBEC3 editing contexts that occurred at particular positions within specific known or predicted epitopes could disrupt peptide binding critical for immune control. Our findings reveal a role for human APOBEC3 in HIV-1 sequence diversification that may influence fitness and evolution of beneficial variants and phenotypes in the population.
| The pathogenesis of HIV-1 infection correlates with the level of active viral replication and relates to a variety of factors specific to the virus, the host, and its immune system. Mutations, insertions, deletions and recombinations that confer changes in the activity of virally encoded genes and gene products affect virus entry and post-entry events [1]. HIV-1 shows high mutational frequency because of a combination of rapid rates of viral replication, error-prone viral reverse transcriptase (RT) and RNA polymerase II replicating enzymes, and recombination during concurrent infection with two or more distinct genomic RNA strands [2]–[4]. In addition, nascent HIV-1 cDNA is vulnerable to mutation by host cell single-stranded cytidine deaminases that edit cytidine to uridine in the minus strand DNA copied from the viral RNA genome, giving rise to G-to-A mutation of the plus strand of viral DNA with a graded frequency of deamination from the primer binding site to the central polypurine tract and the central polypurine tract to the 3′ polypurine tract regions [5]–[8]. The effect of the processes of mutation and recombination in HIV-1 is to promote genetic diversity among the viral variants and thereby allow for a faster rate of adaptation.
Seven related human apolipoprotein B mRNA-editing enzyme, catalytic polypeptide-like 3 (APOBEC3) cytidine deaminases—namely, APOBEC3A, APOBEC3B, APOBEC3C, APOBEC3D, APOBEC3F, APOBEC3G, and APOBEC3H—reside in an expanded 130-kb gene cluster that likely arose through segmental duplication on chromosome 22q13.1 with additional modification by repeated episodes of positive selection during primate evolution [9]. Cytidine deaminases of the APOBEC3 gene family have specificity for single-stranded DNA and inhibit infection by a diverse array of RNA and DNA viruses and retrotransposons by interfering with viral genome replication and littering the genome with deleterious mutations [10], [11]. Mutations mediated by APOBEC3 molecules have a strong preference for a 5′-GG-3′ and 5′-GA-3′ dinucleotide context of the edited sites (target nucleotide underlined) [1], [12], [13].
APOBEC3D, APOBEC3F, APOBEC3G and APOBEC3H are the cellular targets for the HIV-1 accessory protein Vif [14], [15], which can counteract the protective role of this innate immune defense mechanism. The HIV-1 protein Vif induces polyubiquitylation through simultaneously binding to APOBEC3 proteins and the cullin5-elongin B/C-Rbx2 ubiquitin ligase complex. In this manner, the APOBEC3 protein serves as an adaptor that recruits the ligase complex to its substrate and induces the subsequent proteasomal degradation of APOBEC3 proteins. This depletes the pool of APOBEC3 proteins available for incorporation into the assembling viral particle, and thereby minimizes their ability to restrict HIV-1 replication [16]–[19].
Mutation of the HIV-1 genome by cytidine deamination could have a dramatic effect on viral replication. A high rate of mutation could prevent the formation of functional proviruses and explain the G-to-A hypermutation observed in patient samples [20], [21]. Low levels of APOBEC3 activity that survive inhibition by the HIV-1 protein Vif may expose the virus to a broad spectrum of mutations that, rather than impeding virus replication, could provide a source of genetic variation upon which natural selection acts [22], [23]. Indeed, extensive sequencing of transmitted founder HIV-1 variants indicates that sequence variation bearing the hallmark of APOBEC3-mediated G-to-A mutation is commonplace and experiments employing low levels of APOBEC3G expression confirm that modest mutation frequencies, as opposed to inactivating G-to-A hypermutation (5′-UGG-3′ to 5′-UAG-3′; tryptophan-to-stop codon), can be recapitulated in cell culture experiments [21], [24]–[26]. Such sequence changes would have the potential to underlie advantageous alterations in HIV-1 phenotype, such as the appearance of mutations in HLA class I-restricted epitopes that can confer escape from immune recognition or the acquisition of drug resistance [1], [21], [22], [27].
Here, we characterized the relationship(s) between APOBEC3 editing of HIV-1 and the perhaps subtle contribution of mutations that could influence viral adaptation and evolution, as contrasted with destructive G-to-A hypermutation. We applied high depth sequencing to infected cells from single-cycle APOBEC3 titration transfection experiments to confirm and extend the definitions of the DNA sequence context of the edited sites for the four cytidine deaminases of the APOBEC3 gene family that are the cellular targets for HIV-1 protein Vif and their site-specific editing frequencies [5], [7], [8], [12], [13], [28]–[30]. We then followed the evolution of sequence changes and the appearance of the G-to-A signature mutations in the consensus trinucleotide contexts for APOBEC3 protein edited sites in the Gag and Vif genes of HIV-1 in proviral DNA sampled from the peripheral blood in 10 patients with primary HIV-1 infection through time. We found a higher frequency of substitutions within an APOBEC3 trinucleotide context of the edited sites in patients that could often resulted in sequence changes within some major histocompatibility complex (HLA in humans) restricted epitopes that can confer immune escape. Thus, we provide evidence that sub-lethal levels of APOBEC3 deaminases may expose the viral genome to beneficial mutations that influence HIV-1 adaptation and evolution in natural infection.
To make predictions about the potential for APOBEC3 editing of the HIV-1 genome to influence virus diversification in natural infection, it was first necessary to carefully define the nucleotide sequence editing context preferences for the APOBEC3 proteins by titration in virus producing cells (Table 1). Vesicular stomatitis virus G (VSV-G) pseudotyped HIV-1 vif-deficient (HIV-1 pIIIB/Δvif) stocks produced in the presence of escalating doses of each of four human APOBEC3 genes that are the cellular targets for the HIV-1 protein Vif —namely, APOBEC3D, APOBEC3F, APOBEC3G and APOBEC3H—were therefore used to challenge cultured 293T cells.
We screened for mutations in HIV-1 nascent retroviral cDNA with high throughput 454 pyrosequencing using a statistical framework to improve measurement accuracy. Barcoded oligonucleotide primers with unique molecular identifiers were used to amplify a PCR pool that was then sequenced with sufficient depth of coverage to redundantly cover the chosen portion of the viral genome. By stringent filtering and correcting the raw sequencing data, the platform efficiently reduces most sequencing errors generated during pyrosequencing [31]. We separated the resulting viral sequences by sample using the index sequence. Sequence alignments made to the HIV-1 pIIIB/Δvif reference sequence were optimized to reduce alignment errors introduced by insertions-deletions (indels) associated with the pyrosequencing chemistry, correct for incomplete extension miscalls, and filter out less abundant sequencing reads. We used a statistical analytical framework to filter the data for error correction and then built the haplotypes present in the viral populations. With this approach, we found a significantly lower average mutation frequency than the typical analysis of these data (average of 8.45×10−4) (Figure 1A), consistent with the estimations of others [2]. The nucleotide substitution rate for each mutation type (transition or transversion) differed by 1.2 orders of magnitude (Figure 1B). The different rates of nucleotide substitution likely reflect viral RT and RNA polymerase II fidelity and the asymmetrical substitution bias for faster accumulation of G-to-A mutations, the expected result of APOBEC3 editing. Random PCR amplification bias did not affect the reliability of the measurements [32].
Using these results, we identified G-to-A mutations in plus strand DNA as a genetic signature to identify a posteriori APOBEC3 nucleotide contexts of the edited sites. Each G-to-A mutation was considered independently (Figure 1C). Because there is about a two-fold increase in the frequency of adenosine relative to guanosine in the viral genome, we corrected for a bias in the 5′-GpA-3′ and 5′-GpG-3′ context raw numbers. For each of the four human APOBEC3 proteins that are the cellular targets for the HIV-1 protein Vif, we identified positions in the Gag gene region of HIV-1 we sequenced where the site-specific G-to-A substitution frequency increased significantly with increasing APOBEC3 protein abundance (Spearman P<0.05; Figure 2A).
The four APOBEC3 proteins had a clear bias for the plus strand 5′-GpG-3′ or 5′-GpA-3′ in the trinucleotide context of edited sites that can serve as signatures for specific APOBEC3 gene activity (Figure 1C). APOBEC3G exhibited the highest frequency of G-to-A mutations in a 5′-GGD-3′ (where D is the IUPAC code for G, A, or T) and 5′-GAG-3′ context of edited sites and the most hypermutated sequences (Figure 2A). APOBEC3F and APOBEC3H proteins showed high frequencies of G-to-A mutation (range, 0.08 to 0.15) in both 5′-GAD-3′ and 5′-GGA-3′ contexts of edited sites. APOBEC3D showed a high frequency of a 5′-GGD-3′ context of edited sites, but the least activity of the four APOBEC3 deaminases tested under these experimental conditions.
For the trinucleotide context of edited sites within the Gag region of HIV-1 sequenced, G-to-A mutation did not invariably happen in all the potential APOBEC3 trinucleotide context of the edited sites. This observation suggests that other factors inherent to the sequence may affect the activity of these cytidine deaminases. The tryptophan (5′-UGG-3′) to a stop (5′-UAG-3′ or 5′-UAA-3′) codon change, for example, occurred at these two positions at a different frequency in the four APOBEC3 deaminases (Figure S2). Further, the 5′-GCC-3′, 5′-GCG-3′, and 5′-GTC-3′ trinucleotide contexts were not noticeably affected by any of the four APOBEC3 proteins. We did not find a statistically significant enrichment for the rare cytosine-to-thymidine (C-to-T) mutations brought about by a conflict between APOBEC3 editing and guanosine∶uridine (G∶U) mismatched base pair repair in regions of the genome where the plus-strand may become briefly single-stranded during reverse transcription [8].
In the absence of human APOBEC3 protein, the most commonly recovered mutations were random G-to-A or C-to-T transition mutations and −1 frameshift mutations from RT errors. The mutation rate during a single-cycle of HIV-1 replication was approximately 8.45×10−4 mutations per nucleotide (95% confidence interval [7.5, 9.6]×10−4 per nucleotide substitution). In the presence of increasing levels of human APOBEC3 protein, the mutation rate during a single-cycle of HIV-1 replication rose by a factor of between 2 (for APOBEC3D) and 20 (for APOBEC3G), which at sub-lethal levels would likely increase virus diversification and allow HIV-1 to evolve at different rates (Figure 1A) [33]. Once the guanosine in the trinucleotide context of edited sites was removed to correct for the substitution rate in the viral sequence, the average mutation rate corresponded to measurements for single-cycle of HIV-1 replication in the absence of functional APOBEC3 (Figure 2B). Thus, human APOBEC3 deaminase activity is evident in both a directional substitution bias and a higher substitution rate.
To explore the effects of APOBEC3 editing of HIV-1 genomes on virus diversification in patients, we produced longitudinal pyrosequencing data from the Gag and Vif genes of HIV-1 proviral DNA in peripheral blood mononuclear cells (PBMCs) isolated from ten patients during acute or early HIV-1 infection and at a second time point approximately 24 to 26 weeks later before the start of combination antiretroviral therapy. Table S1 shows the clinical characteristics of the ten patients. Transmission risk factors were not able to be determined in two patients (S002 and S006). The range of the estimated duration of infection was between 11 and 70 days [34]. Figure S2 shows the temporal changes in the mean levels of HIV-1 RNA in plasma (range, 779 to 57.6×106 copies per ml) and CD4+ T-cell number counts (range, 311 to 760 cells per ml3). Table S2 shows the high-resolution HLA genotypes. A genetic screen of the APOBEC3D, APOBEC3F, and APOBEC3G genes found no single nucleotide polymorphisms associated with loss-of-function. The poorly expressed APOBEC3H hap I [35], [36] was found for seven patients, four of whom were homozygotes (S001, S005, S007, and S010) and three of whom were heterozygotes (S004, S006, S010; Table S3) [35]–[39].
To look for changes in the genetic structure of the HIV-1 population in each patient through time, we performed high depth gene sequencing with the aforementioned performance improvements for haplotype reconstruction to achieve the sensitivity and molecular resolution necessary for distinguishing among individual viral variants [31]. DNA isolated from patient PBMCs was subjected to HIV-1 Gag (nucleotide positions 977–1564 corresponding to HXB2) and Vif (nucleotide positions 5041–5619) gene sequencing using the 454 Life Sciences' GS-FLX pyrosequencing system. The median number of viral DNA template copies was 23,902 (interquartile range [IQR], 8,909 to 34,103) as measured by quantitative polymerase chain reaction (qPCR).
To estimate the virus population structure of the sample from the pyrosequencing reads, we aligned the reads to a consensus sequence and collapsed the repeated reads to build the haplotypes present in the viral population. We assembled the haplotypes in accordance with the informative de-noised reads of prescribed length such that the fewest haplotypes can account for the most reads. We then estimated the frequency of the reconstructed haplotypes present in the population. After error correction of the sequencing reads and reconstruction of the explaining haplotypes, the median number of the minimal inferred candidate haplotypes present in the population was 30.5 for the Gag (IQR, 5.75 to 67.0) and 35.65 for the Vif (IQR, 9.50 to 59.00) genes of HIV-1 (Table S4). The number of input molecules exceeded the fold depth of sequence data attained. This criterion is necessary to avoid bias caused by selective amplification and artifacts due to sampling and technical variability caused by pyrosequencing [40]. Binomial power calculations suggest that a sample size of 25,000 sequences gives a 96% likelihood of a variant present at 0.02% of the virus population to occur at least twice in the sample [24]. The depth of sequencing reads (range, 28,000 to 84,500) is sufficient to detect viral variants present at or above 0.02% of the population (Table S4).
Nucleotide sequences corresponding to the Gag and Vif genes of HIV-1 sampled at the early and late time points during infection were subjected to maximum-likelihood methods of phylogeny estimation. Figure 3 shows the maximum-likelihood trees of phylogenetic relationships among the aligned haplotype sequences of the Gag and Vif regions of HIV-1 from the ten patients. The topology of the tree shows distinct patient-specific clades, each with >95% of branch support [41], except for patient S007. Consistent with the short time since transmission and rapid expansion of virus from a distinct transmitted founder in the new host, the phylogenetic tree for patient S007 has short branch lengths and few internal branches.
A particular transmitted founder virus, which had been subjected to a stringent genetic bottleneck, successfully established HIV-1 infection in nine of the ten patients studied. For seven of the ten patients, the viral sequences formed distinct patient-specific monophyletic lineages, each with high statistical support (>99% probability). Viral sequences from a multiply infected patient (S007) did not coalesce at a single transmitted founder in the maximum-likelihood tree, consistent with more than one transmitted founder virus being responsible for establishing a productive infection. In two patients of known sexual congress (S004 and S005), each of who had a distinct transmitted founder virus, there was intermixing of viral sequences at the later sample time point, which were valid and did not result from cross-contamination of amplicons.
The Highlighter plots of viral sequences from each patient showed the random distribution of nucleotide polymorphisms across them consistent with a dispersal of variants that arise from a particular transmitted founder virus (Figure S3). Maximum diversity of sequences within the discrete viral lineages from the nine patients with a particular transmitted founder virus was low (mean 0.44%; range 0.15 to 0.73% and mean 0.53%; range 0.10 to 1.56% for the Gag and Vif genes of HIV-1, respectively). The maximum diversity of sequences from the patient with more than one transmitted founder virus (mean 1.38% and 1.87% for the Gag and Vif genes of HIV-1, respectively) exceeded that found in the viral sequences from the other patients.
Based on the evidence for trinucleotide contexts of edited sites for the APOBEC3 deaminases from titration transfection experiments, we examined the potential for APOBEC3 editing of HIV-1 DNA to contribute to adaptation and evolution in natural infection. Analysis of specific editing frequencies at individual guanosines for each of the four APOBEC3 genes revealed a clear overlap with sequence changes observed in patients (Figure 4). Genomic context, such as adjacent nucleotides or local structural constraints, may have moderated against the effects of APOBEC3 editing at certain nucleic acid positions. Because the trinucleotide context of edited sites is shared among APOBEC3G, APOBEC3F and APOBEC3H, the two former cytidine deaminases could affect APOBEC3 editing for the seven patients who carried the less active form of APOBEC3H (haplotype I; homozygotes S001, S005, S007, and S010 and heterozygotes S004, S006, and S009; Table S3).
We next assessed whether APOBEC3 editing contributes to the genetic diversification of the virus populations in these patients. Viral sequences from the first time point differed from their respective consensus sequence by a median value of two (2) nucleotides in the Gag gene of HIV-1 and one (1) nucleotide in the Vif gene of HIV-1. At the second sampling time point 24 to 26 weeks later, the sequences from the Gag and Vif genes of HIV-1 differed from the particular consensus sequence by a median value of three (3) and two (2) nucleotides, respectively. We found relatively low frequencies of per nucleotide site insertions-deletions (indels) (1.47×10−3 and 9.42×10−5 for the Gag and Vif genes of HIV-1, respectively) and stop codons (2.44×10−4 and 1.03×10−4 for the Gag and Vif genes of HIV-1, respectively).
The sequence diversity within each patient was calculated as the Hamming distance between sequences after weighting by the number of collapsed sequences and correcting for the sequence length. The frequencies of nucleotide substitutions measured across the sequenced regions gauges the proportion of nucleotide sites at which the viral sequences being compared are different (p-distance). We excluded the multiply infected patient (S007) and restricted our analysis to the nine patients with infection consistent with a single transmitted founder virus in which the observed sequence diversity is expected to be due to mutations that have happened after HIV-1 transmission. The maximum number of variable nucleotide sites within individual viral populations ranged from 0.6% (with mean diversity 0.3%) at the early time point to 0.9% (mean = 0.6%) at the later time point for the Gag gene of HIV-1 and 0.7% (mean = 0.3%) at the early time point to 2.4% (mean = 0.8%) at the later time point for the Vif gene of HIV-1 (Figure 5). After weighting by the number of collapsed sequences, the inter-sequence pairwise distances for the sequence sets at the early time point were significantly lower than at the later time point during infection (Wilcoxon sum rank test, P<0.05 for both Gag and Vif genes of HIV-1). The overall average ratio of nonsynonymous to synonymous nucleotide substitutions (dN/dS), estimated using the Nei-Gojobori algorithm as implemented in SNAP [42], were consistent with strong purifying selection over the time points sampled (HIV-1 Gag = 0.37, standard error of the mean [SEM] = ±0.07; and HIV-1 Vif = 0.23, SEM = ±0.04).
The overall mutation frequency among the nine patients, after correction for sampling time and frequency of the collapsed haplotype sequences with a maximum-likelihood analysis assuming a strict molecular clock, was estimated to be 3.8×10−3 and 6.5×10−3 per substitution per site per year for the Gag and Vif genes of HIV-1, respectively. When a Bayesian approach was taken and a strict molecular clock was used, the results obtained were very similar with estimated evolutionary rates of 4.1×10−3 and 4.2×10−3 per substitution per site per year for the Gag and Vif genes of HIV-1, respectively. Significantly, the viral sequences were sampled over a time period in which the evolutionary rate does not mirror a compound mutation and substitution rate. These values are consistent with the estimations of others [43]–[45].
To further assess the effects of APOBEC3 editing that happen at sub-lethal levels on the genetic structure of the HIV-1 populations, we analyzed the proportion of viral sequences with nucleotide changes happening in this way as well as their contribution to the genetic diversity of the viral populations. It follows that among the APOBEC3 trinucleotide contexts of the edited sites, which are distinguishable from the more random RT-induced G-to-A mutations, viral diversification should increase through acquisition of neutral or beneficial substitutions all the while circumventing the introduction of a deleterious stop codon or loss of an initiation codon. The frequencies of G-to-A mutations in the HIV-1 Gag gene sequences from the first to the second time point averaged: 31% (1.9/6.2 potential sites) for 5′-GGA-3′; 35% (3.1/8.5) for 5′-GGG-3′; 17% (0.5/2.8) 5′-GGT-3′; and 19% (1.2/6.5) for 5′-GAG-3′. In the HIV-1 Vif gene sequences, these frequencies averaged: 18% (1.6/8.9) for 5′-GGA-3′; 13% (1/7.8) for 5′-GGG-3′; 6% (0.3/4.8) for 5′-GGT-3′; and 15% (0.9/5.4) for 5′-GAG-3′.
To confirm that the APOBEC3 activity in virally infected cells may influence the substitution biases that could increase the substitution rate, we compared the pairwise genetic distance (nucleotide changes per site) with the complete alignments after removing the guanosine position from the APOBEC3 trinucleotide contexts of the edited sites identified in the cell culture experiments (5′-GAD-3′ and 5′-GGA-3′ for APOBEC3F and APOBEC3H and 5′-GGD-3′ and 5′-GAG-3′ for APOBEC3G) from the patients' collapsed alignments. Importantly, this process resulted in the pairwise distance between sequences being decreased significantly (Wilcoxon rank sum test P-value<0.05; Figure 5). Transition (purine to purine or pyrimidine to pyrimidine) and transversion (purine to pyrimidine or pyrimidine to purine) median ratio values were 4.66 (range 1.73 to 57.44) in the Gag and 5.34 (range 1.45 to 11.27) in the Vif genes of HIV-1. G-to-A (and C-to-T) transitions accumulated 5-fold faster than A-to-G (and T-to-C) transitions, an inequality in the evolutionary trajectory [32], [46]. In sum, these results demonstrate that stochastic or transient changes in APOBEC3 deaminase activity could have relevance for the directionality of HIV-1 evolution in natural infection.
To determine whether natural selection acting on G-to-A mutations found in the APOBEC3 trinucleotide context of the edited sites could facilitate evasion of host immunity, we screened known or potential cytotoxic T lymphocyte (CTL) epitopes for positively selected sites. CTL epitopes had been established experimentally by interferon-γ enzyme-linked immunospot (ELISPOT) or predicted on the basis of amino acids that could serve as anchors to enable HLA binding or affect proteosome cleavage sites that abolish peptide binding, lessen T cell receptor recognition, or generate antagonistic CTL responses [47]–[49]. As an indicator of amount of natural selection operating on these CTL epitopes, we undertook a site-specific analysis of dN/dS in the Gag and Vif genes of HIV-1 using the Single Likelihood Ancestor Counting (SLAC) method implemented in HyPhy [50]. We focused on the G-to-A changes among the positions identified by reason of their significant selection signal (complete list and description in Table S5) and grouped them into positions appearing within or outside APOBEC3 trinucleotide contexts of the edited sites. We found an overrepresentation of positively selected positions within the APOBEC3 trinucleotide context of the edited sites in the Vif gene of HIV-1 (Fisher's exact test, P = 0.02; Table 2). Moreover, some of these positively selected positions in the Vif gene of HIV-1 appeared within a known or predicted HLA-appropriate epitope (S004 and S009, Table S5).
HIV-1 sequences encoding variants that could result in a lower predicted peptide binding score which would potentially confer a diminished or immune escape phenotype, were found within a HLA-appropriate epitope in the Gag (19 of 42; per patient range, 0 to 6) or Vif genes of HIV-1 (37 of 99; per patient range, 0 to 4) at the later time point during infection (Tables 3 and S6). Most epitope escape mutations were found in those patients that carry HLA-A01:01:01, HLA-A02:05:01 or HLA-A03:01:01 or HLA-B07:02:01, HLA-B08:01:01 or HLA-B57:01:01 (S001, S002, S003, S009, and S010). Of the sites under positive selection, 7 of 19 sites in the Gag and 8 of 42 sites in the Vif genes of HIV-1 were a result of G-to-A mutations in APOBEC3 editing contexts. Clusters of G-to-A mutations in known or predicted HLA-appropriate epitopes were higher in some patients (S001 and S009) than in others (S002, S003, and S010). We found statistically significant evidence for G-to-A mutations in APOBEC3 trinucleotide contexts of edited sites that cause nonsynonymous substitutions in the amino acid residues located at the epitopes (Fishers exact test P<0.05). These data demonstrate that APOBEC3-induced mutation embedded in the HLA-restricted epitopes can accumulate over time as a consequence of immune selection pressure.
To identify G-to-A mutations in APOBEC3 trinucleotide contexts of edited sites in known or predicted HLA-appropriate epitopes in relation to the most common haplotype in the first time point in each patient, we compared the epitopes at both early and late time points during infection. At the second time point during infection, G-to-A mutations were preferentially found at 1/13 sites in the Gag and 6/23 sites in the Vif gene of HIV-1 (Table S6). Though only direct experimental studies can establish which of the G-to-A mutations are associated with the evasion of host immunity, we infer that many of these are positively selected sites at low frequency variants at the first time point that transition to fixation at the second. In this manner, APOBEC3 editing can affect the crucial interaction between the virus and the host during the earliest stages of infection, and thereby potentially influence the natural history of HIV-1 infection.
In this study we define at unprecedented depth the specific APOBEC3 trinucleotide contexts of edited sites in cell culture experiments and show that the equivalent mutations that accrue in viral DNA in cells from patients through time provide a source of genetic variation upon which natural selection acts; thus, resolving the widely debated contribution of APOBEC3 editing to the genetic changes underlying the evolution of HIV-1 populations in natural infection [13], [22], [33]. Using a statistical framework that detects and corrects pyrosequencing errors, we show that APOBEC3D, APOBEC3F, APOBEC3G, and APOBEC3H, the cytidine deaminases of the human APOBEC3 gene family that are the cellular targets for HIV-1 protein Vif, have distinct, but overlapping trinucleotide contexts of the edited sites associated with antiviral defense. Mapping these APOBEC3-mediated G-to-A mutations onto the viral sequences from ten patients with primary HIV-1 infection through time is informative of the impact that these human genes can have on virus diversification. The over-representation of G-to-A mutations in the viral sequences compared with A-to-G or C-to-T or T-to-C mutations (in a A-rich, C-poor genome) suggests that accumulation of APOBEC3 mutations is well tolerated in diversifying sites and could account for the skewing of nucleotide and codon usage in the viral genome. We note that the total number and location of the APOBEC3 trinucleotide context of the edited sites within the viral genome and the extent to which they can accumulate through time need to be accounted for in evolutionary inference at the population-level.
The tandem array of the seven human cytidine deaminases of the APOBEC3 gene family on chromosome 22, which we distinguish by their target sequence consensus, suggest that multiple, related antiviral functions can contribute to the control of virus infection. Differences in single-stranded DNA binding, as well as translocation along engaged templates, may explain the sequence specificity of APOBEC3 activity and processing accuracy [51], [52]. Nucleotides adjacent to the APOBEC3 editing context likely influence the kinetics of G-to-A mutation. Functional biases in cytidine deaminase activity suggest that people may differ in the predominant expression of APOBEC3 and that these functionality-related genes may play a role in the spectrum of innate resistance that protects against invading viruses and contributes to phenotype. This conclusion, which could apply to other types of viruses or retroviral elements, suggests that human APOBEC3 proteins have clear impact at the boundary between the virus and its host.
It has been posited that limiting-levels of APOBEC3 activity could result in lethal mutations rather than rapid adaptation through acquisition of neutral or potentially beneficial mutations [33]. Further, that a single incorporated APOBEC3 unit is likely to cause extensive and inactivating levels of HIV-1 hypermutation. These conjectures are based on in silico analyses of optimized reference sequences that would be estimated to account for the mutation levels of 39 near-full length patient-derived hypermutated viral sequences selected from the HIV Sequence Database. As the analyses began with highly mutagenized HIV-1 genomes with 5′-GpG-3′ or 5′-GpA-3′ signatures of APOBEC3 editing from which the non-hypermutated reference sequences were derived, the estimated effect is distorted by a clear selection bias. Studies that have original patient-derived non-hypermutated reference sequences clearly corroborate the relevance of small increases in mutation frequency affected by APOBEC3 for genetic changes underlying virus evolution [53].
Natural selection during virus infection can create advantageous mutations or eliminate deleterious ones. The rate of fixation of advantageous mutations, which is faster than the rate of fixation of neutral mutations, increases with the strength of selection. We found statistically significant evidence for positive selection acting on the Vif region of HIV-1 in the APOBEC3 trinucleotide context of the edited sites. Even though blockade by the HIV-1 protein Vif effectively counters the action of certain human APOBEC3 proteins that could lead to the lethal accumulation of mutations, some G-to-A mutation is produced that can increase genetic diversity and facilitate adaptation; an apparent shortcoming of APOBEC3 editing that should caution against the use of a HIV-1 Vif antagonist as a virus inhibitor.
When the trinucleotide context of the edited sites rests within an HLA-appropriate epitope so that G-to-A mutation affects peptide binding or T-cell antigen receptor recognition, the APOBEC3 proteins could be an important driver of mutations that enable a virus (of reduced replicative fitness) to evade host immunity [27]. G-to-A mutations within APOBEC3 trinucleotide contexts of the edited sites may be constrained, however, by pressure to retain possible useful structural elements or functional sites [54]. We show here that a number of diversifying codon sites in an APOBEC3 trinucleotide context, as indicated by the accumulation of non-synonymous nucleotide changes, were clustered within a number of HLA-restricted epitopes that could act as anchors for HLA binding or in the proximal three amino acid regions that could affect peptide processing. These data reveal that APOBEC3 can generate viral mutations in immune-susceptible locations that are subjected to strong positive selective pressure during the acute phase of infection.
Our analyses provide strong statistical evidence for an association between G-to-A mutation rates and HIV-1 diversification in natural infection. Consistent with APOBEC3 evolutionary footprints in the viral genome, we find a higher frequency of mutations in APOBEC3 than non-APOBEC3 edited sites introduced during sequential generations of HIV-1 within patients. The sub-lethal APOBEC3 editing that make subtle contributions to viral sequence diversity can lead to mutational fitness effects that should facilitate host adaptation, having been associated with the evasion of host immunity and evolution of resistance to antiretroviral drugs [21], [55]. The longitudinal analysis of HIV-1 infection in these ten patients gives important new insight into the causes and consequences of virus diversity upon which selection can act. The findings described here, therefore, suggest that the genetic conflict caused by the APOBEC3 innate immune effectors is an important determinant in explaining the mutational dynamic and directionality that underlies HIV-1 evolution.
293T cells, which express little or no endogenous APOBEC3, were cultured in Dulbecco modified Eagle medium (DMEM) supplemented with 10% fetal bovine serum plus penicillin-streptomycin. Sub-confluent monolayers of 293T cells seeded in 35-mm plates were co-transfected with 3.0 µg of the vif-deficient HIV-1 pIIIB/Δvif construct, a vesicular stomatitis virus G (VSV-G) protein expression vector, and between 0.01 µg and 3 µg of pcDNA3.1-based expression vectors for APOBEC3D, APOBEC3F, APOBEC3G or APOBEC3H (haplotype II) using polyethylenimine (PEI; Table 1). Forty-eight hours later, viral supernatants were harvested, treated with 20 U/ml RQ1 DNase (Promega) in 10 mM MgCl2 for 3 h at 37°C, and then purified by pelleting through a sucrose cushion. Virus was quantified by a HIV-1 p24 Gag enzyme-linked immunosorbent assay (ELISA; Perkin-Elmer).
Sub-confluent layers of 293T cells were infected with VSV-G-pseudotyped virus stocks equivalent to 50 ng HIV-1 Gag p24. The input virus was removed four hours later, and the cells were thoroughly washed before the addition of fresh medium. Two days after infection, the supernatant was harvested for quantification by HIV-1 Gag p24 ELISA. Viral infections were determined in single-cycle assays as described [56]. Total genomic DNA was isolated with the QIAamp DNA cell mini kit (Qiagen) and purified DNA was digested with Dpn I to remove any residual plasmid DNA. To normalize the input amount of viral DNA for sequencing using the 454 Life Sciences' GS-FLX pyrosequencing system (Roche), we measured the amount of the HIV-1 Gag gene DNA by qPCR as described [21]. The relative amount of HIV-1 target DNA was normalized to the quantification cycle for a concentration calibrator by using an external standard curve of serial 10-fold dilutions of a reference Gag gene of HIV-1 DNA.
Over a period of up to 26 weeks, we tracked changes in the nucleotide sequences from the Gag and Vif genes of HIV-1 in peripheral blood sampled from ten infected patients. All ten patients had confirmed HIV-1 infection, were enrolled in a study of early HIV-1 infection. Acute HIV-1 infection was defined by the presence of HIV-1 RNA in plasma and a negative or weakly positive HIV-1 ELISA followed by a positive one. Early HIV-1 infection was defined by the presence of a positive HIV-1 ELISA confirmed by a detuned negative HIV-1 ELISA. All participants in the study had symptoms compatible with the acute retroviral syndrome and were treated with a combination of potent antiretroviral drugs (one protease inhibitor and two nucleoside reverse transcriptase inhibitors) within a median of 2 years (range, 1.5 to 3 years) from the time of diagnosis. Clinical and laboratory data and sample collection begins at enrollment and at prescribed interval study visits thereafter.
All patients provided written informed consent according to the guidelines of the Human Subjects Protection Committee of the University of California, San Diego. The University of California, San Diego Institutional Review Board approved the study.
Genomic DNA was isolated from frozen PBMC samples (approximately 2 million cells) using the QIAamp DNA blood mini kit (Qiagen) according to manufacturer protocol. DNA was eluted in nuclease-free water (100 ul) and stored at −80°C until use. The amount of HIV-1 DNA was measured by qPCR of the Gag gene of HIV-1 with the TaqMan Universal PCR Master Mix (Applied Biosystems) on the 7900HT sequence detector (Applied Biosystems) as described [21]. The relative amount of HIV-1 target DNA was normalized to the quantification cycle for a concentration calibrator by using an external standard curve of serial 10-fold dilutions of reference HIV-1 linear full-length plasmid DNA derived from the pNL-43 plasmid. The amount of input cell DNA was normalized to the amount of human CCR5 amplified using the forward primer CCR5-F 5′-ATCGGAGCCCTGCCAAAA-3′, the reverse primer CCR5-R 5′-TGAGTAGAGCGGAGGCAGGAG-3′, and probe CCR5-P 5′-FAM-CGGGCTGCGATTTGCTTCACATTG-BHQ-3′. All reactions were performed in quadruplicate.
High resolution HLA genotyping was performed by next generation sequencing of exonic amplicons using the 454 Life Sciences' GS-FLX pyrosequencing system (Roche) with Conexio Assign ATF 454 software as described [57].
We performed genotype and haplotype analysis of APOBEC3H for previously identified variants that cause the amino acid polymorphisms R18L, G105R, K121D, and E178D by means TaqMan SNP Genotyping assays with an Applied Biosystems 7500 Real-time PCR detection system. For the N15Δ, K121D, K121N, and K121E polymorphisms, primers were designed to produce a PCR-product DNA that could be sized and sequenced. First round PCR was performed with A3H_EK2852F (5′-AGGCAGGAGAATCGCTTGAACTTG-3′) and A3H_EK4571R (5′-CCTCCCGGGTGGTGTCAGAT-3′) to amplify exon 1 and exon 2 for 30 cycles (94°C-30 sec; 58°C, 30 sec; 72°C, 1.5 min). For 121 polymorphisms, PCR-product DNA was diluted and directly sequenced with A3H_EK4112F (5′-CCCCTGCTCCTCCTGTGCCT-3′) and A3H_EK4522R (5′-CTTCCTGGCCTCCCACAGACC-3′). For N15Δ polymorphism, primers A3H_EK3068F-FAM (5′-FAM-ACAGCCGAAACATTCCGCTTACAG-3′) and A3H_EK3204R (5′-TTGTTTTCAAAGTAGCCTCTCGTGGG-3′) were used with Taq polymerase for initial denaturing for 2 min at 94°C, followed by 30 cycles (94°C, 15 sec; 58°C, 1 min) and a final 5 min extension at 72°C. DNA fragment analysis was performed on an Applied Biosystems 3730xl DNA Analyzer with 36 cm capillary array and analyzed by GeneMapper 4.0 software (Applied Biosystems/Life Technology).
To facilitate quantitative sampling of the viral population, we performed viral DNA amplification by PCR using high template volume, low cycle numbers, and multiple replicates that were pooled for sequencing. PCR primer design was predicated upon the alignment of multiple sequences from the HIV Sequence Database [58] to minimize biased amplification of the target DNA. We selected highly conserved regions in the Gag and Vif genes of HIV-1 that encompassed the APOBEC3 trinucleotide context of edited sites and known or predicted epitopes, which were an appropriate distance apart and within the read-length limits of the 454 sequencing technology employed. Primers design considered preexisting alignment covering the region of interest. Conserved primer locations were selected based on alignment positional entropy. Within the selected conservative sequences, we used degenerative bases for the APOBEC3 editing context. Unique molecular identifiers that label individual molecules in the pool moderate against erroneously attributing multiple identical sequences to low viral diversity and allelic skewing by biased PCR amplification.
Viral DNA isolated from 293T cells was amplified using the HIV-1 NL4-3 Gag gene-specific degenerate forward primer gag_F1329dg (5′-CGTATCGCCTCCCTCGCGCCATCAG [fusion primer A]- [multiplex identifier sequence (MID)]-CCCCACAARATTTAAACACCAT-3′, corresponding to positions 1328→1349) and reverse degenerate primer gag_R1785dg (5′-CTATGCGCCTTGCCAGCCCGCTCAG [fusion primer B]- [MID]-GTYTTACAATYTGGGTTYGCAT-3′, corresponding to a 1784→1763 reverse complement) to generate 457 bp of the Gag gene of HIV-1 (HXB2 genome 1328→1784).
Viral DNA in PBMC samples from patients was amplified using the HIV-1 Gag gene-specific forward primer A-Gag_977F_degEK (5′-primer A-GCTACAACCAKCCCTYCAGACAG-3′, corresponding to positions 977→1000) and the reverse primer B-Gag_1564R_degEK (5′-primer B-CTACTGGGATAGGTGGATTAYKTG-3′, corresponding to a 1564→1541 reverse complement) to generate 588 bp of the Gag gene HIV-1 (HXB2 genome 977→1564) or the HIV-1 Vif gene-specific forward primer A-Vif_5041F-EK innerF (5′-primer A-ATGGAAAACAGATGGCAGGTG-3′, corresponding to positions 5041→5061) and the reverse primer B-Vif_5623R-EK_innerR (5′-primer B-AGCTCTAGTGTCCATTCATTGTATG-3′, corresponding to a 5623→5599 reverse complement) to generate 583 bp of the Vif gene of HIV-1 (HXB2 genome 5041→5623).
PCR was performed using the High Fidelity Platinum Taq DNA Polymerase (Invitrogen) with thermal cycling conditions of 94°C for 2 mins, followed by 35 cycles of 94°C for 15 sec, 54°C for 15 sec, 68°C for 1 min, with a final extension step at 68°C for 5 mins. DNA amplicon libraries were resolved on a pre-cast 2% agarose gel and purified with QIAquick Gel Extraction kit (Qiagen) and AMPure XP SPRI beads (Beckman Coulter Inc.). To determine amplicon library quality, a Bioanalyzer (Agilent) was used and quantity for amplicon samples along with KAPA Library Quant Kit (KAPA Biosystems). An equimolar mix of the amplicon libraries was subjected to emulsion PCR and DNA sequencing using the 454 Life Sciences' GS-FLX pyrosequencing system (Roche). Multiplex identifiers were used to bin the sequence reads before analysis.
As the quality of sequence may decrease across a sequence read, we quality filtered the sequence data before analysis [31]. Each sequence read had to pass a series of standard metrics to ensure the output of high quality sequence reads while maintaining the maximum possible length of the output sequences. We excluded sequences that had a frameshift relative to a reference sequence, were too short, contained long direct repeats, were a recombinant, or had a close pair of matched-length indels that created a short compensated frameshift. In this analysis, we used the same methods and parameters for obtaining a clean alignment as in previous work [31]. Viral sequences that were observed only once were also excluded to further reduce technical noise. After k-mer mapping, reads were pairwise aligned to the consensus template pIIIB/Δvif reference sequence for the cell culture experiments, taking into account data-specific indels, and thereby reducing dependency on a generic template.
We aligned the viral sequences obtained from peripheral blood from patients by Segminator II (version 0.1.1) using the HXB2 sequence (GenBank accession number K03455) as a reference for assembly [31], [59], [60]. We generated a consensus sequence for each patient, which was then used as a reference sequence for the patient-specific re-alignment. We used a statistical model that accounts for site-specific error rates to separate errors from true variations and remove chimeric molecules that arise from PCR or pyrosequencing errors by applying the Predator algorithm default implemented in Segminator II. This statistical framework maintains the reading frame and corrects for length errors in homopolymeric runs of nucleotides, the characteristic error of 454 pyrosequencing. Nucleotide sequence alignments used in this study were deposited in GenBank with the accession numbers (KJ016272–KJ017738).
We performed a tally of G-to-A mutations among the viral sequences from the titration transfection experiments that contain a guanosine within a binucleotide, trinucleotide, tetranucleotide, or hexanucleotide context of the edited sites for the HXB2 reference sequence using a sliding window. Each G-to-A mutation was considered independently. A tally of C-to-T mutations was used to assess the extent of noise in the analysis. The viral sequences were treated as character arrays, and therefore each trinucleotide context of the edited sites was compared separately with the same location in the reference sequence.
For each position, we separately checked for the fraction of mutations increasing with the concentration of APOBEC3 by calculating the significance of the Spearman rank correlations, without correcting for multiple testing. Only those positions that changed significantly with increasing concentrations of the different enzymes for both forward and reverse reads were chosen as possible APOBEC3-induced G-to-A change. Motifs that were statistically significantly overrepresented in this dataset were designated as APOBEC3 trinucleotide editing contexts and used in further analyses. We did not find any significant associations with longer motifs. The strength of the effect was evaluated by calculating the slope from a logistic regression of the mutation probability against the concentration of the enzyme.
To analyze APOBEC3 editing in patients, we calculated the sequence divergence for each position in the Gag and Vif genes of HIV-1 and determined base frequencies at each position in the alignment. We analyzed the G-to-A mutations in the defined APOBEC3 trinucleotide editing contexts that increased significantly between the two time points. We used the previously cleaned alignments to calculate the frequency of G-to-A mutations at each position with guanosine in the early consensus sequence by using an in-house Perl script. These ratios were compared between time points by Fisher's exact test and G-to-A mutations increasing in the later time point with a P-value<0.01 were taken into account for further analysis.
To avoid using uninformative sequence repeats, the viral sequences were collapsed into shared non-recombinant haplotypes representing only unique sequences using the tools in the FASTX-toolkit (version 0.0.13) implemented in Galaxy (https://main.g2.bx.psu.edu). The collapsed haplotypes in each patient were realigned using the alignment method implemented in MUSCLE (version 3.8.31) [61]. We kept sequences corresponding to variants present above 0.2% of the total existing variants in the collapsed alignments. The best-fit model of nucleotide substitution was estimated for each of the collapsed alignments with a maximum-likelihood method using PhyML version 3.0 as implemented in jModelTest2 [62]. We constructed maximum-likelihood trees for the Gag and Vif genes of HIV-1 with and without the APOBEC3 trinucleotide context of the edited sites for each patient. Because maximum-likelihood tree error may increase when unreliable sites are included, we estimated trees on viral sequence sets from which gaps in the alignment were removed and considered as missing data. Ancestral states of each node of the trees constructed with the complete alignments (including the trinucleotide context of the edited sites) were estimated by a maximum-likelihood method using PhyML version 3.0 [62] applying approximate likelihood ratio test (aLRT) for branch support [41]. Nucleotide sequences from all ten infected patients were analyzed with phylogenetic trees using the neighbor-joining method together with the Highlighter sequence visualization tool (www.HIV.lanl.gov) to trace commonality between sequences in an alignment based on individual nucleotide changes.
To assess the type of evolutionary forces operating on the patient derived viral sequences, we estimated the overall ratio of dN/dS using Synonymous Non-synonymous Analysis Program (SNAP) (www.HIV.lanl.gov) [42]. Site-specific analyses of dN/dS were undertaken using the SLAC method implemented in HyPhy at the Datamonkey webserver (http://www.datamonkey.org/) [50], [63]. We conducted this analysis with the general reversible (REV) model of nucleotide substitution on phylogenetic trees using the neighbor-joining method (cut-off P-value = 0.1). We selected the G-to-A changes among all the identified positions under selection and then analyzed their possible association with APOBEC3 activity by testing whether there was an overrepresentation of G-to-A mutations in an APOBEC3 editing context.
The mutation rates were estimated from the maximum-likelihood phylogeny of viral sequences collected at the early and late time points during infection. To account for the time-dependency of evolution rate estimates, we rooted the tree at a position most compatible with a strict molecular clock and analyzed the slope of the regression of root-to-tip distances against the dates of sampling with Path-O-Gen version 1.4 (http://tree.bio.ed.ac.uk/software/pathogen/). Evolutionary rates obtained using this maximum-likelihood method were confirmed by a Bayesian approach using phylogenetic reconstructions of the complete alignments of the ten most represented haplotypes for each patient sampled at early and late time points during infection. This analysis was performed using BEAST version 1.7.4 [64] with a GTR + gamma evolutionary model assuming a strict molecular clock and a constant population size. We computed the posterior probability of the model to obtain the Bayes factors to discriminate among the models.
We calculated the number of substitutions per site, performing a pairwise comparison of every sequence to the first time point consensus for viral sequences from each patient's collapsed alignments using MEGA (version 5.2.2), weighting the obtained values with the number of sequences of each haplotype, and then determining the weighted average of each time point. Two alignments of the collapsed haplotypes in the sequenced populations were generated: One without the sites containing guanosine in the 5′-GAD-3′ and 5′-GGA-3′ trinucleotide contexts in the first time point consensus (corresponding to APOBEC3F and APOBEC3H editing), and the other without the sites containing guanosine in 5′-GGD-3′ and 5′-GAG-3′ trinucleotide context (corresponding to APOBEC3G). We removed the guanosines on APOBEC3 trinucleotide contexts of the edited sites from the alignments using the Unix stream editor command ‘sed’. Pairwise distances of the complete alignments in each time point were compared with the ones obtained in the absence of guanosine in the trinucleotide context of the edited sites using a Wilcoxon signed-rank sum test (P<0.05). Values were normalized with their respective sequence length prior to comparisons.
We genotyped the six-digit HLA types using a next-generation sequencing method as described [65] (Table S2). To investigate whether the CTL epitope variants or escape mutations were mediated by APOBEC3 activity, we retrieved known or predicted MHC class I-restricted epitopes from the HIV Molecular Immunology Database [66] (http://www.hiv.lanl.gov/content/immunology/variants/variant_search.html).
Standard descriptive statistics were performed with the use of the STATA, GraphPad or R packages (version 1.1-1, http://CRAN.R-project.org/package=binom) [67], [68].
|
10.1371/journal.ppat.1003170 | Dynamic Epigenetic Regulation of Gene Expression during the Life Cycle of Malaria Parasite Plasmodium falciparum | Epigenetic mechanisms are emerging as one of the major factors of the dynamics of gene expression in the human malaria parasite, Plasmodium falciparum. To elucidate the role of chromatin remodeling in transcriptional regulation associated with the progression of the P. falciparum intraerythrocytic development cycle (IDC), we mapped the temporal pattern of chromosomal association with histone H3 and H4 modifications using ChIP-on-chip. Here, we have generated a broad integrative epigenomic map of twelve histone modifications during the P. falciparum IDC including H4K5ac, H4K8ac, H4K12ac, H4K16ac, H3K9ac, H3K14ac, H3K56ac, H4K20me1, H4K20me3, H3K4me3, H3K79me3 and H4R3me2. While some modifications were found to be associated with the vast majority of the genome and their occupancy was constant, others showed more specific and highly dynamic distribution. Importantly, eight modifications displaying tight correlations with transcript levels showed differential affinity to distinct genomic regions with H4K8ac occupying predominantly promoter regions while others occurred at the 5′ ends of coding sequences. The promoter occupancy of H4K8ac remained unchanged when ectopically inserted at a different locus, indicating the presence of specific DNA elements that recruit histone modifying enzymes regardless of their broad chromatin environment. In addition, we showed the presence of multivalent domains on the genome carrying more than one histone mark, highlighting the importance of combinatorial effects on transcription. Overall, our work portrays a substantial association between chromosomal locations of various epigenetic markers, transcriptional activity and global stage-specific transitions in the epigenome.
| Malaria is a devastating parasitic disease caused by the protozoan protist Plasmodium falciparum. The complex life cycle of P. falciparum comprises various morphological and functionally distinct forms and is completed in two different hosts. Various regulatory mechanisms are employed by these parasites to complete their life cycle and survive in human hosts. Epigenetic mechanisms, though not fully explored, have been implicated as one of the key players in gene regulation, morphological differentiation and antigenic variation. Here, we present a comprehensive epigenetic map of 12 histone post-translational modifications during the intraerythrocytic life cycle of P. falciparum. We have been able to identify at least eight histone modifications whose dynamic patterns correlate with the transcriptional regulation across the life cycle. In particular, we have shown that a set of euchromatic histone marks work in synergy, creating a dynamic unique histone code that is linked with gene expression during the progression of the Plasmodium intraerythrocytic developmental cycle. These findings enhance our knowledge of complex gene regulation and will help to identify novel targets for fighting malaria.
| In spite of worldwide efforts, malaria remains one of the most devastating illnesses with an estimated 216 million episodes leading to 655,000 deaths in 2010 [1]. The effectiveness of current treatment strategies is attenuated by increasing resistance of malaria parasites to the available chemotherapeutic drugs. The emergence of artemisinin resistance [2], [3] has motivated researchers to develop alternate control mechanisms by identifying new drug targets. As such, there is a rapid advancement of genomic and epigenomic research to unveil unique molecular mechanisms associated with the growth and development of malaria parasites. Plasmodium falciparum, the causative agent of the most severe form of malaria, is also the model organism to study the parasite development due to its ability to be grown in vitro. The clinical manifestations of malaria are a result of the parasite development in the red blood cells where it completes its asexual intra-erythrocytic developmental cycle (IDC). Even though transcriptional regulation is important for all developmental stages, the IDC transcriptome revealed a particularly distinct temporal transcriptional regulatory system in P. falciparum [4]. Such a broad and dynamic character of transcriptional regulation where each gene is expressed only at a specific time is unprecedented amongst known living organisms and likely represents a unique evolutionary adaptation of the parasite to its host. The presence of plant-like apicomplexan AP2 (Api-AP2) transcription factors [5] and the general paucity of many other types of specific transcription factors [6] further contributes to the unique character of the parasite regulatory machinery. P. falciparum also displays several diverse features of its epigenome such as the absence of linker histone H1 [7], the absence of RNA interference machinery [8], the presence of DNA cytosine methyltransferase but apparent absence of DNA methylation [9], [10] and the presence of unusual histone variants with a unique set of modifications [11]. Unlike the majority of higher eukaryotes, P. falciparum chromatin is predominantly in a euchromatic state with only a few heterochromatic islands marked by trimethylation of lysine 9 of histone 3 (H3K9me3) [12], [13], [14]. Unlike Saccharomyces cerevisiae where K16 acetylation is the dominant modification present at 80% of all H4 molecules [15], K8 and K12 are the favored acetylation sites in P. falciparum H4 [11]. Nevertheless, consistent with several studies from yeast and mammalian models showing that regulation of gene expression is mediated by chromatin structure [16], [17], epigenetic states in P. falciparum have been shown to affect transcription [18], [19]. In our previous study, we have shown that a potent histone deacetylase inhibitor, apicidin, induces severe alterations in histone modifications as well as gene expression [20]. Recently, it was also shown that epigenetic factors affect clonally variant transcription in P. falciparum likely via switching between hetero- and euchromatic structures at several genetic loci that mainly encode factors involved in host-parasite interactions [21]. Moreover, there is also evidence suggesting links between the mode of action of artemisinin as well as its resistance mechanism with factors affecting histone modifications [3]. Taken together, these lines of evidence highlight the contribution of the chromatin environment in regulating transcriptional control in P. falciparum and stress the need to characterize the overall chromatin landscape as well as its effect on transcriptional regulation during the life cycle.
A total of 44 different post-translational covalent modifications on P. falciparum histones including acetylations and methylations have been recently identified [11]. Here, we provide insights into the temporal relationship between twelve of these post-translational modifications and their effect on global transcriptional regulation associated with the complex IDC. The dynamic changes in the transcript pattern during the P. falciparum IDC were reflected in the genome wide epigenomic profiles of eight of the studied histone marks. The transcription linked patterns were associated with enrichment of most histone marks predominantly at the start of coding regions while only one modification, acetylation of H4 at lysine 8 (H4K8ac) was found predominantly at the putative promoter regions. Our data also demonstrate co-operative binding of acetylation marks in modulating gene expression across the IDC.
In order to understand the dynamics of chromatin remodeling and its role in gene expression, we generated an epigenomic map comprising the genome-wide distribution of twelve histone modifications throughout the P. falciparum IDC. The main rationale was to recapture the transcriptional cascade of the IDC [4] and to investigate correlations between the occupancy of each histone modification and transcriptional activity along the genome. For this purpose, a large scale culture of highly synchronized cells was grown and samples were collected every 8 h across the 48 h IDC (Figure 1A). It is important to note that the same starting culture was used for the entire set of experiments including chromatin immunoprecipitation combined with microarrays (ChIP-on-chip), transcriptome and western blots to achieve the highest level of comparability within the dataset. For the genome-wide studies, we utilized a P. falciparum DNA microarray that contains probes representing both the open reading frames (ORF) and the upstream intergenic regions (IGR) (see Materials and Methods for further details). This microarray was used to carry out ChIP-on-chip to study the chromosomal distributions of thirteen histone marks that included twelve individual histone modifications involving acetylation (ac) of lysine (K) residues on histones H4 and H3, namely H4K5ac, H4K8ac, H4K12ac, H4K16ac, H3K9ac, H3K14ac and H3K56ac; methylation (me) of K or arginine (R) residues, namely H4K20me1, H4K20me3, H3K4me3, H3K79me3 and H4R3me2; and finally one combination of histone modifications H4ac4 (H4 tetra-acetylated at lysines 4, 8, 12 and 16). Apart from H3K9ac and H3K4me3 which have been studied in P. falciparum previously [13], [22], other histone marks were selected based on evidence about their roles in transcription from other eukaryotic systems [23], [24], [25], [26], [27], [28], [29]. Using commercially available antibodies directed against modified histones, we confirmed the presence of the thirteen epitopes on P. falciparum histones by western blotting directly (Figure S1A) or by peptide competition assays (Figure S1B) and by immuno-fluorescence microscopy (Figure S1C). The designed ChIP-on-chip strategy (Figure 1A) allowed us to generate abundance profiles of histone modification occupancy across the genome and at the same time global mRNA levels during the IDC.
The overview of the occupancy of histone modifications along the P. falciparum genome during the IDC revealed that the most abundant histone marks are H4K5ac, H4K12ac, H3K14ac, H4K8ac, H3K4me3, H3K56ac and H3K9ac which associate with >80% of the genome (Figure 1B). This is followed by H4R3me2, H4K20me1 and H4K16ac that associate with 65 to 80%, and finally H4ac4, H4K20me3 and H3K79me3 that associate with less than 60% of the genome (represented by the 14,773 microarray probes). This shows that the individual histone modifications exhibit specific occupancy patterns that reflect their distinct roles in the parasite chromatin structure and function. The visual display of the chromosomal distribution of histone mark occupancy further supports this observation showing distinct patterns of histone marks across the chromosomes but also the existence of some genetic loci marked by more than one modification (Figure 1C, black boxes).
Next we determined the enrichment of histone marks represented as log2 ChIP/input ratios with respect to their position within the Plasmodium genes. This was done separately for genes with different levels of expression: top, middle and bottom 10% in the rank of their mRNA levels in each IDC time point (Figure 1D). Overall we could divide the studied histone modifications into two groups, (i) those with a biased distribution in the IGRs and/or 5′ termini of the ORFs and (ii) those with no preference in their occupancy within the gene structures. The first group comprises four H4 (K8ac, K16ac, ac4 and K20me1) and four H3 (K9ac, K56ac, K4me3 and K79me3) modifications with higher enrichment at IGRs and gradual decrease towards the 3′ end of the genes. The extreme example is H3K4me3 with sharp IGR occupancy, which is consistent with previous suggestion that the primary role of this modification is to demarcate the non-coding regions in between P. falciparum genes [22]. Interestingly, while some modifications such as H3K4me3 and H3K79me3 retained this IGR preference throughout the IDC, others showed stage specific changes in their positional enrichment. These include H4K8ac and H3K56ac that were found at the IGRs predominantly in trophozoites and early schizonts but show essentially no gene position preference in the extremes of the IDC, early rings and late schizonts. For most of the histone marks, there were only small, likely insignificant, differences in their occupancy between genes with high, medium or low levels of expression. The exceptions are, H4K8ac, H4K16ac, H3K9ac and H3K56ac that exhibited somewhat higher IGR enrichment for genes with high mRNA levels (Figure 1D). In addition there was a slight tendency for all H4 acetylations to increase their enrichment towards the 3′ end of genes with low levels of mRNA during 16 to 30 hours post invasion (hpi). In summary the differential occupancy of histone marks within genes suggest their distinct roles as chromatin remodeling factors that may be linked with gene expression during the P. falciparum IDC.
The most significant observation made by these studies is the broad and dramatic dynamics of the occupancy of histone modifications across the IDC. Essentially all thirteen histone marks show some degree of variable occupancy at least for a small portion of the genetic loci with which they associate. For the purpose of this study, we define the occupancy variability by two criteria: (1) The statistical significance of the measured change in occupancy across the experimental time points with respect to experimental replicas (P<0.05), and (2) in addition to statistical significance (P<0.05), a minimum 1.5 fold change in the occupancy profiles across the IDC (Figure 1B, Table S1). Below, we refer to these as “dynamic histone marks” and “dynamic occupancy profiles”, respectively. Quantitative real time PCR was carried out to validate the dynamic occupancy profiles for three modifications in three genes (Figure S2A). We also wished to evaluate the performance of the microarray probes representing the IGRs in comparison to the ORFs. The signal-intensity/signal-ratio distributions between the sets of ORF and IGR probes show essentially identical profiles with no measurement bias towards any ratio/intensity interval (Figure S2B). This supports the fidelity of the applied microarray technology and ensures that the dynamic range and thus ChIP-on-chip measurements of histone occupancy are directly comparable between IGRs and ORFs. The two most dynamic histone modifications were found to be H4K8ac and H3K4me3 that showed significant changes of ChIP-on-chip signal at more than 50% of the loci with which these histone marks associate. Moreover H3K56ac, H3K9ac, H4ac4, H4K16ac, H4K12ac, H4K20me1 and H4K20me3 exhibited dynamic occupancy profiles at more than 25% of their loci. In contrast, H4R3me2, H3K14ac, H3K79me3 and H4K5ac represented the other side of the spectrum with a constitutive pattern of occupancy at the majority of the loci with only 20% or less showing variation across the IDC. The Chi-square test revealed a preference for localization of the dynamic histone marks, with H4K16ac, H4ac4, H4K8ac, H4K12ac, H3K56ac, H4K20me1, H4K20me3 and H3K4me3 showing overrepresentation in the ORFs, whereas H4R3me2 showed overrepresentation in the IGRs (Figure S3). The ORF preference occupancy of the dynamic histone marks is surprising as it is in contrast to their overall (constitutive and dynamic) occupancy in IGRs (Figure 1D). This may suggest that while the constant occupancy of these histone marks at the IGRs may function as general demarcation elements, it is the nucleosomes linked with the 5′ termini of the ORF regions that play a dynamic role in the chromatin remodeling and possibly transcriptional regulation during the IDC (see below).
Similar to mRNA, the occupancy patterns of histone modifications exhibited single peak profiles with each locus being marked once at a specific time during the IDC (Figure 2A). Investigating the time of peak occupancy, we observed no general trends, but instead each histone mark exhibited a distinct pattern (Figure 2B). In particular, more than 50% of genetic loci are associated with the dynamic occupancy of H4K20me1, H4K20me3 and H3K14ac which reach maximum levels between 0 and 16 hpi, whereas 40% of loci associated with the dynamic occupancy of H3K4me3 and H4K5ac peaked between 17 and 32 hpi. In contrast, H4K8ac, H3K9ac and H4R3me2 showed maximum occupancy (more than 40%) at the late schizont stage. Other histone marks were evenly distributed amongst the three stages. This global variation in the occupancy of histone marks again suggests their distinct role in chromatin remodeling with multiple events during the IDC affecting their overall distribution across the IDC.
The dynamic character of the histone modifications and its similarity to the mRNA abundance profiles during the IDC suggests their possible role in transcription. To investigate this, we evaluated the correlations between the occupancy profiles of the dynamic histone marks and steady state mRNA levels of the corresponding genes. Here we hypothesize that a synchrony between the histone marks and mRNA substantiates a link between their deposition at a particular gene and transcriptional activity. Hence we calculated Pearson Correlation Coefficients (PCC or r) between mean-centered profiles of the dynamic histone modification occupancy and the corresponding mRNA. For example, 49% (out of 7260) of the H4K8ac dynamic occupancy profiles showed positive correlations (r≥0.4) with transcription while 22.8 and 22.5% showed no or negative correlation, respectively (Figure 3A). The overall skew of the H4K8ac correlation values to the positive side suggests that this histone mark plays a role in transcriptional induction. Using the Kolmogorov-Smirnov test (KS test) against randomized data, we were able to evaluate the significance of the occupancy profile correlations with mRNA for all histone marks (Figure 3B). In addition, we utilized the degree of skewness (S) to identify all histone marks that are positively correlated with transcription. With P<0.0005 and S>0.05, we identified eight histone marks including H4K8ac, H4K16ac, H4ac4, H3K56ac, H3K9ac, H3K14ac, H3K4me3 and H4K20me1 that showed positive correlations with transcription and thus we refer to these as “transcription-linked” histone marks. For these eight histone marks, the percentage of probes that show a positive correlation with expression (r≥0.4) varied from 35 to 48% (Figure S4A). The genes associated with these transcription-linked histone marks show no bias to any particular developmental stage but instead are more or less evenly distributed amongst all stages of the IDC (Figure S4B). Interestingly, the histone marks which followed transcription are also amongst the most dynamic, with at least 25% of the loci changing their occupancy across the IDC (see Figure 1B). There was a statistically significant link between one histone modification (H4K5ac) and transcription that is skewed towards a negative correlation. Although H4K5ac is predominantly a constitutive histone mark, this observation opens the possibility that this otherwise euchromatic mark may play a role in transcriptional repression in a small group of genes. Four histone modifications (H4K20me3, H4R3me2, H4K12ac and H3K79me3) show essentially no association with transcription during the IDC (Figure 3B). One interesting example is H4K20me3 which was shown to be present at both heterochromatic and euchromatic domains of the P. falciparum genome [12]. In the future, it will be interesting to study their potential roles in chromatin structure and remodeling which may be distinct from transcription.
Next, we were interested in the biological significance of the transcription-linked histone modifications. We analyzed the distribution of mRNA correlating histone-marked loci with respect to their position in the gene and subsequently investigated the functional involvement of these genes (Figure 4). Interestingly only one dynamic, transcription-linked histone mark (H4K8ac) showed a strong presence for the IGRs and/or 5′ untranslated regions (5′UTR). This is in good agreement with its overall distribution in the genome (see Figure 1D). All other modifications associated with transcription, including H3K9ac, H3K4me3, H3K56ac, H3K4me3, H4K16ac and H4ac4, appear to accumulate mainly at the 5′ ends of the ORFs. Transcription-linked occupancy of H4K20me1 and H3K14ac showed no positional preference. We did not observe any positional bias for probes negatively correlated with expression (r≤−0.4).
To assess the functional relationship of transcription-linked histone marks, we identified significantly represented functional groups (P<0.05) based on Gene Ontology (GO), Kyoto Encyclopedia of Genes and Genomes (KEGG) and Malaria Parasite Metabolic Pathway (MPMP) (Figure 4). The transcription-linked histone marks were mainly enriched in genes associated with growth (ribosomal structure and assembly), metabolism (fatty acid metabolism, nucleotide biosynthesis) and host-parasite interactions (Maurer's cleft, invasion). A small subset of genes associated with H4K8ac, H4K16ac and H4ac4 which negatively correlated with expression belonged to molecular motors or genes coding for kinetochore and centrosome organization. Interestingly, transcription profiles of genes involved in DNA replication correlated positively with H4K20me1 but negatively with H4K8ac and H4K16ac occupancy profiles. This implies that the DNA replication genes may be deacetylated and methylated at the onset of DNA replication. These observations are consistent with previous studies in other eukaryotic systems that have shown both co-existence [30] as well as competition [31] between H4K20me1 and H4K16ac, and imply the presence of a similar histone code in P. falciparum. In summary, these results clearly demonstrate the association of transient histone modification states with transcriptional activation where at least eight histone marks either individually or in various combinatorial patterns, have the potential to modulate gene expression during the P. falciparum IDC.
Presently, very little is known about the mechanisms of chromatin remodeling in Plasmodium parasites. Given the highly dynamic character of histone modifications observed by this as well as previous studies [13], [20], [22], these mechanisms are likely to be highly evolutionarily diverse. Transcription factors bound to promoter and upstream regions are known to recruit chromatin modifiers in other species. We therefore investigated the role of promoter regions in the recruitment of H4K8ac that we found mainly in the upstream regions of active genes. In particular, we wanted to assess the presence of any DNA elements which help to establish histone marks in promoter regions. Four promoters (1.5–2 Kb upstream of the ATG) marked by H4K8ac were selected and cloned into luciferase reporter constructs including upstream regions of ring-specific (MAL13P1.122), trophozoite-specific (PF14_0705), schizont-specific (PFD0240c) and sporozoite-specific (PFC0210c) genes. Here, we made use of the strain Dd2attB [32] in which transgenes can be integrated at the cg6 locus (Figure 5A). We found that the occupancy profile of H4K8ac was recapitulated on three of the four ectopic promoters (Figure 5B). These profiles override an existing profile of the endogenous cg6 gene (dashed line) that is normally characterized by high levels in rings and gradually declines through trophozoites and schizonts. The luciferase activity profiles were also similar to the acetylation patterns of all transfected promoters (data not shown). For one of the promoters (PF14_0705), there was an incomplete “carry-over” of the H4K8ac occupancy profile that was matched only in the ring stage. This may be due to unknown factors like insufficient promoter length. Overall our data suggest that the promoter regions of P. falciparum genes carry DNA regulatory elements that establish H4K8ac independently of their endogenous chromatin environment.
The abundance of dynamic temporal regulation of individual histone marks suggests the existence of combinatorial patterns forming a putative “dynamic histone code”. To investigate this possibility, we carried out pair-wise analysis of occupancy profiles with all thirteen histone marks. To this end, we evaluated the concordance of the occupancy profiles using PCC distributions and subsequently the skewness and KS-test P value as described above (Table S2). Figure 6A shows examples of highly positive (S>1), moderately positive (S between 1 and 0) and highly negative PCC distribution (S<0) between the histone marks. Overall our analysis revealed that most of the overlapping marks (present at the same loci) exhibited a high level of correlation between their occupancy profiles. Here it is important to note that the mean size of ChIP DNA product generated by our protocol is 500 bp which corresponds to approximately three nucleosomes positioned in the vicinity of the genetic locus represented by a microarray probe. Hence the correlations in the occupancy profiles represent either a combinatorial histone modification at the same nucleosome or co-occurrence of these at directly adjacent nucleosomes. High correlations of the occupancy profiles are particularly evident for acetylations that exhibited positive correlations at essentially all overlapping loci (Figure 6B). This suggests that, each nucleosome predominantly undergoes only one set of modifications during the IDC, presumably for one purpose (such as transcriptional regulation). This situation contrasted with the lysine methylation profiles that showed loose or no correlations with each other or with acetylations. Interestingly, H4R3me2 exhibited a strong negative correlation with most of the other marks. This methylation is thought to be mediated by PfPRMT1 and might be playing a similar role to H3R2me2a (asymmetrical histone H3 arginine 2 dimethylation) which has been shown to have a mutually exclusive pattern with H3K4me3 in budding yeast [33].
Acetylation clusters (Figure 6B) comprising H4K8ac, H4K16ac, H4ac4, H4K12ac, H3K9ac and H3K56ac were found to be enriched for specific functional groups which define biologically related genes. Hence, the histones within the nucleosomes associated with the genes within these groups appear to be acetylated at most of the (studied) lysine residues during active transcription. Comparing the associations of all thirteen modifications with mRNA, the most represented gene families associated with ribosome structure, protein biosynthesis, tRNA modifications, Maurer's cleft and invasion displayed positive correlations with the acetylation marks (Figure S5). One of the striking examples involves genes of the early transcribed membrane proteins (ETRAMPs) whose mRNA abundance displayed strong correlations to virtually all acetylated histone marks. Interestingly, genes coding for histones themselves showed an anti-correlation between expression and the histone mark occupancy profiles. However, the majority of the gene groups associated with IDC functionalities was mostly in strong positive correlation with the majority of these euchromatic histone marks. An interesting example involves the group of Api-AP2 transcription factors whose mRNA levels are mostly in positive correlation with all euchromatic marks with the exception of H4K16ac which appears to correlate negatively. Hierarchical clustering of members of each group showed that while the majority of genes in the basic cellular and biochemical pathways expressed during the IDC exhibit good correlation between histone marks and transcription, each of the functional groups contains at least a small subset of genes whose mRNA levels are negatively correlated with at least some histone marks. It will be interesting to investigate the implication of these histone marks on transcriptional regulation and thus the functional involvement of these outlying members.
In order to validate the co-occupancy of more than one histone mark on a genomic region, we carried out sequential ChIP at ring stage parasites to identify bivalent domains having two histone marks: H3K56ac and H3K9ac (Figure 6C, Table S3). From individual ChIP results, a total of 2,560 probes common to H3K56ac and H3K9ac yielded ChIP-to-input ratios above 1. Out of 3,318 probes recognized by sequential ChIP, 2,230 (67%, P = 0) overlapped with probes common to the individual ChIPs. As an example, the distribution of ChIP signals on chromosome 4 defined common areas of enrichment between chromatin immunoprecipitated with either H3K56ac or H3K9ac independently, or H3K56ac followed by H3K9ac sequentially (Figure 6C).
The dynamic morphology during the life cycle development is believed to be an evolutionarily unique feature of Plasmodium as well as other eukaryotic parasitic organisms and likely reflects their adaptation to a specific host environment. It is clear now that these morphological switches are underlined by broad transcriptional shifts that, in the case of Plasmodium species, affect essentially the entire genome [4], [34], [35]. There is mounting evidence that epigenetic mechanisms contribute to the regulation of gene expression across the IDC by maintaining hetero- and euchromatin domains within the Plasmodium genome [12], [14], [22], [36], [37]. Although most previous studies focused on heterochromatic domains, here we provide a comprehensive epigenetic atlas of thirteen predominantly euchromatic histone marks and provide evidence suggestive of their unique functions during the parasite's asexual blood stage development. Our data suggests that at least 8 modifications are linked with the transcriptional activity of up to 76% of P. falciparum genes.
The dynamic nature of histone modifications and their link with transcription during the P. falciparum IDC has been previously suggested. Using ChIP-on-chip, it was shown that H3K9ac and H3K4me3 exhibit highly dynamic patterns of chromosomal distribution between rings and schizonts, and that their occupancy correlates positively with transcription [13]. In the follow-up study by the same group, ChIP-seq results showed that both of these histone modifications associate mainly with promoter regions but only H3K9ac is correlated with transcription while H3K4me3 appears uncoupled from transcription [22]. In agreement with these two reports, our results show a strong accumulation of both H3K9ac and H3K4me3 in the IGRs (Figure 1D). However, for both of these modifications, it is mainly their association with the 5′ end of coding regions that correlates with expression (∼22% and ∼28% of H3K9ac and H3K4me3 modifications associated with the 5′ ends of ORFs are positively correlated with transcript levels, respectively) (Figure 4). On the other hand, only ∼12% and ∼10% of H3K9ac and H3K4me3 within IGRs correlate with transcript levels, respectively. This further highlights previous findings showing that H3K9ac and its association with IGRs plays a greater role in transcription compared to H3K4me3. In addition to these two modifications, we demonstrated that at least six other euchromatic marks of H3 (K14ac and K56ac), and H4 (K8ac, K16ac, K20me1 and ac4) play roles in transcriptional regulation during the P. falciparum IDC. With the exception of H4K8ac which shows maximum enrichment at IGRs and/or 5′UTRs, the five others show maximum enrichment at the 5′ ends of ORFs of transcriptionally active genes. This is surprising as in other eukaryotes such as Toxoplasma gondii, Caenorhabditis elegans and Saccharomyces cerevisiae, most acetylations and methylations that are positively correlated with expression typically localize at the promoters and transcriptional start sites [24], [38]. From this perspective, the chromatin structure of Plasmodium resembles that of plants where most of the euchromatic histone marks accumulate within the start of ORFs as compared to IGRs [26], [27]. In the future, it will be interesting to study these features of epigenetic regulation, possibly in combination with another plant-like phenomenon in Plasmodium, the Api-AP2 transcription factors [5]. On the other hand, the striking shift in the accumulation of H4K8ac towards upstream regions is indicative of a distinct role in transcription that is more related to that of other eukaryotes such as mammals [39].
In eukaryotic organisms, distinct chromatin states are defined by multiple histone modifications acting sequentially and/or in combination, a phenomenon referred to as the histone code. Although a full understanding of the histone code is pending, distinct patterns of modified histones define groups of biologically related genes [29], [39], [40]. Here we show a good concordance of many transcription-linked histone marks suggesting a combinatorial effect in the regulation of gene expression during the P. falciparum IDC. Interestingly, these euchromatic marks (individually or in combination) associate with distinct functionalities that could be broadly divided into two main biological categories: (i) growth (e.g. protein biosynthesis, nucleotide metabolism, and DNA replication), and (ii) host parasite interaction (e.g. merozoite invasion, Maurer's cleft proteins) (Figure S5). This may reflect a regulatory link between the two most crucial functions determining parasite virulence during infection: multiplication rate and interaction with the host immune system.
Given the essential role of epigenetic regulation in gene expression, unique factors associated with these processes are presently considered as drug targets for malaria as well as other human parasitic diseases [41], [42]. One such factor is histone deacetylase (HDAC) which plays a pivotal role in chromatin remodeling and thus transcriptional activity. In P. falciparum, the HDAC inhibitor apicidin causes a massive hyperacetylation of H3K9 and H4K8 (and demethylation of H3K4) residues leading to global deregulation of the IDC transcriptional cascade [20]. This deregulation can be induced by at least three other HDAC inhibitors including a 2-aminosuberic acid derivative, Trichostatin A and SAHA, the latter being currently approved for cancer therapy [43]. Moreover, these inhibitors are able to effectively inhibit Plasmodium HDACs [44] and oral administration of apicidin at 2–20 mg/kg for 3 days cures P. berghei infection in mice [45]. This suggests that inhibition of epigenetic mechanisms in Plasmodium represents a promising target area for malaria drug development, and more efforts in developing new compounds with higher selectivity as well as bioavailability are ongoing [46]. However, the development of new antimalaria “epi-drugs” may not be restricted to HDACs, but could also target their opposing histone acetyl transferases or other factors such as chromatin remodeling complexes and signaling pathways impinging on these processes. Our results with the histone modification landscape in the most pathogenic malaria parasite, P. falciparum, will provide a solid reference for all epi-drug development in malaria as well as other parasitic diseases in the future.
Highly synchronized cells of P. falciparum strain T996 were cultured at 5% parasitemia and 2% hematocrit under standard conditions [47]. For ChIP, saponin-lysed parasites were cross-linked with 0.5% formaldehyde and harvested at 8, 16, 24, 32, 40 and 48 hpi. Samples were also collected for RNA and protein isolation from the same time points.
Equal amounts of total protein lysate obtained from parasite pellets from the 6 time points were separated by 12% SDS PAGE and transferred onto nitrocellulose membrane. Western hybridizations were carried out using antibodies (Millipore, Upstate) directed against the modified histones. Horseradish peroxidase conjugated secondary antibody was purchased from GE Healthcare. We also performed immuno-localization with these antibodies as described [48].
Cross-linked cells were homogenized with 200 strokes of a dounce homogenizer and lysed using 1% SDS. The resulting nuclear extract was sonicated with 8 bursts of 10 sec with 50 sec rest between bursts to shear DNA to a final length of 200 to 1000 bp. The sonicate was then centrifuged for 10 min at 13,000× g, and sheared DNA incubated with the immunoprecipitating antibody overnight at 4°C followed by incubation with salmon sperm DNA/Protein A agarose slurry (Millipore) for 1 h at 4°C. Protein A agarose was gently pelleted followed by extensive washes. The DNA bound to protein of interest was reverse cross-linked using 0.2 M NaCl and incubation overnight at 65°C. Recovered DNA was purified using the QIAEX II kit (QIAGEN). Amplification of immunoprecipitated DNA as well as sonicated genomic DNA (input) was carried out as described [49] with a few modifications [50]. Equal amount of Cy5-labeled amplified ChIP DNA was hybridized to Cy3-labeled amplified input DNA. For sequential ChIP, the eluted complex from the first ChIP was subjected to immunoprecipitation using second antibody as described [51]. During the second round of immunoprecipitation, no antibody control was included.
RNA was isolated to carry out transcriptional profiling at the appropriate time points. RNA extraction and cDNA synthesis were carried out as described [52]. Cy5-labeled cDNA was hybridized against a Cy3-labeled reference pool which was made by combining equal amounts of RNA from each time point.
Equal amounts of Cy5 and Cy3 labeled samples were hybridized to P. falciparum microarrays containing 5,402 50-mer intergenic oligonucleotide probes and 10,416 70-mer ORF probes representing 5,343 coding genes [53]. The intergenic regions were represented by one highly specific probe (up to 1.5 kb upstream of the start codon) whose microarray hybridization parameters were matched to the intragenic probe set using the OligoRankPick algorithm [53]. Using PlasmoDB version 8.2, we were able to remap 14,773 probes to the P. falciparum genome providing an even coverage with at least one probe per 1.542 kb. P. falciparum strain T996 was chosen to carry out these experiments due to an exact IDC length of 48 h and the ease with which it can be synchronized. Since the probes on the array have been designed for 3D7, we excluded vars, rifins and stevors from our analysis. The microarray hybridization was carried out at 63.5°C or 65°C in the automated hybridization station (MAUI, USA) for ChIP DNA or cDNA, respectively, as described [4]. The microarrays were scanned using the GenePix scanner 4000B and GenePix pro 6.0 software (Axon Laboratory).
Lowess normalized data was processed to filter out spots with signal intensity less than twice the background intensity for both Cy5 and Cy3 fluorescence. The relative occupancy of histone marks is represented by log2 ChIP/input ratios where high and low ratios represent strong and weak enrichment respectively of modified histones. For expression analysis, each gene profile was represented by an average expression value calculated as an average of all probes representing a particular gene. For ChIP-on-chip, all microarrays were done in triplicates. For each time point, probes with data present in at least 2 out of 3 triplicates were included. Data was presented as an average of triplicates after Kth nearest neighbor (KNN) imputation. Data was further filtered to include only those probes where signal from ChIP DNA was obtained in at least 2 consecutive time points. To address dynamics across the life cycle, the average ChIP/input ratio of each time point was used to detect the summit and bottom time point of enrichment for every probe. P value was assigned based on a student's t-test between the replicates. Significantly oscillated occupancy profiles (relative occupancy defined by log2 ChIP/input ratio) between the summit and bottom time points were defined as P value<0.05 and fold change ≥1.5 across the IDC referring to all detectable dynamic probes (within the limit of the applied ChIP-on-chip technology) and probes showing the highest level of change in marked histone occupancy, respectively.
To assess overall ChIP enrichment at every time point, probes were divided into 3 groups based on expression of the respective genes at each time point: top, middle and bottom 10% of total probes with highest, intermediate and lowest levels of expression, respectively. For each group, ChIP/input log2 ratios were plotted against probe position from −1000 bp to +3000 bp with respect to ATG at every time point.
Phaseograms for expression and ChIP data were generated by fast Fourier transform method where probes/genes were sorted according to phase from −π to π with the mean-centered log2 ratios across all the time points.
Pearson's Correlation coefficient (PCC or r) was calculated between the histone mark profiles and corresponding mRNA profiles across the IDC and skewness of correlation distribution was calculated. Negative skew values indicate a long tail on the negative side (higher frequency on positive side). For ease of understanding, skew is represented as the negative of skew throughout the manuscript such that positive skew means a higher frequency on the positive side. To test how statistically significant the histone levels correlate with expression levels at each probe, we randomly generated marked histone and expression profile pairs by randomizing profiles between probes 100 times. The two-sample Kolmogorov-Smirnov (KS) test was used to test whether our observed correlation distributions are different from the random ones (P value<0.0005). The same analysis was also performed for correlations between histone modifications. Dice's coefficient was calculated in order to assess the overlap between any two histone modification profiles across the IDC.
ChIP enrichment with respect to position in the gene was calculated using probes with oscillated profiles (P<0.05 and fold change ≥1.5 across the IDC). For each histone mark, the number of ChIP-enriched probes showing positive (r≥0.4) and negative (r≤−0.4) correlation with expression profiles of corresponding genes were normalized to total probes in the input. Data were arranged into bins ranging from −1500 bp upstream of ATG to +3000 bp into the gene and plotted against the percentage of probes falling in each bin.
RTQ-PCR was carried out on immunoprecipitated and input DNA using the SYBR Green PCR Master Mix (Roche) according to manufacturer's instructions. ChIP enrichment was calculated by using the ΔCt method (Ct of immunoprecipitated target gene - Ct of input target gene) where Ct is the threshold cycle. All PCR reactions were done in duplicates or triplicates.
Vector pLN-ENR-GFP and P. falciparum strain Dd2attB were provided by D. Fidock. The GFP cassette from the vector pLN-ENR-GFP [32] was replaced by the firefly luciferase gene and hsp 86 3′ UTR from the plasmid pPF86 [54] at the Bam HI/Apa I sites to create the plasmid pLN-Luc. All constructs were confirmed by sequencing. For transfection, promoter regions of various genes were cloned upstream of the luciferase gene at the Bam HI/Sph 1 sties of pLN-Luc. Transfections of P. falciparum strain Dd2attB and subsequent drug selection were carried out as described [32]. Vehicle vector lacking the luciferase gene was used as a negative control and all transfectants were checked for firefly luciferase activity 48 h post infection using a reporter assay from Promega. Plasmid integration at cg6 locus was confirmed by PCR.
All primer sequences used in the current study are listed in Table S4.
The microarray data have been submitted to NCBI GEO with accession number GSE39238.
|
10.1371/journal.pgen.1001226 | LaeA Control of Velvet Family Regulatory Proteins for Light-Dependent Development and Fungal Cell-Type Specificity | VeA is the founding member of the velvet superfamily of fungal regulatory proteins. This protein is involved in light response and coordinates sexual reproduction and secondary metabolism in Aspergillus nidulans. In the dark, VeA bridges VelB and LaeA to form the VelB-VeA-LaeA (velvet) complex. The VeA-like protein VelB is another developmental regulator, and LaeA has been known as global regulator of secondary metabolism. In this study, we show that VelB forms a second light-regulated developmental complex together with VosA, another member of the velvet family, which represses asexual development. LaeA plays a key role, not only in secondary metabolism, but also in directing formation of the VelB-VosA and VelB-VeA-LaeA complexes. LaeA controls VeA modification and protein levels and possesses additional developmental functions. The laeA null mutant results in constitutive sexual differentiation, indicating that LaeA plays a pivotal role in inhibiting sexual development in response to light. Moreover, the absence of LaeA results in the formation of significantly smaller fruiting bodies. This is due to the lack of a specific globose cell type (Hülle cells), which nurse the young fruiting body during development. This suggests that LaeA controls Hülle cells. In summary, LaeA plays a dynamic role in fungal morphological and chemical development, and it controls expression, interactions, and modification of the velvet regulators.
| Numerous fungi have the potential to infect immunocompromised patients or to contaminate and spoil our nutrients. They represent an increasing danger that threatens public health and agriculture. This requires improved understanding of fungal growth, development, dissemination of spores, and mycotoxin production. We have discovered two related fungal specific protein complexes that provide a molecular link among spore formation, fungal development, and secondary metabolite production. The subunit allocation of both complexes depends on each other, and they share a common subunit. These complexes comprise three related and in fungi conserved proteins of the velvet family that function in concert with a known regulator of secondary metabolism, LaeA. This protein controls the formation of both complexes but is only a part of the trimeric complex. We found that this regulator of secondary metabolism also possesses several developmental control functions in gene expression. These protein complexes discovered in the fungal model system Aspergillus nidulans are conserved in fungal pathogens where they might provide novel insights for understanding growth, development, and interaction with their respective hosts.
| Multicellular organisms have developed a variety of different cell types, which become apparent during the ontogenesis of an organism to its adult form. Cell differentiation requires the coordinated interplay of key regulators, which respond to internal and external cues. Cell type specificity often requires specific physiology and metabolism to allow the formation of tissues and organs exhibiting various functions for the organism. Early cells are often omnipotent or pluripotent and lose potential during differentiation except for those misregulated or uncontrolled for cell-differentiation, which might result in tumorogenesis or cancer [1].
Higher fungi produce a limited number of specialized cells and serve as simple and easily tractable models to study cell differentiation. Filamentous fungi grow by forming polar hyphae where similar cellular units are reiterated. The tip as well as branch points of the filamentous hyphae show increased cellular activity. Highly specialized cells include the ubiquitous asexual or sexual spores that are often dispersed into the air for propagation, and specialized cells that are required to form spores. Especially, sexual spore formation can require complicated fruiting bodies consisting of additional specialized cells that form various tissues [2]–[4]. Furthermore fungal differentiation is coupled to the production of various secondary metabolites including mycotoxins and antibiotics [5], which are assumed to provide a chemical shield against competitors [6].
The model fungus Aspergillus nidulans grows vegetatively as a filament with two developmental options: it can either enter the asexual or the sexual developmental pathway (Figure 1A). Sexual development produces closed spherical fruiting bodies (cleistothecia) where meiotic sexual spores are generated. The maturing fruiting body is embedded in a tissue of globose Hülle cells that are proposed to provide protection and nourishment [7]. The molecular mechanism triggering the developmental switch from a vegetative to globose fungal cell is presently unknown [4], [8].
Formation of sexual fruiting bodies and production of certain secondary metabolites occur preferentially in darkness in A. nidulans and are coordinately inhibited by light as an external signal [9], [10]. In contrast, formation of the asexual spores is promoted by light. Light is perceived by various receptors [11] including the red light receptor FphA [12], the blue light receptors LreA-LreB [13] or the blue-UVA receptor CryA [14]. The molecular mechanism of light signal transduction is yet unknown as well as the exact function of the conserved VeA (velvet A) protein, which is the founding member of the velvet family [9]. CryA controls the levels of the VeA mRNA [14], whereas FphA, LreB and LreA act through physical interaction with VeA by a yet unknown molecular mechanism [12], [13]. Strains lacking veA fail to produce cleistothecia and undergo asexual sporulation under both light and dark conditions.
VeA is a part of the heterotrimeric velvet complex [9], which is assembled in the nucleus in darkness and contains the VeA-related developmental regulator VelB (velvet-like B) and LaeA, the global regulator of secondary metabolism [15]. All three proteins are conserved in various fungi [16]–[19]. VelB interacts with the N-terminus of VeA, whereas LaeA interacts with the C-terminus of VeA. Illumination reduces the cellular amounts of VeA [9]. VelB and LaeA are unable to interact with each other and need VeA as a bridging factor. In addition, VeA supports the transport of VelB into the nucleus, whereas nuclear localization of LaeA does not depend on the other subunits of the velvet complex. This suggests that the complex fulfills its function in coordinating sexual development and secondary metabolism in darkness primarily by controlling gene expression in the nucleus [9].
In this study, we show that the coordination of development and secondary metabolism is only one function of the velvet complex subunits. VelB is a part of a second novel light-regulated complex, which includes VosA (viability of spores A). VeA, VelB and VosA are related members of the fungus-specific novel velvet family regulatory proteins [16]. The VelB-VosA complex can repress asexual development and is essential for asexual as well as sexual spore maturation and trehalose biogenesis. Moreover, besides being a global regulator of secondary metabolism, LaeA executes three important novel developmental functions: (i) LaeA controls the VelB complex allocation between VosA-VelB and VeA-VelB. (ii) LaeA is required for the transition from filamentous cells to globular Hülle cells, and (iii) LaeA is a key factor in light control of fungal development.
Functionally tagged versions of all three proteins of the velvet complex VelB-VeA-LaeA are able to recruit the respective other subunits from a fungal protein extract. In addition, the phenotypes of the corresponding velB or veA deletion strains are similar: both mutants are unable to perform sexual development and are impaired in light control and secondary metabolism [9]. However, only a tagged VelB, but neither VeA nor LaeA, is able to recruit another related protein, VosA [9]. VosA was isolated as a high copy repressor of asexual development and is also required for spore maturation, trehalose biogenesis and long-term viability of asexual and sexual spores [16]. We analyzed whether VelB has an additional yet unexplored function in fungal development.
We initially examined whether VosA is the fourth subunit of the velvet complex during the establishment of developmental competence. Developmental competence describes the phenomenon that A. nidulans spores require at least 20 hours of growth after germination to respond to external signals when placed on the surface of a medium [20]. A. nidulans strain expressing a functional vosA::ctap fusion driven by its native promoter was cultivated in liquid medium and induced on the surface of solid medium for asexual or sexual development by incubation in light and dark, respectively. Purification of VosA::cTAP was performed from 12 hours post- induction cultures on surface of solid medium after developmental competence was achieved. Tagged VosA was only present in the dark and co-purified exclusively with the VelB protein, but neither with VeA nor LaeA (Figure 1B and Table S4). VelB is not only a part of the VelB-VeA-LaeA velvet complex, but also a part of the second complex VelB-VosA when developmental competence is established.
Heterologous expression of VelB in Escherichia coli resulted primarily in dimers suggesting that VelB is able to form homodimers (data not shown) in addition to the VelB-VosA heterodimer. We employed a split-YFP system to determine the in vivo compartment where the subunits of the VelB-VosA heterodimer or of the VelB-VelB homodimer interact. An mRFP histone fusion served as control to track the nuclei within the hyphae. The VosA-VelB YFP signal colocalized predominantly to the nuclear RFP signal, indicating that the VosA-VelB complex is formed in the nucleus (Figure 1C). In contrast, we found the combined signal of N-YFP::VelB and C-YFP::VelB in vivo in the cytoplasm as well as in the nucleus (Figure 1D).
These data suggest that VelB is not only a component of the nuclear VelB-VeA-LaeA complex, but can also (i) form a VelB homodimer in the cytoplasm as well as in the nucleus, and (ii) be part of the nuclear VosA-VelB heterocomplex, which is hardly detectable in the cytoplasm.
VosA is not only a high-copy repressor of asexual development but also plays an essential role in the maturation and viability of spores primarily by coupling trehalose biogenesis and sporogenesis [16]. We analysed whether VelB plays a similar role, as it forms the nuclear VelB-VosA heterodimeric complex. The viability of spores, trehalose biosynthesis and tolerance against various stresses were compared between the velBΔ, wild type, and veAΔ or vosAΔ strains (Figure 2A). The conidia of both velBΔ and vosAΔ strains displayed severe viability defects, whereas viability of the veAΔ conidia was similar to that of wild type, indicating that VelB and VosA play a specific role in conferring spore viability. VelB is needed for the proper biogenesis of trehalose in conidia, because trehalose was undetectable in the velBΔ and vosAΔ conidia (Figure 2B). The mRNA levels of two genes (tpsA and orlA) associated with trehalose synthesis [21], [22] revealed that the velBΔ and vosAΔ strains both exhibited reduced tpsA and orlA transcript levels during the late phase of development and in conidia (Figure 2C). These results indicate that both VelB and VosA are necessary for trehalose biogenesis and viability of spores.
As trehalose plays an important protective role in response to various stresses, we tested whether the absence of velB would result in decreased tolerance of the spores against various stresses, and examined two-day old conidia of wild type, veAΔ, velBΔ, and vosAΔ strains. Serially diluted spores were cultivated on solid medium containing various H2O2 concentrations. The velBΔ conidia were the most sensitive among those tested (Figure 2D). At 0.25 M H2O2, 90% of the velBΔ conidia were non-viable, whereas only about 40% of wild type and the veAΔ conidia lost viability. After being treated with 0.5 M H2O2, most of the velBΔ and vosAΔ conidia were non-viable, whereas about 60% and 50% of wild type and the veAΔ conidia, respectively, were viable (Figure 2D). These data were further confirmed by testing the tolerance against UV, where both the vosAΔ and velBΔ conidia were more sensitive than those of wild type. Being exposed to 100 J/m2 UV only about 30% of the velBΔ and vosAΔ conidia were viable, whereas 80% of wild type conidia could survive. The veAΔ conidia were also more sensitive compared to wild type (Figure 2E). While the velBΔ and vosAΔ conidia were more sensitive to thermal stress than wild type, the mutant and wild type conidia were equally tolerant to high osmolarity (data not shown). These data indicate that both VelB and VosA are required for trehalose biogenesis in spores, thereby conferring the viability and stress tolerance of spores. The VelB-VosA heterodimer might be the functional unit for these critical biological processes.
The finding that both heteromeric complexes are located in the nucleus suggested that there might be a competition for VelB between the nuclear VelB-VeA-LaeA velvet complex and the nuclear VosA-VelB complex. VelB and VosA protein levels were monitored using functional TAP-fusions and the α-calmodulin antibody to address the developmental time window during which both subunits are expressed simultaneously and the VelB-VosA complex can be formed. In wild type cells VelB and VosA are present abundantly during vegetative cultivation in submerged cultures but upon transfer to solid medium in the light both proteins became undetectable. In the dark both proteins were present at the beginning of sexual development (12 h sexual) and then undetectable during later stages of development (Left panels, Figure 3A and 3B). This suggests a potential role of the VosA-VelB complex during vegetative growth and at the beginning of sexual development in the dark when the velvet complex VelB-VeA-LaeA is also present. Simultaneous overexpression of VelB and VosA under an inducible promoter resulted in repression of asexual development, which further supports a common role of both proteins (Figure S1).
We analysed whether the VosA and VelB protein levels depend on VeA or LaeA. Expression analysis in a veAΔ strain did not result in significant changes of the VelB or VosA protein levels in comparison to wild type (data not shown). However, in a laeAΔ strain, both VosA and VelB were still present after 12 hours incubation in the light. Moreover they also appear during mid sexual stage (24, 48 h sex) (Right panels, Figure 3A and 3B). We performed VosA-TAP purification using a laeAΔ strain to determine whether the absence of LaeA also resulted in formation of the VelB-VosA complex in fungal extracts (Figure 3C). TAP purification of VosA from cultures grown in either the light or the dark in the absence of LaeA demonstrated that the VosA-VelB association occured predominantly in the light (Table S5), which is contrary to wild type where we only found the complex in the dark (Figure 1B). Formation of the VosA-VelB nuclear complex in the light in a laeAΔ strain was further corroborated by BiFC (Figure 3D). velB::ctap and vosA::ctap mRNA levels in wild type and laeAΔ did not correlate with the protein levels (Figure S2). These results suggest that there is a posttranslational control for the VosA-VelB proteins and LaeA plays a key role in light-dependent control of the VosA and VelB protein levels.
We monitored the cellular levels of the VeA protein during development to explore whether the protein levels of all three members of the velvet family are controlled by LaeA. While it was previously reported that veA expression is upregulated in the laeAΔ [9], the VeA protein levels have not been analyzed.
α-VeA antibodies revealed that the cellular levels of the native 63 kDa VeA protein were comparable in wild type and the laeAΔ strain in crude cell extracts (Figure 4A and 4B). In addition, a small subpopulation of a VeA isoform of a higher molecular weight (72 kDa) could be detected in wild type cultures during vegetative growth or sexual development in the dark. During the light-mediated asexual development this isoform was hardly detectable. The VeA antibody specifically recognized VeA-63 kDa as well as VeA-72 kDa, because neither bands were present in a veAΔ strain (Figure S3A).
This VeA-72 kDa isoform accumulated to higher levels than VeA-63 kDa in the laeAΔ strain in vegetative growth and early development with or without light. The total amount of the VeA protein in the absence of LaeA is therefore significantly higher in comparison to wild type. This suggests that LaeA inhibits the overall protein levels of all three members of the velvet family members and specifically inhibits the formation of the 72 kDa VeA isoform.
VeA1 is a peculiar light-insensitive mutant variant of the VeA protein. The veA1 mutant produces significantly reduced levels of sexual fruiting bodies and constantly high amounts of asexual spores in the dark as well as in the light [23]. The veA1 mutant phenotype develops by an unknown mechanism and depends on the truncation of the first 36 N-terminal amino acids in comparison to the full-length VeA [24]. This shortened VeA1 mutant protein exhibits reduced protein interaction with VelB and decreased nuclear import of both proteins [9], [25]. In contrast to wild type, the veA1 mutant did not accumulate VeA-72 kDa (Figure 4B) suggesting that this LaeA dependent molecular shift correlates with light regulation and depends on an intact N-terminal part of VeA. In the presence of VeA1, actin levels decreased presumably due to the increased asexual conidiation (Light 12 and 24), (Figure 4B).
In the absence of LaeA we analyzed complex formation of VeA in the light, when the modified VeA-72 kDa, VosA and VelB proteins accumulated. A VeA::cTAP laeAΔ strain was shifted from vegetative liquid growth to solid medium in the light or in the dark for 12 hours to achieve developmental competence. We detected high levels of the VelB-VeA dimer associated with the α-importin KapA under both conditions (Figure 4C and Table S6). The reciprocal experiment using VelB::cTAP recruited VeA and KapA, in addition to VosA. These proteins all co-purified with VelB in the dark as well as in the light (Figure 4C and Table S7). However, VeA::cTAP in wild type recruits these proteins only in the dark, but fails to recruit VosA and only small amounts of VelB in the light [9]. BIFC localization studies revealed that the VeA-VelB interactions in the laeAΔ background took place in nuclei of fungal hyphae both in the light and the dark (Figure 4E).
The data suggest that LaeA not only controls the amounts of VosA, VelB and VeA in the light, but also prevents the shift of VeA to the 72 kDa isoform, which presumably represents a post-translational modification. This LaeA controlled VeA modification does not impair the transport of VeA-VelB into the nucleus assisted by the importin KapA. The finding that the importin KapA was only recruited together with VeA(-TAP)-VelB but not with VosA(-TAP)-VelB supports our earlier finding that VelB is preferentially transported into the nucleus together with VeA [9].
LaeA has been identified as a global regulator of secondary metabolism [15] in light-insensitive veA1 laboratory strains [24]. The veA1 allele represents an artificial situation that could be misleading for the understanding of the molecular function of VeA. Therefore we analyzed the laeA deletion mutant in the veA wild type background, which revealed distinct differences in colony morphology for veA+ and veA1. The laeAΔ veA+ colony is white, whereas laeAΔ veA1 exhibits the typical green color of wild type colonies, which is due to the pigmentation of the asexual spores (Figure 5A). All analyzed laeAΔ strains irrespective of the veA allele were unable to produce the mycotoxin sterigmatocystin (ST) underlining the well-known LaeA function as a global regulator of secondary metabolism (Figure 5B).
Microscopic examination revealed two major differences between the laeAΔ veA+ strain and the other strains. Wild type as well as laeAΔ veA1 strain produced higher number of conidiophores bearing the asexual spores (conidia) than laeAΔ veA+ strain in the light and dark. Quantification of the conidia indicated that conidia production in laeAΔ in the veA+ background was significantly decreased in the light to approximately 20% of the wild type and asexual development was unresponsive to illumination (Figure 5A). This suggests that there is a yet unexplored LaeA control for asexual spore formation, which only works in combination with an intact VeA N-terminus.
In addition to a reduced number of conidia, the whitish appearance of laeAΔ colonies originated from significantly elevated levels of sexual structures both in the dark and light (Figure 5A). Wild type veA+ strain generated few cleistothecia (seen as black or white round structures) and many conidiophore heads (green structures) in the light, but more cleistothecia and less conidiophores in the dark. The veA1 strain produced only few cleistothecia in the dark, therefore formed predominantly conidia under both light and dark conditions (Figure 5A).
The unresponsiveness of the laeAΔ strain to the white light does not depend on specific light receptors. We determined photon fluence-rate response curves for the photoinhibition of fruiting body formation under near UV for CryA, blue light spectra for LreA-LreB, and red-light spectra for FphA [13], [14]. Wild type strain reduced cleistothecia formation with increasing photo dosage to below 20%. In contrast, the photoinhibition in the laeA mutant was lost under all irradiation conditions (UVA 366 nm, blue 460 nm, red 680 nm) (Figure 6). The lack of photoinhibition caused by a loss of LaeA was regardless of high or low light intensity, suggesting that laeAΔ strains are entirely blind and LaeA is required for light mediated inhibition of cleistothecia formation of all three known light qualities.
The functional relationship between laeA and veA was examined by creating the laeAΔ veAΔ double mutant. The double mutant exclusively manifested the veAΔ phenotype characterized by only asexual development. Thus, the veA mutation is epistatic to laeAΔ and sexual development of laeA mutants depends on VeA (Figure 5A). These results demonstrate that LaeA has an additional developmental role besides being a major regulator of secondary metabolism and is an essential part of the light-dependent control mechanism of fungal development. Double mutant strains of laeAΔ with fphAΔ, lreAΔ, lreBΔ or cryAΔ representing photoreceptor genes always resulted in an epistatic laeAΔ phenotype (data not shown). The LaeA dependency of an intact VeA is essential to promote the asexual developmental program and to inhibit the sexual program of A. nidulans in the light. Truncation of the N-terminus part of VeA, which interacts with VelB, abolishes this LaeA mediated regulation. This suggests that LaeA controls the protein levels of the members of the regulatory velvet family but also the balance between VelB-VeA, VelB-VeA-LaeA or VosA-VelB complexes within the fungal cell.
We compared in more detail the constitutively produced fruiting bodies of laeAΔ veA+ and wild type. This resulted in the discovery of two remarkable phenotypes. Both were verified by complementation of the laeAΔ strain by the laeA wild type allele (Figure 7A). First, the laeAΔ mutant produced more fruiting bodies than wild type but they were significantly smaller in size. Detailed inspection with scanning electron microscope (SEM) unveiled that the wild type fruiting bodies of a diameter of approximately 200 µm were reduced to 40 µm diameter cleistothecia in the laeAΔ strain (Figure 7A). In agreement with their small size, cleistothecia of laeAΔ contained only 20% of the ascospores compared to wild type fruiting bodies (Figure 7A). The small laeAΔ cleistothecia contained meiotically formed viable ascospores which germinated on appropriate medium, indicating that the fertility of ascospores was not affected (data not shown).
Second, wild type cleistothecia are normally covered by spherical Hülle cells forming a tissue which is proposed to nurse the maturing fruiting bodies. In contrast to wild type where cleistothecia were entirely surrounded by hundreds of Hülle cells, the cleistothecia in laeAΔ were in contact with only two to five Hülle cells per cleistothecium (Figure 7A).
We examined the influence of various degrees of LaeA overproduction on fungal development for a more comprehensive picture of the LaeA regulatory function in sexual development. We expressed laeA under the nitrate inducible niiA promoter [26] in the veA+ backgound (Figure 7B). Induction of laeA expression was verified by Northern blot hybridization. The ipnA and stcU genes were used as control because ipnA was previously shown to increase by high levels of LaeA [15] whereas stcU, a gene of the ST gene cluster, was not affected. Increasing degrees of LaeA expression did not disturb light inhibition of sexual development which was functional as in wild type (data not shown). Only high levels of LaeA resulted in a significant developmental phenotype in the dark. This overexpression strain produced twice more cleistothecia than wild type, when the niiA promoter was activated by cultivation on nitrate medium (Figure 7B). This further corroborates a developmental role of LaeA to control cleistothecia, which might be mediated by the Hülle cells.
Hülle cells were analyzed in more detail by monitoring the expression of cell specific genes in the laeAΔ strain. The α-mutanase encoded by mutA is particularly expressed in Hülle cells [27]. A mutA promoter fusion to sgfp (synthetic green fluorescence protein) was constructed in wild type and laeAΔ strains. Whereas wild type showed an sGFP signal during late phases of vegetative growth and development, laeAΔ strain failed to generate detectable sGFP signal (Figure 7C). The GFP fluorescence of 100 Hülle cells for each strain was measured to analyze whether the single Hülle cell of the laeAΔ strain differs from the Hülle cell tissue of wild type. Approximately 35 of the 100 wild type Hülle cells showed a specific sGFP signal originating from the cytoplasm of the Hülle cells (Figure S3C). In contrast, there was hardly any specific sGFP in the Hülle cells of laeAΔ strains except for a weak autofluorescence. Transcript analysis of the mutA gene in wild type and the laeAΔ strains further supported the failure of laeA mutants to express the Hülle cell specific mutA gene. Regardless of the veA+ or veA1 alleles, the mutA mRNA levels were drastically reduced in laeAΔ strains in comparison to wild type (Figure S4).
Our data suggest that LaeA affects VeA on gene expression and on protein levels potentially by inhibiting the modification of the VeA-63 kDa protein. The N-terminally truncated VeA1 protein is impaired in this control and also impaired in the interaction with VelB. Consistently, LaeA also controls the cellular levels of VelB and VosA as further members of the VeA regulatory protein family. This regulatory network is involved in the promotion of asexual spore formation in the light (presumably by releasing the repressor function of VosA-VelB) as well as the light-dependent inhibition of sexual development. In addition, LaeA has functions which do not specifically require the VeA N-terminus but require some VeA activity. These include Hülle cell formation and/or controlling the Hülle-cell specific mutA gene activity (Figure 7) but also secondary metabolism control including aflR expression [15]. These findings predict that there might be more regulatory developmental genes controlled by LaeA either in a VeA N-terminus dependent or independent way.
The screening of transcripts of various fungal developmental regulator genes (Figure S4) revealed that the asexual regulator abaA is one of the genes controlled by the LaeA when VeA N-terminus is intact. abaA encodes a transcription factor which is conserved from filamentous fungi to yeast [28], [29] and which is required for asexual spore formation. abaA expression levels were almost abolished during development of a veA+ laeAΔ strain. The effect seems to be specific because another key regulator of asexual development, brlA [30] was significantly less affected in its expression in the same mutant strains.
Various regulator genes of sexual development exhibited only subtle VeA dependent changes in gene expression during development. The two sexual regulatory genes nosA and steA [31], [32] were exceptions because they were transiently reduced in the veA1 laeA and the veA+ laeA deletion strains during vegetative growth (20 h). This effect is therefore independent of the N-terminus of VeA and seems to be specific, because the mRNA for the GATA type transcription factor NsdD, which is essential for sexual development [33], was not significantly changed in wild type in comparison to both laeA mutant strains. Indeed, overexpression of nosA in laeAΔ moderately rescued the small cleistothecia phenotype (Figure S5).
Our data support that LaeA is required not only for differentiation of asexual spores but also for Hülle cells and their activity. It seems plausible that without LaeA and therefore without Hülle cells the cleistothecia are not nursed properly and can not reach their wild type regular size. These results also indicate that formation of the Hülle cells is not an absolute prerequisite for fruiting body formation. Moreover, our results further support that LaeA is involved in the control of regulatory genes in development and secondary metabolism and this control can be dependent or independent of the VeA N-terminus.
The velvet family regulatory proteins are fungus-specific and highly conserved among ascomycetes and basidiomycetes [16]. Fungi represent one of the largest groups of eukaryotic organisms on earth with an estimated 1.5 million, mostly unknown, species including human and plant pathogens [34]–[38]. The understanding of the molecular mechanisms of the VeA family proteins function might play a key role to understand fungal development. The VeA family includes VeA, VelB, VelC and VosA. VeA, as the first identified light regulator of this family [23], regulates morphological development coupled with secondary metabolism [10], [17]–[19], [39]. VosA is not only able to repress asexual development in A. nidulans, but is also essential to link sporogenesis and trehalose biogenesis [16]. VelB was discovered by its ability to interact with VeA and characterized as a light-dependent developmental regulator [9]. In this study, we also identified the VelB-VosA complex. The appearance of VelB correlates with the VosA protein. VelB and VosA seem to share at least parts of their functions, because overexpression of the dimer represses asexual development and the velBΔ strain exhibits similar reduced survival rates as the vosA deletion. The genetic data suggest that VelB and VosA are inter-dependent in executing trehalose biogenesis, spore maturation and long-term viability. This may be associated with the formation of the nuclear VelB-VosA heterodimeric complex. Therefore VelB has dual functions within asexual as well as sexual development.
The roles of VelB and VosA in spore maturation are similar to those found in other filamentous fungi including A. fumigatus and Histoplasma capsulatum. In H. capsulatum, Ryp2 and Ryp3, are homologs of VosA and VelB, respectively, and play a role in regulation of sporulation and inter-dependent expression of the RYP genes [40]. In A. fumigatus, the deletion of vosA and velB caused ∼50% reduction of the spore trehalose content and viability (Park & Yu, unpublished). Preliminary functional studies of velC in A. nidulans indicate that this fourth member of the velvet family positively functions in sexual development (Park et al, unpublished).
Heteromeric proteins play vital roles in the development of fungi, plants or animals. Fungal examples involved in the development of sex-specific cells include the heterodimeric α2-a1 complex which represses haploid specific gene expression or the α2-MCM1 complex which turns off alpha-specific genes in yeast cells [41]. Combinations of bE (East) and bW (West) heterodimeric complexes promote the switch from the haploid yeast phase to the pathogenic dikaryotic phase of the corn smut fungus Ustilago maydis [42]. Our studies demonstrated that the velvet family proteins form a novel class of fungal regulators that also establish heteromeric complexes and have interdependent functions in determining cell fate.
The VeA-VelB heterodimeric complex of A. nidulans presumably forms in the cytoplasm and serves as the major pathway for the VelB entry into the nucleus. The VeA nuclear transport is controlled during development by the light which increases the cytoplasmic fraction of VeA and reduces the nuclear population [25]. The bipartite nuclear localization signal (NLS) is located at the N-terminus of the VeA protein and is disrupted in VeA1, which is derived from a truncation of 36 amino acids of the N-terminus of VeA. This results in the constitutive but reduced VeA nuclear import with reduced interaction with VelB without being controlled by illumination. Light control of VeA might be activated during development by a direct interaction of VeA to the phytochrome FphA. This light sensor is connected to the white collar homolog proteins LreB and LreA as additional light sensors [13]. CryA, another fungal light sensing system, functions in a distinct way. It does not interact with VeA, but reduces veA mRNA accumulation and therefore reduces the VeA protein levels within the fungal cell during development [14]. Whereas VelB can form homodimers in both cytoplasm and nucleus, VosA-VelB is preferentially located in the nucleus. If VeA provides the major nuclear import pathway for VelB, this suggests that VeA can be exchanged for VosA or another VelB within the nucleus.
The VosA-VelB heterodimer complex appears to have multiple functions. It can repress asexual spore formation and also controls genes associated with trehalose biogenesis for the spore. The VosA-VelB complex may act as a transcription factor as the C-terminal domain of VosA has transcription activation activity and the VosA protein might bind to the promoter regions of various genes [16]. It will be interesting to reveal the genes regulated by the VosA-VelB complexes among filamentous fungi including human or plant pathogens. While our in vivo biochemical studies never identified VelC as an interacting partner of the three velvet regulators, a yeast two hybrid screen followed by GST pull-down assay suggested that VosA and VelC interact and form a heterodimer complex (Ni et al, unpublished data). It appears that velC might be expressed at very low levels under specific environmental or developmental conditions.
LaeA fulfills two distinct yet related functions within the fungal cell. One function includes the control of the amount of velvet family proteins and therefore the potential to form various complexes. We found here a specific regulatory role of LaeA for all three velvet family members. This novel regulatory role of LaeA for fungal development exceeds its previously reported function as a global regulator of secondary metabolism [15].
LaeA controls the amount of VosA and VelB in a light dependent manner. In the light the wild type fungus would normally reduce the VosA-VelB complex to release asexual inhibition and to promote the asexual program. In parallel, the sexual program which also requires VelB is repressed. Without LaeA we find, even in the light, high amounts of VosA and VelB and consistent with the VosA-VelB complex, the asexual program is repressed and the sexual pathway is constitutively activated. It is not yet understood why the truncation of the N-terminus of the VeA1 mutant protein results in constitutively high asexual and low sexual development independent of illumination. Activation of sexual development by excessive amounts of the VelB-VosA dimers even under the light conditions further supports that a major function of the VelB-VosA complex after successful germination of spores is to repress fungal development during vegetative growth.
LaeA does not only control VosA and VelB protein levels but also controls simultaneously VeA protein levels and the formation of different VeA forms. VeA is constitutively expressed during different phases of fungal development and normally represents a 63 kDa protein. An additional higher molecular weight VeA of 72 kDa is inhibited in the light where asexual development is promoted, and is only detectable during vegetative growth or in the dark during sexual development. The increased amounts of VelB and VosA in the absence of laeA somehow correlate with an accumulation of the VeA-72 kDa version. This accumulation can not be observed when the N-terminus is truncated as in the VeA1 mutant protein. The 72 kDa shift from VeA-63 kDa presumably represents a modification which is inhibited by LaeA in a light dependent manner. VeA is known to be a phosphoprotein [13] and phosphatase treatment does not affect VeA-63 kDa or the VeA1 mutant version but resulted in a partial reduction in the mobility of the 72 kDa version (Figure S3B). Furthermore α-phosphoserine and α-phosphothreonine recognized the immunoprecipitated phosphorylated 72 kDa VeA protein in the laeAΔ background supporting that the serine and threonine residues of VeA are phosphorylated (data not shown). However, the LaeA dependent VeA modification is even more complex and includes at least one yet unknown modification. LaeA associates with the VelB-VeA dimer forming the heterotrimeric velvet complex. LaeA might protect VeA from modification by occupying the C-terminus of VeA, and thereby controlling the balance between VosA-VelB and VelB-VeA-LaeA (Figure 8). There might be another level of control that limits the overall VeA protein levels. It will be interesting to analyze whether LaeA is able to interfere with the interaction of VeA to the light receptor complex FphA-LreA-LreB [43] to confer its light control function.
Further LaeA regulatory functions are independent of the N-terminus of VeA. It is tempting to speculate that the N-terminus dependent LaeA functions involve VosA and VelB, whereas the independent functions concern LaeA alone or in concert with VeA and/or VelB. The LaeA-VeA1 complex can at least partially fulfill the LaeA control of secondary metabolism, which has been investigated in veA1 laboratory strains [5], [15].
In a striking contrast to the veA and velB mutants, loss of LaeA does not abolish the potential to form fruiting bodies. We found it remarkable that without LaeA almost no Hülle cells can be formed, and hardly any expression of the Hülle cell specific mutA gene occurs. The function of Hülle cells are proposed to protect and nourish the maturating nests which are the primitive structures of cleistothecia [2]. Consistently to the proposed nursing function, the fungal fruiting bodies of a laeA deficient strain are only one fifth of the normal size. The size of an average cleistothecium is around 200 µm. In literature there are few genes affecting the size of cleistothecia including tryptophan auxotrophic mutants [44], hisB gene deletion [45] as well as sumO mutant. SumO is a small ubiquitin like modifier of A. nidulans [46]. The laeA deletion mutant constitutively produces these high amounts of small cleistothecia, even in the presence of light, further corroborating the key role that LaeA plays in light dependent fungal development.
Another remarkable finding is that the expression of the transcriptional regulatory genes steA [31] and nosA are LaeA dependent during vegetative growth. Both genes are involved in the sexual pathway. Without SteA there are no fruiting bodies [31]. Even more interesting is that nosA mRNA is completely absent in vegetative cells of laeAΔ. Deletion of nosA gene also results in very small cleistothecia which are about 30 µm in size but still contain fertile ascospores [32]. nosAΔ strain has almost no Hülle cells, a phenotype similar to laeAΔ strains. It is therefore likely that LaeA dependent expression of nosA during the vegetative stage is required for Hülle cell formation. This is further supported by the findings that overexpression of nosA under nitrate inducible niiA promoter in laeAΔ partially rescued the lack of Hülle cells, small cleistothecia and ascospore production (Figure S5). This results in abundant expression of NosA in vegetative cells in a laeA deletion (Figure S5D). The reason why the rescue is only partial might be due to the fact that some other regulators acting in the parallel pathway with nosA for Hülle cell formation are still less expressed or misregulated in a laeAΔ. It will be interesting to examine whether and how this LaeA dependent temporal control of transcription factor genes like nosA depends on the members of the velvet family.
The parental generation of multicellular organisms normally has to provide nourishment as well as protection for the next generation. Hülle cells of the mold A. nidulans are associated with cleistothecia and provide this function for the fungal fruiting body. Our major finding here is that LaeA in combination with the velvet family of related regulatory proteins is involved in both lines of support for the next generation. LaeA was first discovered to be the global regulator of secondary metabolite genes including sterigmatocystin, penicillin and many other compounds. All these chemicals might confer a certain advantage to the fungus during growth under substratum in the soil. Aspergillus produces asexual conidiation on the surface of the soil, but sexual development takes place under substratum where numerous eukaryotic or prokaryotic organisms compete for nutrients and represent a threat to vulnerable sexual fruiting bodies. Carcinogenic sterigmatocystin might protect fungal cleistothecia against eukaryotic competitors. Consistently, laeAΔ strains are the preferred food source of insect larvae in comparison to a wild type strain [6].
Similarly, penicillin might help to defend against various bacteria in the soil. All these responses regulated by LaeA might be considered as the chemical protection of fruiting bodies. At the same time, LaeA is essential for the Hülle cells and therefore controls feeding of the fruiting bodies by providing these cells. Thus, LaeA promotes both the production of chemicals to protect fruiting bodies and the production of nourishing cells for developing fruiting bodies.
The LaeA functions exerted on maturating cleistothecia in combination with the heteromeric protein complexes of the velvet family represent an unexpected scenario in fungal development. It will be interesting to see how much convergent evolution there is and whether there are molecular counterparts of LaeA in other higher organisms which are involved in the protective as well as the nutritional function for preparing the next generation for future life.
Strains used in this study are listed in Table S1. Aspergillus nidulans strains; TNO2A3 (nkuAΔ) [47], AGB152 [48], AGB154 [49] served as wild type transformation hosts for the deletion and epitope tagging as well as overexpression experiments. Transformation of the vosA::ctap linear construct into AGB152 yielded AGB509 strain. laeA deletion cassette containing ptrA marker was transformed into TNO2A3 generating laeAΔ/veA1 (AGB468) which was then crossed with AGB154. This crossing gave rise to prototrophic deletion strains laeAΔ/veA1 (AGB512) and laeAΔ/veA+ (AGB493), respectively. AGB493 and AGB509 strains were crossed in order to obtain vosA::ctap, laeAΔ/veA+ combination (AGB510). The velB::ctap, laeAΔ/veA+ hybrid (AGB511) was created by crossing AGB493 with AGB389 strain. The presence of wild type veA+ allele was verified by analytical PCR of the locus followed by BstXI digestion. laeA deletion as well as vosA- and velB-tap loci were confirmed by Southern blot (Figure S6). AGB513 strain that contains veA::ctap in laeAΔ strain was created by introducing pME3711 into AGB512. pmutA::sgfp reporter plasmid, pME3296, was introduced into AGB152 (wt) and laeAΔ (AGB493) strains yielding AGB514 and AGB515, respectively. The BIFC plasmids, pME3714 (nyfp::velB/cyfp::vosA), pME3715 (cyfp::velB/nyfp::vosA), and pME3717 (nyfp::velB/cyfp::velB) were introduced into the recipient strain AGB506 yielding AGB516 (velB-vosA), AGB517 (vosA-velB), and AGB543 (velB-velB) BIFC strains, respectively. pME3715 was transformed into laeAΔ (AGB468), resulting in AGB544 (velB-vosA, laeAΔ). nosA OE construct (pME3719) was placed in AGB493, which led to AGB545. Integration of the plasmids into the genome was confirmed by diagnostic PCR. DH5α and MACH-1 (Invitrogen) Escherichia coli strains were applied for recombinant plasmid DNA. Aspergillus and E. coli strains were cultured as described previously [14].
Tranformation of E. coli and A. nidulans was performed as explained in detail [50], [51].
During processing and construction of linear and circular DNAs, standard recombinant DNA technology protocols were followed as given in detail [52]. Plasmids and oligonucleotides (Invitrogen) employed in the course of this study are listed in Table S2 and Table S3, respectively. PCR reactions [53] were performed with various DNA polymerase combinations including Pfu (MBI Fermentas), Phusion (Finnzymes), Platinum-Taq (Invitrogen) and Taq polymerases.
In order to create laeA deletion construct 5′ UTR region of laeA was amplified from the wild type genomic DNA with primers OSB22/24 and 3′ UTR region was amplified with OSB25/27. The two amplicons were fused to the ptrA marker (from pPTRII) with fusion PCR [47] (nested oligos OSB23/26) yielding 4324 bp linear deletion construct which was used to transform TNO2A3 to AGB468. For complementation of laeAΔ, the laeA genomic locus (3.7 kb), containing 1.5 kb promoter and 1 kb terminator regions, was amplified from genomic DNA (OSB22/27) and cloned into the StuI site of pAN8-1 (phleoR) which yielded pME3635. Then pME3635 was introduced into laeAΔ strains, (veA+, AGB493) and (veA1, AGB512), resulting in AGB494 and AGB518, respectively. In order to overexpress laeA gene, laeA cDNA was amplified from cDNA library (OZG61/62) and inserted into the PmeI site (pME3718) under nitrogen source regulable niiA promoter, generating pME3716. This plasmid was eventually introduced into AGB152, which resulted in AGB519.
To replace the vosA locus with vosA::ctap, vosA ORF including 1 kb of the vosA promoter (oligos VosA–A/C) and 1 kb vosA terminator (VosA–D/F) were amplified from genomic DNA and the resulting amplicons were fused to the ctap::natR module via fusion PCR (VosA–B/E). Gene replacement cassette was introduced into AGB152 and the substitution of the vosA locus by vosA::ctap was verified by Southern blot hybridization (Figure S6).
velB cDNA was amplified (OZG397/64 for n-yfp, OZG63/64 for c-yfp fusion) from sexual cDNA library. Then n- (OZG73/387) and c-yfp (OZG75/77) amplicons were fused to velB cDNAs with oligos OZG397/64 (n-yfp::velB) and OZG63/64 (c-yfp::velB), respectively. n-yfp::velB and c-yfp::velB were cloned into the PmeI site of pME3160 yielding plasmids pME3712 and 3713, respectively. vosA cDNA was also amplified (OZG436/438 for n-yfp, OZG437/438 for c-yfp fusion) from sexual cDNA library. vosA cDNA amplicons (OZG436/438) and (OZG437/438) were fused to n-yfp (OZG73/387) and c-yfp (OZG75/388) via fusion PCR [54]. c-yfp::vosA and n-yfp::vosA fragments were inserted into SwaI site of pME3712 and 3713, respectively. Plasmids bearing n-yfp::velB/c-yfp::vosA and c-yfp::velB/n-yfp::vosA were named as pME3714 and pME3715, respectively. For the analysis of VelB-VelB dimer formation, c-yfp::velB fragment was cloned into the SwaI site of pME3712 generating pME3717. nosA cDNA, which was amplified from sexual cDNA library (OZG320/321), was cloned into the PmeI site of pME3718 yielding pME3719.
The veA::ctap fusion construct encompassing the promoter and terminator sequences was amplified from pME3157 with oligos OZG304/305. This amplicon was cloned in the blunted ApaI site of pNV1 [55] generating pME3711.
Northern [56] and Southern [57] hybridization experiments were performed as given in detail [9]. Band densities in the Northern blots were analyzed with IMAGEJ (National Institutes of Health) and normalized against rRNA. DNA and amino acid sequences were analyzed by using Lasergene software (DNAstar). Northern blot probes were generated by PCR amplification of the following genes (primer sets): abaA cDNA (abaA5/abaA3), brlA cDNA (brlA5/brlA3), mutA cDNA (mutA5/mutA3), nosA cDNA (nosA5/nosA3), steA gDNA (steA5/steA3), nsdD cDNA (nsdD5/nsdD3), aflR gDNA (aflR5/aflR3), laeA cDNA (OZG61/OZG62), gpdA gDNA (gpdA5/gpdA3), tpsA gDNA (OMN176/OMN177), orlA gDNA (OMN182/OMN183), ipnA gDNA (ipnA5/ipnA3), and stcU gDNA (stcU5/stcU3).
Viability of spores was examined as described [16]. Two-day old conidia (105 per plate) of wild type and the mutants were spread on solid minimal medium (MM) and incubated at 37°C. After 2∼10 days the conidia were collected and counted in a hemocytometer. Approximately 200 conidia were inoculated on solid MM and incubated for 2 days at 37°C. Survival rates were calculated as a ratio of the number of growing colonies to the number of spores inoculated. This test was performed in triplicate.
Trehalose was extracted from conidia and analyzed as described previously [16], [58]. Two-day old conidia (2×108) were collected and washed with ddH2O. Conidia were resuspended in 200 µl of ddH2O and incubated at 95°C for 20 min and the supernatant was collected by centrifugation. The supernatant was mixed with equal volume of 0.2 M sodium citrate (pH 5.5) and samples were incubated at 37°C for 8 h with or without 3 mU of trehalase (Sigma), which hydrolyzes trehalose to glucose. The amount of glucose generated was assayed with a glucose assay kit (Sigma). The amount of glucose by deducting trehalase untreated sample from trehalase-treated sample was converted into the trehalose amount (pg) per conidium (triplicate).
Oxidative stress tolerance test was carried out as described previously [59]. Hydrogen peroxide sensitivity of conidia was tested by incubating 1 ml of conidial suspensions containing 105 conidia with varying concentrations (0.0, 0.25 or 0.5 M) of H2O2 for 30 min at RT. Each conidia suspension was then diluted with ddH2O, and the conidia were inoculated into solid MM. After incubation at 37°C for 48 h, colony numbers were counted and calculated as a ratio to the untreated control. Sensitivity to oxidative stress was also tested by spotting 10 µl of serially diluted conidia (10 to 105) on solid MM with 0, 2.5, 5 M of H2O2 and incubated at 37°C for 48 h.
UV tolerance test was carried out as described previously [60] with a slight modification. Two-day old conidia were collected in ddH2O and plated out on solid MM (100 conidia per plate). The plates were then irradiated immediately with UV using a UV crosslinker and the plates were further incubated at 37°C for 48 h. The colony numbers were counted and calculated as a ratio to the untreated control. UV sensitivity was also tested by spotting 10 µl of serially diluted conidia (10 to 105) on solid MM, which were then irradiated with UV and incubated at 37°C for 48 h.
For detection of GFP signal in 80 µg protein extracts, α-gfp mouse antibody (SantaCruz) was used in combination with One-Hour Western kit (Genscript). α-Calmodulin rabbit antibodies (Millipore) in 1∶1000 dilution in TBS 5% (w/v) non fat dry milk and secondary goat α-rabbit antibodies 1∶1000 in dilution in TBS 5% (w/v) milk were used for the recognition of TAP tag fusion proteins in 80 µg protein extracts. Polyclonal α-VeA antibody recognizing the native VeA protein was raised in rabbit (Genscript). α-VeA antibody (5 µg) in TBST 5% (w/v) milk 0.2% (v/v) Tween-20 was used for the detection of the VeA protein in 80 µg protein extracts in immunoblotting.
Protein extracts were prepared in B buffer (100 mM Tris pH 7.5, 300 mM NaCl, 10% Glycerol, 0,1% NP-40, 1 mM DTT, protease inhibitor mix (Roche)) without phosphatase inhibitors. Total protein extract (1 mg) was treated with 10 units of Shrimp Alkaline Phosphatase (SAP, MBI Fermentas) at 37°C for 30 min. SAP-treated extracts were used for immunoblotting.
Tap tag experiments and preparation of the protein crude extracts were performed as explained in detail [9].
Protocols given elsewhere [9] were followed for further data processing and analysis of the proteins.
A. nidulans spores (2000) were inoculated in 8 chambered borosilicate coverglass system (Nunc) supplemented with liquid medium. Fluorescence photographs were taken with an Axiovert Observer. Z1 (Zeiss) microscope equipped with a QuantEM:512SC (Photometrics) digital camera and the SlideBook 5.0 software package (Intelligent Imaging Innovations). For BIFC and GFP studies the following parameters were used; YFP filter 1000 milliseconds (ms), RFP filter 600 ms, DAPI filter 40 ms, DIC filter 200 ms, and GFP filter 400 ms.
Extraction of ST and running on TLC plates were performed as described in detail elsewhere previously [49].
|
10.1371/journal.ppat.1007574 | TDP-43 proteinopathy in Theiler’s murine encephalomyelitis virus infection | TDP-43, an RNA-binding protein that is primarily nuclear and important in splicing and RNA metabolism, is mislocalized from the nucleus to the cytoplasm of neural cells in amyotrophic lateral sclerosis (ALS), and contributes to disease. We sought to investigate whether TDP-43 is mislocalized in infections with the acute neuronal GDVII strain and the persistent demyelinating DA strain of Theiler’s virus murine encephalomyelitis virus (TMEV), a member of the Cardiovirus genus of Picornaviridae because: i) L protein of both strains is known to disrupt nucleocytoplasmic transport, including transport of polypyrimidine tract binding protein, an RNA-binding protein, ii) motor neurons and oligodendrocytes are targeted in both TMEV infection and ALS. TDP-43 phosphorylation, cleavage, and cytoplasmic mislocalization to an aggresome were observed in wild type TMEV-infected cultured cells, with predicted splicing abnormalities. In contrast, cells infected with DA and GDVII strains that have L deletion had rare TDP-43 mislocalization and no aggresome formation. TDP-43 mislocalization was also present in neural cells of TMEV acutely-infected mice. Of note, TDP-43 was mislocalized six weeks after DA infection to the cytoplasm of oligodendrocytes and other glial cells in demyelinating lesions of spinal white matter. A recent study showed that TDP-43 knock down in oligodendrocytes in mice led to demyelination and death of this neural cell [1], suggesting that TMEV infection mislocalization of TDP-43 and other RNA-binding proteins is predicted to disrupt key cellular processes and contribute to the pathogenesis of TMEV-induced diseases. Drugs that inhibit nuclear export may have a role in antiviral therapy.
| TDP-43 is a widely expressed nuclear protein that shuttles between the nucleus and cytoplasm, and regulates many aspects of RNA processing, such as splicing, trafficking, stabilization, and miRNA production. In almost all cases of ALS, neuronal and glial TDP-43 is phosphorylated, cleaved, and mislocalized to the cytoplasm, where it aggregates into stress granules and insoluble inclusion bodies. Although the mechanisms involved in TDP-43 proteinopathy remain unclear, impaired nucleocytoplasmic trafficking is thought to play an important role. Here we investigated whether TDP-43 proteinopathy also occurs during TMEV infection since TMEV L protein is known to perturb nucleocytoplasmic transport. We found evidence of TDP-43 proteinopathy in both TMEV-infected cultured cells, with predicted splicing abnormalities, as well as in neural and glial cells of TMEV-infected mice. The findings suggest that TDP-43 may contribute to the pathogenesis of TMEV-induced diseases, including TMEV-induced immune-mediated demyelination.
| Trans-activation response (TAR) DNA-binding protein of 43 kDa (TDP-43) is an RNA-binding protein (as well as DNA-binding protein) primarily present in the nucleus and important in RNA processing, mRNA transport/stability, and mRNA translation [2–4]. A variety of cellular stresses normally triggers TDP-43 to transiently shuttle into the cytoplasm and assemble into stress granules (SGs). Due to an abnormality of nucleocytoplasmic transport that is known to occur in amyotrophic lateral sclerosis (ALS), TDP-43 accumulates in insoluble aggregates in the cytoplasm of glia and degenerating neurons in the central nervous system (CNS) [5–7]. The mislocalized TDP-43 is cleaved into C-terminal fragments (CTFs), phosphorylated, and/or ubiquitinated [8–10].The importance of TDP-43 in disease pathogenesis is evidenced by the fact that mutant TDP-43 is a rare cause of familial ALS and, like wild type (wt) TDP-43, is mislocalized to the cytoplasm.
TDP-43 proteinopathy has been described in a number of diseases in addition to ALS [11]. Since the leader (L) protein of Theiler’s murine encephalomyelitis virus (TMEV), a member of the Cardiovirus genus of Picornaviridae, is known to disrupt nucleocytoplasmic transport [12, 13], we wondered whether TDP-43 proteinopathy occurs in infections with this pathogen; however, it is known that different RNA binding proteins and different protein compositions of the nuclear pore complex are present in different cell types [14]. TMEV includes strains of two subgroups with different disease phenotypes in mice [15]. GDVII strain and other members of the GDVII subgroup do not persist, but cause an acute fatal gray matter disease. In contrast, DA strain and other members of the TO subgroup induce a subclinical acute gray matter disease followed by an immune-mediated demyelinating myelitis with virus persistence in the CNS for the life of the mouse. DA-induced demyelinating disease serves as an experimental model of multiple sclerosis (MS).
Here we report that TMEV infection of cultured cells causes L-dependent mislocalization of TDP-43, and L-independent cleavage and phosphorylation of TDP-43 along with splicing abnormalities. Mislocalization and phosphorylation of TDP-43 also occurs in neuronal cells following early TMEV infection of mice, and in oligodendroglia and other glial cells in demyelinated areas 6 weeks after DA virus infection. These results suggest that TDP-43 mislocalization occurs and presumably contributes to cellular dysfunction and death in TMEV infections. An important role for TDP-43 mislocalization in TMEV-induced demyelinating disease is suggested by recent findings that TDP-43 binds to mRNAs encoding myelin genes, and that a knockdown of TDP-43 in oligodendrocytes of mice leads to demyelination and the death of this neural cell [1].
In control mock-infected BHK-21 cells, expression of TDP-43 was primarily restricted to the nucleus (Fig 1A). Following infection with DA or GDVII virus, which was detected by positive staining for TMEV VP1 capsid protein, TDP-43 was depleted from the nucleus and aggregated in the cytoplasm (Figs 1A, 1B and S1). The location of TDP-43 was juxtanuclear in structures that resembled aggresomes (see below), which have been previously observed in TMEV-infected cells [16, 17]. In addition, phosphorylated TDP-43 (pTDP-43) was present in the cytoplasm of TMEV-infected cells (Figs 1C and S2).
We questioned whether other RNA-binding proteins were also mislocalized to the cytoplasm in TMEV-infected cells. For this reason, we investigated the localization in cells of i) fused in sarcoma (FUS), which like TDP-43 is a cause of familial ALS when mutated, and ii) polypyrimidine tract binding protein (PTB), which is known to be mislocalized in TMEV infections, where it plays a role in TMEV translation [18, 19]. DA infection induced cytoplasmic mislocalization of both FUS and PTB1, one of PTB isoforms, along with TDP-43 (Fig 1D and 1E).
Since TMEV L protein is known to disrupt nucleocytoplasmic trafficking, we investigated TDP-43 localization following infection with mutant TMEV that had an L deletion. As predicted, DAΔL and GDVIIΔL infection failed to induce mislocalization of TDP-43 in VP1-positive cells (Fig 1A and 1B), demonstrating that TDP-43 mislocalization is indeed L-dependent. In order to further confirm the importance of TMEV L in TDP-43 mislocalization, we transfected eukaryotic expression constructs pDA L and pGDVII L into BHK-21 cells. Although both of these expression constructs caused cytoplasmic mislocalization of TDP-43 in the three cell lines that were tested (Figs 1F and S3), TDP-43 was present in small aggregates in the cytoplasm rather than the aggresome that had been detected in wild type (wt) TMEV-infected cells. The different effect of the TMEV L expression constructs was not a result of a different level of L protein expression when compared to TMEV L protein expression (S4 Fig).
In order to confirm the cytoplasmic mislocalization of TDP-43 in TMEV-infected cells, we separated the nucleus and cytoplasm of cultured cells infected with TMEV (S5 Fig). The results confirmed the prominent TDP-43 mislocalization in infected cells. Some TDP-43 is present in the cytoplasm of mock and TMEVΔL-infected cells presumably due to the normal shuttling of this protein from the nucleus.
As noted above, the juxtanuclear location of TDP-43 seen following TMEV infection had a morphology typical of an aggresome. Vimentin surrounded these juxtanuclear structures (Fig 2A), as is true in the case of aggresomes [20]. TMEV infections of L929 cells also induced a juxtanuclear aggresome that contained PTB1 (Fig 2B). In contrast, TDP-43 was diffusely present in the nucleus and cytoplasm of DA- and GDVII-infected HeLa cells (Figs 2C and S6), and not in an aggresome, perhaps related to the poor growth of TMEV in these cells [21].
Aggresomes result from a remodeling of intracellular membranes to generate sites of virus replication [20]. Fig 3A shows that VP1 and double-stranded RNA (ds-RNA), produced during TMEV replication, decorated the margins of aggresomes in TMEV-infected BHK-21 cells; an orthogonal view demonstrates that there is only very partial colocalization of VP1 and TDP-43 (S7 Fig). DA L was present within the aggresome’s vimentin cage, while DA L*, a non-structural protein that inhibits RNase L, was in the cytoplasm, but outside the aggresome (Fig 3B). Although ds-RNA was detected in DAΔL virus-infected BHK-21 cells, it tended to be present in small aggregates throughout the cytoplasm (Fig 3C). VP1 generally had a similar localization to that found with dsRNA in DAΔL virus-infected cells, however, at times it was diffusely distributed in the cytoplasm, presumably related to increasing virion production over time (see later).
In order to assess the importance of aggresomes in TMEV infection, we made use of nocodazole, a microtubule inhibitor that interferes with aggresome formation. BHK-21 cells were exposed to nocodazole (10 μM, 1hr), and then infected with DA virus. Compared to levels obtained with no nocodazole treatment, nocodazole led to a 10-fold reduction in virus genome at an MOI of 1, and 100-fold reduction at an MOI of 0.25 (Fig 3E). As expected, nocodazole treatment decreased the virus titer by more than 10-fold at 12 HPI (Fig 3F). In contrast, the effect of nocodazole on the level of viral genome and infectivity was relatively small in HeLa cell (S8 Fig). These findings suggest that the effect of nocodozole on TMEV replication is not related to this drug’s general disruption of the cytoskeleton, but a more specific effect on aggresomes.
SGs are mainly composed of stalled translation preinitiation complexes, markers such as G3BP1, eIF3A, and TIA1, and RNA-binding proteins including TDP-43. These structures are cytoplasmic non-membranous structures that appear in cells exposed to various stresses, including virus infections [22]. Certain viruses are known to induce SGs while others inhibit SG formation [23]. At times of stress or following treatment with a SG inducer, there is formation of SGs < 1 μm or 1–2 μm in size (S9 Fig).
Borghese and Michiels [24] previously reported that DA L inhibits SG formation in HeLa cells, a human cell line. We examined this issue in HeLa cells as well as two rodent cell lines. Uninfected control BHK-21 cells have homogeneous cytoplasmic immunostaining of SG markers G3BP1, eIF3A and TIA1 (Figs 4 and S9). In DA- and GDVII-infected (rodent) BHK-21 and L929 cells, but not in infected HeLa cells, these markers are located in the aggresome of VP1-expressing cells and not in SGs (Figs 4A–4C and S10); the lack of aggresome formation in TMEV-infected HeLa cells may be associated with the inefficient TMEV infection described in these cells [21]. At times, a VP1-expressing BHK-21 cell expressed these markers in what appeared to be typical SG structures as well as aggresomes, suggesting that the markers (and RNA-binding proteins) may transiently assemble in SGs, and then over time, when there is increasing virus production, relocalize in aggresomes (Fig 4D). Other picornavirus infections are reported to also transiently induce SG formation, followed by an inhibition of SGs later in infection [23]. In the case of TMEVΔL virus-infected cells, typical SGs were induced that immunostained with G3BP1, eIF3A and TIA1 (Figs 4E, 4F and S10), indicating that L interferes with SG formation. The SGs induced by TMEVΔL virus infections rarely colocalized with TDP-43 and PTB1 (Fig 4G and 4H).
To determine whether TMEV infection induces cleavage of TDP-43, as in the case of ALS, we carried out Western blots on RIPA-soluble and insoluble (but urea soluble) fractions extracted from TMEV-infected BHK-21 cell lysates at 8 HPI. Following infection with both wt and TMEVΔL virus, ~35-kDa and ~25-kDa bands as well as the expected 43-kDa band of full-length TDP-43 were detected in the urea-soluble, but not RIPA-soluble fraction, of BHK-21 cell lysates (Fig 5A). These findings suggest that L-independent cleavage of TDP-43 occurs in BHK-21 cells. Of note, there was no clear correlation between TDP-43 cleavage and TMEV infection, as monitored by VP1 immunodetection.
TDP-43 is known to have an important role in alternative splicing, including cystic fibrosis transmembrane conductance regulator (CFTR) exon 9 skipping [25]. In order to assess splicing abnormalities in infected cells, we transfected L929 cells with a CFTR minigene construct. Compared to uninfected cells, DA or GDVII virus-infected cells had a decrease of the lower band, which corresponds to the exon 9 spliced product (Fig 5B and 5E). These results provide evidence of impaired splicing regulatory activity in the infected cells, presumably because of abnormal TDP-43 localization and aggregation associated with TMEV infection.
In the case of ALS and ALS/FTD, TDP-43 is depleted in nuclei of neural cells, and mislocalized and phosphorylated in inclusions in the cytoplasm (Fig 6A and 6B). In order to determine whether the findings that we observed in cultured cells were also present in TMEV-induced disease, we carried out immunohistochemical staining of the CNS of mice 1 week following infection with GDVII virus, a time when mice are paralyzed and near moribund. Neurons in the CA1 region of the hippocampus had VP1 immunostaining (Fig 6C) with mislocalization of TDP-43 to the cytoplasm (Fig 6D and 6E). pTDP-43 was present in the nucleus (Fig 6G) and cytoplasm (Fig 6H), at times in a compact cytoplasmic inclusion body (Fig 6I). Approximately 60% of VP1-positive cells in TMEV-infected mice had evidence of pTDP-43 (S11 Fig). In contrast, TDP-43 maintained its normal nuclear localization in uninfected CA3 region neurons from the same TMEV-infected mouse (Fig 6F) and in normal uninfected mice (S11 Fig). The spinal cord of infected mice showed perivascular mononuclear infiltrates (Fig 6J) with numerous VP1-positive anterior horn cells (Fig 6K and 6L) that had TDP-43 and PTB2 depleted from the nucleus (Fig 6M and 6N). Immunofluorescent staining confirmed the presence and aggregation of TDP-43 in the cytoplasm of VP1-positive motor neurons (Fig 6O).
DA virus produces a biphasic disease in SJL mice with minimal or subclinical disease within the first two weeks post-infection, followed by progressive paralysis from an inflammatory demyelination that peaks at 6 weeks post-infection. In the acute phase of DA virus infection, VP1-positive neurons and axons were present in the CA2 region of the hippocampus (Fig 7A and 7B); however, the severity of infection and inflammation was mild compared to that seen in GDVII virus-infected mice. Some cells appeared to have cytoplasmic as well as nuclear staining of TDP-43 and PTB2, a splicing isoform of PTB that is increased in neurons compared to other cell types (Fig 7C and 7D). The infected regions generally had a decrease in TDP-43 staining, perhaps partly because many of the infected cells had pTDP-43 (Fig 7E), which was not stained by the anti-TDP-43 antibody that was used.
Six weeks after infection with DA virus, the ventral region of the thoracic spinal cord showed perivascular mononuclear cell infiltrates, (Fig 7F), demyelination, and vacuolation (Fig 7G). Activated microglia clustered within or around the demyelinated areas (Fig 7H). In these demyelinated areas, TDP-43 was depleted from the nucleus and mislocalized to the cytoplasm of VP1-positive white matter glial cells (Fig 7I and 7J), including oligodendrocytes (Fig 7K).
TDP-43 is a ubiquitously expressed RNA-binding protein that predominantly resides in the nucleus, but shuttles across the nuclear membrane in association with mRNAs [26]. A hallmark of almost all cases of ALS is disruption of nucleocytoplasmic trafficking with cytoplasmic mislocalization, aggregation, cleavage, and phosphorylation of TDP-43 in neural cells [5, 7, 9]. TDP-43 mislocalization is thought to lead to abnormalities of splicing and RNA metabolism with subsequent neuronal dysfunction [4, 27, 28]. It is likely that the cytoplasmic mislocalization of other RNA-binding proteins also contributes to the abnormalities of splicing in ALS [29]. In the present study, we demonstrate that TMEV infection leads to cytoplasmic mislocalization of TDP-43 (as well as FUS and PTB) along with cleavage into products similar in size to those found in ALS [7] and TDP-43 phosphorylation. Importantly, TDP-43 mislocalization was also found in neural cells following acute infections of mice, and in oligodendrocytes and other glial cells in demyelinated regions 6 weeks after DA infection.
As is true of many pathogens, picornaviruses disrupt nucleocytoplasmic trafficking during infection, leading to cellular dysfunction as well as the redistribution and hijacking of nuclear proteins into the cytoplasm for use during virus replication [30, 31]. For example, in infections of cultured cells by coxsackievirus B3 (CVB3), a member of the Enterovirus genus of Picornaviridae, TDP-43 is mislocalized (by viral protease 2A) and cleaved (by viral protease 3C) [32]. In CVB3 infections, TDP-43 colocalized with SGs in the cytoplasm at 3 HPI, the longest time observed. Human immunodeficiency virus-positive neural cells have also been reported to have TDP-43 in cytoplasmic inclusions [33].
In ALS, TDP-43 is thought to shuttle into the cytoplasm initially into SGs, and then remain aggregated in the cytoplasm. In the case of TMEV-infected BHK-21 and L929 cells, we detected the normal markers for SGs (G3BP1, TIA1 and eIF3A) in aggresomes rather than SGs. Of note, SGs were present following TMEVΔL infections, suggesting that L inhibits SG formation, as has been reported by others [24]. The aggresomes in TMEV-infected cells also contained TDP-43, FUS, PTB1, TMEV proteins (VP1, L), and dsRNA. Nocodazole, a microtubule inhibitor that interferes with aggresome formation, decreased viral replication, suggesting that TMEV uses the aggresome as a “viral factory,” perhaps by concentrating proteins and genome in one region of the cell, as described for other virus infections [34]; however, nocodazole’s disruption of the cytoskeleton with a resultant disturbance of cell physiology may also have had a substantial indirect effect on viral replication. In contrast, TMEV infection of HeLa cells led to minimal cytoplasmic translocation of TDP-43 with no aggresome formation, perhaps a reflection of the reported inefficient infection of these cells [21]; the reasons for the lack of aggresome formation and inefficient infection remain unclear. Cytoplasmic TDP-43 aggregates in ALS have also been referred to as aggresomes [35–37]. In the latter case, the aggresome is thought to be a cytoprotective response that sequesters potentially toxic misfolded proteins and facilitates their clearance by autophagy [20, 38].
Mislocalization and phosphorylation of TDP-43 occurred in TMEV-infected cultured cells as well as neuronal and glial cells of TMEV-infected mice. In DA virus-induced demyelinated regions, TDP-43 and other RNA-binding proteins were mislocalized in glial cells, including oligodendrocytes. The fact that TDP-43 was not present in the cytoplasm following infection with TMEVΔL virus, suggests that L interfered with nucleocytoplasmic transport. TDP-43 mislocalization in neural cells may also be influenced by inflammatory stimuli since tumor necrosis factor-α can lead to mislocalization of TDP-43 [39]. In addition, interferon-γ leads to hnRNP A1 mislocalization and accumulation into the cytoplasm [40].
The mislocalization of RNA-binding proteins in TMEV infections may disrupt cellular splicing and mRNA translation, thereby contributing to neuronal dysfunction and death in GDVII and DA early disease as well as oligodendrocyte dysfunction in the late demyelinating disease of DA-infected mice. Our previous study suggested the possibility that PTB mislocalization in TMEV-infected neurons played a role in neuronal dysfunction [41]. The deleterious effect of PTB2 mislocalization in neurons may be compounded by the FUS mislocalization that was also present, since the latter RNA binding protein is important in axonal transport [42]. Importantly, recent studies have demonstrated that: i) TDP-43 binds to mRNAs of myelin proteins, ii) knockdown of TDP-43 in oligodendrocytes of mice leads to demyelination and death of this neural cell [1]. TMEV L-dependent nucleocytoplasmic trafficking defect is likely to also interfere with other RNA binding proteins in addition to the three that were investigated as well as to disrupt the proper subcellular localization of a number of key transcription factors and proteins in oligodendrocytes and oligodendrocyte precursor cells that are needed for efficient myelination and remyelination [43–46]. Altered nucleocytoplasmic transport leading to mislocalization of RNA-binding proteins and other macromolecules with associated cellular dysfunction may underlie a number of disease states, both infectious as well as non-infectious. The importance of this mechanism of cell dysfunction highlights the potential relevance of antiviral drugs that target nucleocytoplasmic transport.
The study involving the analysis of human subjects was approved by The University of Chicago Institutional Review Board for Clinical Research. Informed written consent for an autopsy was obtained from an immediate member of the deceased’s family. Animal use was approved by The University of Chicago Institutional Animal Care and Use Committee (IACUC) under the Protocol Number 71772. Animal work conducted at the University of Chicago complies with all applicable provisions of the Animal Welfare Act (AWA) and the Public Health Service (PHS) Policy on Humane Care and Use of Laboratory Animals. The PHS Policy incorporates the standards in the Guide for the Care and Use of Laboratory Animals and the U.S. Government Principles for the Utilization and Care of Vertebrate Animals Used in Testing, Research and Training and requires euthanasia be conducted according to the AVMA Guidelines for the Euthanasia of Animals. The University of Chicago Animal Care Program has an approved Assurance with the National Institute of Health (NIH), is registered with the United States Department of Agriculture (USDA) and is accredited by the Association for Assessment and Accreditation of Laboratory Animal Care International (AAALAC).
DA and GDVII viruses were derived from a full-length infectious cDNA clone [15]. DAΔL virus has a deletion of amino acids 2 to 67 of L [47]. GDVIIΔL virus (originally referred to as dl-L virus) [48] has a deletion of amino acids 2 to 71 of L, and was previously received as a gift from M. K. Rundell.
Infections were carried out in BHK-21 (ATCC, CCL-10), L929 (ATCC, CRL-6364) or HeLa cells (ATCC, CCL-2), usually with a multiplicity of infection (MOI) of 10. BHK-21 cells were used for plaque assays and the growth of virus stocks, as previously described [49].
For the study of TDP-43 cleavage, cells were treated with 1μM of the proteasome inhibitor MG-132 (Cell Signaling Technology, Danvers, MA) for 16hs prior to harvest. For induction of SGs, BHK-21 cells were treated for 45 min with 0.5mM sodium arsenite (Sigma Aldrich, St Louis, MO). In investigations of the aggresome, nocodazole (Sigma Aldrich), a microtubule inhibitor, was solubilized in DMSO and added at varying concentrations to the culture medium for 1h prior to infection.
pDAL and pGDVIIL, which are eukaryotic expression constructs of DA L and GDVII L respectively, with myc/His epitope tags at the carboxyl terminus [47], were transfected into BHK-21 cells using Lipofectamine 3000 (Thermo Fisher Scientific, Waltham, MA).
Cells on coverslips were harvested 8hs post infection (HPI) or 48hs after transfection, fixed in 4% paraformaldehyde for 5 min, and then permeabilized with phosphate buffered saline (PBS) with 0.1% Triton X-100 for 20 min at room temperature. The coverslips were then incubated overnight at 4°C with primary antibodies (S1 Table). After rinsing, cells were incubated for 30 min with Alexa 594-conjugated goat anti-mouse IgG and Alexa 488-conjugated goat anti-rabbit IgG (Invitrogen, Carlsbad, CA), and then counterstained with 4',6-diamidino-2-phenylindole (DAPI). Images were captured using a confocal laser microscope system (Leica TCS SP5, Leica Microsystems, Wetzlar, Germany). A sequential multiple fluorescence scanning mode was used to avoid nonspecific overlap of signals. In some experiments, manual counting of infected cells was carried out in five different regions of the coverslips.
Cells were lysed 8 HPI with a radioimmunoprecipitation assay (RIPA) buffer containing a protease inhibitor and phosphatase inhibitor cocktail (Thermo Fisher Scientific, Waltham, MA). Lysates were centrifuged at 14,000 rpm for 30 min at 4°C, and supernatants collected as RIPA buffer-soluble proteins. The pellets were sonicated and centrifuged twice at 14,000 rpm for 30 min at 4°C to obtain RIPA buffer-insoluble pellets. Pellets were dissolved in urea buffer (8 M urea, 50 mM Tris-HCl, pH 8.5) and then sonicated again prior to electrophoresis. Ten μg of total protein quantified by a Pierce BCA Protein Assay Kit (Thermo Fisher Scientific, Waltham, MA) was subjected to electrophoresis on 10% SDS polyacrylamide gels, and then transferred to Amersham Hybond P 0.45 μm PVDF membrane (GE Healthcare, Buckinghamshire, UK). The membrane was first blocked with 5% non-fat skim milk in Tris-buffered saline (TBS) containing 0.05% Tween-20 for 30 min at room temperature, and then incubated for 1h at room temperature with a rabbit antibody directed against C-terminal TDP-43 (1:1000, Proteintech, Rosemont, IL) or a mouse monoclonal antibody against TMEV VP1 (1:2000), which was previously called GDVII mAb2 [50], or a mouse monoclonal antibody against Lamin A/C (1:1000, Cell Signaling Technology, Danvers, MA), or a mouse monoclonal antibody against β–actin (1:5000, Sigma Aldrich, St Louis, MO). Following washing, the membrane was incubated with anti-rabbit or anti-mouse horseradish peroxidase–conjugated secondary antibodies (GE Healthcare, Buckinghamshire, UK) for 1h at room temperature. The signal was detected using SuperSignal West Dura Extended Duration Substrate (Thermo Fisher Scientific, Waltham, MA), and analyzed using ChemiDoc MP Imaging System (Bio-Rad Laboratories, Hercules, CA).
DA RNA was extracted from BHK-21 and HeLa cell homogenates using an RNeasy Plus minikit (Qiagen). A region between nt 1485 and 1684 was amplified using forward primer TACTATGGCACCTCTCCTCTTGGA and reverse primer CAGCCGCAAGAACTTTATCCGTTG with a Superscript III Platinum two-step qRT-PCR kit with SYBR green (Invitrogen). A region between nt 182 and 721 of the murine β-actin gene, which was used for normalization and determination of the quality of total mRNA, was amplified using forward primer GTGGGCCGCTCTAGGCACCAA and reverse primer CTCTTTGATGTCACGCACGATTTC. qRT-PCR was conducted on a CFX96 Real-Time System (Bio-Rad). The ΔΔCT method of relative quantitation was used to calculate fold change of DA with β-actin.
In order to assess splicing in TMEV-infected cells, a cystic fibrosis transmembrane conductance regulator (CFTR) minigene construct designed to evaluate CFTR exon 9 splicing (a gift from Virginia Lee’s lab and described in Buratti et al.[25]) was transfected into L929 cells using Lipofectamine LTX reagent (Invitrogen). Twenty-four hours later, the cells were infected separately with DA or GDVII viruses at an MOI of 10. Total RNA was prepared from cells 12 h after infection of viruses, and RT-PCR was performed with 1 μg of total RNA and 1 μl of resulting cDNA. The relative exclusion of exon 9 was evaluated by primer extension from the flanking sequence of exon 9 using the following primers, as previously described [25]: TAGGATCCGGTCACCAGGAAGTTGGTTAAATCA; CAACTTCAAGCTCCTAAGCCACTGC. The PCR products were visualized on a 2% agarose gel. Relative amounts of different splice products were quantified and visualized using Image J. The experiments were repeated in triplicate.
Immunohistochemical studies were performed on autopsied brain specimens of a patient with ALS.
TMEV was inoculated intracerebrally in weanling SJL mice (Jackson Laboratory, Bar Harbor, ME), and mice were sacrificed at 1, 2 or 6 weeks post infection (PI). At the time of sacrifice, mice were deeply anesthetized and perfused transcardially first with PBS, and then with 4% paraformaldehyde in 0.1 M phosphate buffer. CNS tissues were fixed in 10% buffered formalin and processed into paraffin sections (5 μm thick). Deparaffinized sections were hydrated in ethanol and then incubated with 0.3% hydrogen peroxide in absolute methanol for 30 min at room temperature to inhibit endogenous peroxidase. After rinsing with tap water, sections were washed twice using Tris–HCl with 0.1% Triton X-100 for 5 min, and then with Tris–HCl for 5 min. Sections were then incubated at 4°C overnight with primary antibody (S2 Table) diluted in 5% normal goat serum, 50 mM Tris-HCl (pH 7.6) and 1% BSA. After rinsing, sections were subjected to labeling by an enhanced indirect immunoperoxidase method. The reaction product was developed using a solution of 3, 3’-diaminobenzidine (DAB). Sections were counterstained with hematoxylin. Double immunostaining was carried out with two enzyme systems, peroxidase and alkaline phosphatase, followed by staining with Vector Red (Vector Laboratories).
Paraffin sections were also used for immunofluorescent staining. Sections were deparaffinized in xylene, rehydrated through an ethanol gradient, and then incubated with primary antibody for 1h at room temperature. The following antibodies were used: mouse monoclonal anti-VP1, rabbit anti-TDP-43, and rabbit anti-2',3'-cyclic nucleotide-3'-phosphodiesterase (CNPase) (S2 Table); rabbit anti-TDP-43 was pre-conjugated with Zenon Alexa Fluor 647 rabbit IgG (Invitrogen). After rinsing, sections were incubated for double immunofluorescence with Alexa 488-conjugated goat anti-rabbit IgG and Alexa 594-conjugated goat anti-mouse IgG (Invitrogen) and for triple immunofluorescent staining with Alexa 488-conjugated goat anti-rabbit IgG, Alexa 555-conjugated goat anti-mouse IgG (Invitrogen), and DAPI. Images were captured using a confocal laser microscope system (Leica TCS SP5, Leica Microsystems, Wetzlar, Germany) with sequential multiple fluorescence scanning mode to avoid non-specific overlap of colors. All photographs were captured under the same magnification, laser intensity, gain and offset values, and pinhole setting.
Statistical analysis was performed by an unpaired t-test or one-way ANOVA with Tukey's multiple comparisons test using GraphPad Prism version 7.0a. A P-value of <0.05 was considered significant. The data are presented as the mean ± standard deviation (S.D.).
|
10.1371/journal.pgen.1007838 | Evolution of maternal and zygotic mRNA complements in the early Drosophila embryo | The earliest stages of animal development are controlled by maternally deposited mRNA transcripts and proteins. Once the zygote is able to transcribe its own genome, maternal transcripts are degraded, in a tightly regulated process known as the maternal to zygotic transition (MZT). While this process has been well-studied within model species, we have little knowledge of how the pools of maternal and zygotic transcripts evolve. To characterize the evolutionary dynamics and functional constraints on early embryonic expression, we created a transcriptomic dataset for 14 Drosophila species spanning over 50 million years of evolution, at developmental stages before and after the MZT, and compared our results with a previously published Aedes aegypti developmental time course. We found deep conservation over 250 million years of a core set of genes transcribed only by the zygote. This select group is highly enriched in transcription factors that play critical roles in early development. However, we also identify a surprisingly high level of change in the transcripts represented at both stages over the phylogeny. While mRNA levels of genes with maternally deposited transcripts are more highly conserved than zygotic genes, those maternal transcripts that are completely degraded at the MZT vary dramatically between species. We also show that hundreds of genes have different isoform usage between the maternal and zygotic genomes. Our work suggests that maternal transcript deposition and early zygotic transcription are remarkably dynamic over evolutionary time, despite the widespread conservation of early developmental processes.
| Genetic control of embryonic development in all animals requires precise coordination between mother and zygote. The mother provides gene products to the egg to drive the earliest stages of development, until the zygote is able to transcribe its own genome. Many processes of early development are highly conserved over evolutionary time, and are critical for organism survival. Here, we determine how conserved the pools of transcripts provided by the mother and transcribed by the zygote are over evolutionary time, by sequencing all the transcripts at each stage in 14 species of fruit flies at different stages of development. We find a substantial conservation of transcripts represented, with the transcripts deposited by the mother being especially highly conserved. However, we also find considerable variation in these early transcript pools between species, suggesting that while early developmental processes are highly conserved, the gene products driving them may not be. This may be necessary for carrying out the processes of development in different environments.
| Most early developmental processes, such as rapid cleavage cycles and the establishment of body axes, are shared across multicellular animals, but the extent to which the mechanisms and the genes involved are also shared remains an open question.
Throughout the animal kingdom, the first stages of development are controlled by mRNA transcripts and proteins deposited by the mother during oogenesis. Genetic control is subsequently transferred from the maternal genome to the zygotic genome. This is accomplished through a precise and elegant series of regulatory steps, in which the zygotic genome is transcriptionally activated while maternal transcripts are degraded, in a process known as the maternal to zygotic transition (MZT; [1,2]. This handoff between mother and zygote has the appearance of a functional logic that dictates which genome is in control. Genes involved in processes unique to the earliest stages of development, such as rapid cleavage cycles, are necessarily transcribed by the mother. Genes that control processes such as patterned gene expression in the developing embryo require zygotic transcription from specific subsets of cells. And genes performing essential housekeeping functions required at all stages of life are transcribed by both the mother and the zygote. For these genes, the maternal and zygotic genomes are able to coordinate during the MZT to such a degree that the transcript levels of these genes can be relatively constant, despite the transition between genomes of origin for these transcripts.
In general, the logic underlying the partitioning of gene expression between the maternal and zygotic genomes is unclear. While we have examples of particular genes that are transcribed by either (or both) the maternal and zygotic genomes, in accordance with the requirements discussed in the previous paragraph, it is unknown whether these requirements play out to shape maternal and zygotic gene expression genome-wide. One way to address this question is to analyze evolution of transcript pools at these stages on short to moderate timescales. If the genome of origin is a constraining factor for many genes, we would expect to see a high degree of conservation of maternal or zygotic expression for those genes. Alternatively, if transcripts of some genes may be supplied by either mother or zygote, we might expect to see control of expression vary between the two genomes across species.
The current evidence for conservation of maternal and early zygotic regulatory factors is mixed. One of the most critical maternal genes in the fruit fly Drosophila melanogaster, bicoid [3], is of relatively recent evolutionary origin. Bicoid controls axis formation (determines the anterior pole of the egg), and is not found outside of higher Diptera [4], having resulted from a gene duplication. This demonstrates that conserved early developmental processes can incorporate new genes. Some theory and empirical studies suggest that maternal genes might be expected to evolve more quickly than zygotic genes, as selection will be less effective since these are (largely) autosomal genes that are expressed in only one sex [5–7]. These studies examined coding region changes in a limited number of genes, and might not fully account for the significant developmental constraint imposed by the need to build functional offspring. However, recent genomic studies demonstrate a high degree of maternal transcript level conservation relative to zygotic gene expression [8,9].
Mechanistically, maternal deposition and zygotic expression are subject to different constraints. Prior to zygotic genome activation, deposited maternal gene products are the only mRNA transcripts available, and transcript level cannot be dynamically increased to respond to the rapidly changing environment of the earliest developmental stages. Perhaps because of this, post-transcriptional control mechanisms play an especially important role in the levels of gene products produced from maternal transcripts [10–15]. On the other hand, since the products of many zygotic genes are needed soon after transcription is activated, the accumulation of sufficient transcripts in a short period of time (especially in species with rapid development such as Drosophila) can be difficult. For this reason, genes expressed in Drosophila at the early zygotic stage tend to be short in length and contain few or no introns [9,16,17], allowing them to be transcribed quickly. Since both the maternal and zygotic phases of early development have unique constraints, predicting the level of conservation of each stage is challenging.
In order to assess the extent of transcript level conservation in early development across evolutionary time, we sequenced transcriptomes from early embryonic stages from 14 Drosophila species. These species span divergence times of approximately 250,000 to close to 60 million years. We determined mRNA levels both before and after zygotic genome activation, analyzing the conservation and divergence of transcript representation across species over evolutionary time.
Our findings show strong levels of evolutionary conservation of both maternal and zygotic transcripts. However, genes that are represented at only one stage (either before or after zygotic genome activation) show strikingly high degrees of transcript level evolution compared to genes represented at both stages. The results suggest that in most cases robust transcript levels may be achieved through regulatory mechanisms that rely on both the maternal and zygotic genomes. Nevertheless, when all transcripts at the maternal stage are compared to all the genes at the post-genome activation stage, the maternal stage has a higher degree of conservation. This is in contrast to the pattern observed with stage-specific transcripts, where transcripts present at the maternal stage that are entirely degraded at the MZT evolve faster than zygotic genes with no maternal contribution. Furthermore, we find that expression levels of a small proportion of zygotic genes with no maternal contribution are tightly regulated, and the use of stage-specific isoforms may be a hitherto unrecognized method of partitioning the contributions of two genomes. Finally, combining our results with data from the mosquito Aedes aegypti, we show the conservation of a core set of zygotic genes over 250 million years. Our study demonstrates the power of transcriptomic phylogenetics to identify the key players regulating core developmental processes.
In order to characterize how transcript representation and abundance across the maternal to zygotic transition changes over evolutionary time, we have created an RNA-Seq dataset from 14 Drosophila species at two developmental timepoints. The 14 species chosen represent divergence times of 0.25–57 million years [18], considerable life history variation [19], and fully sequenced genomes [20–22]. We sampled two developmental timepoints, one (stage 2, Bownes’ stages) where all transcripts are maternally derived, and another (the end of stage 5/blastoderm stage) after the onset of zygotic transcription, and just prior to gastrulation. Since these developmental stages are morphologically distinct [23,24] and highly conserved [25], comparable timepoints can be identified across species. Each species and timepoint was represented by at least 3 single-embryo replicates (S1 Table; see Methods for full description of the dataset), although in one case (Drosophila ananassae) only two stage 2 replicates were used in the analysis.
Our results confirm our previous finding that single-embryo RNA-seq is highly reproducible [26–28]. Spearman rank correlation coefficients of transcript abundance (in fragments per kilobase per million reads mapped, or FPKM) for replicates from the same stage are high at both developmental stages (S1 Table), though slightly lower at stage 5 (post-zygotic genome activation) than stage 2 (maternal transcripts only). For stage 2, the correlation coefficients within species range from 0.94–0.98 across 13 of the 14 species (Fig 1A; S1 Table), with a mean of 0.96. D. erecta stage 2 transcriptomes have slightly lower correlations (0.91–0.96). Spearman coefficients for stage 5 replicates (Fig 1A; S1 Table) are always lower than their stage 2 counterparts and are also more variable, ranging from 0.80 to 0.95, with a mean of 0.90. They are still higher, however, than correlation coefficients of replicates from different stages (for any given species), which range from 0.67 to 0.83, with a mean of 0.76 (Fig 1C; S1 Table).
There are a few reasons why stage 5 transcript levels may be expected to be less correlated than stage 2, even within replicates. First, stage 5 transcript levels are the result of both zygotic transcription and maternal transcripts that have yet to degrade. Approximately 50% of maternal transcripts are still present at this timepoint [1,17,29], and degradation of maternal genes may vary between species [28]. As full activation of the zytogic transcriptome begins during stage 5, small differences in developmental timing between replicates in this stage may also produce differences in zygotic transcripts present. To address this, our stage 5 timepoint is a precise point at the end of this stage, when cellularization has completed, but prior to gastrulation (see Methods). Finally, embryo sex may begin to play a role by the end of stage 5. As the sex determination pathway has been activated by this stage, it is possible to distinguish males from females by comparing levels of known female- and male-specific transcripts (Sxl and msl-2) [30]. While sex-specific differences in stage 5 transcript abundance across species have been described previously [28], there are no consistent differences between Spearman rank sum comparisons of same-sex vs. opposite-sex stage 5 embryos (S1 Table), suggesting that these differences are overwhelmed by transcript levels from non-X-linked genes when comparing whole transcriptomes.
Since replicate FPKM levels are highly correlated, our analysis in the remainder of this paper focuses on mean replicate FPKM values for a given species and stage.
While transcript levels within each stage are highly correlated across species (Fig 2; S2 Table), they diverge as evolutionary distance increases. Spearman rank correlation coefficients decrease when comparing a species from within the melanogaster subgroup (e.g. D. melanogaster) with more distantly related flies, but only drop to around 0.7 for divergence times of approximately 57 million years (e.g. D. melanogaster to D. virilis, Fig 2C). As was found with replicates from the same species, stage 5 correlation coefficients are usually slightly lower than the equivalent stage 2 coefficients (Fig 2B and 2C; S2 Table). In contrast, cross-species comparisons of different stages show striking differences. Clustering the transcriptomes from all species and timepoints (S1 Fig), the two stages fall out as the first two distinct clusters, demonstrating that the biological distinctiveness of these developmental stages in Drosophila transcends interspecific variation.
While divergence in transcript levels generally increases with evolutionary distance, transcriptome comparisons do not recapitulate the phylogeny (S1A Fig). In particular, the obscura group shows strong transcriptomic divergence in excess of its phylogenetic distance from other species. The obscura group has experienced a number of fusions of sex chromosomes, resulting in a larger proportion of the genome being sex-linked [31]. To determine if this could be driving the pattern we observe, we performed the clustering analysis on autosomal genes, removing all gene groups where one or more orthologs were located on a sex chromosome in any of the species with chromosomal-level annotations. This required removing gene groups with an ortholog on the X chromosome, or Muller element A, in D. melanogaster, D. simulans, D. mauritiana, D. yakuba, D. pseudoobscura, and D. miranda, in addition to gene groups with an ortholog on Muller D in D. pseudoobscura or D. miranda or an ortholog on Muller C in D. miranda [32–34]. We found that the obscura group still clusters together outside the rest of the species when sex-linked genes were removed from the analysis (S1B and S1C Fig). As this group represents the only clade in our study adapted to a temperate environment (D. virilis is also considered a temperate species, but has no close relatives in our dataset), some of the differences in this group may be ecologically relevant. We return to this observation below.
We then focused on stage-restricted genes, either maternal-only (maternally deposited and entirely degraded by stage 5) or zygotic-only (not maternally deposited, transcribed from the zygotic genome and present at stage 5); see Methods for further definitions. The number of maternal-only genes is relatively small; on average this group represents ~6% of transcripts present at either or both of these stages (S3 Table). In comparison, zygotic-only transcripts represent ~20% of those present at these stages (S3 Table). When comparing species pairs, we found that as evolutionary distance increases, the number of orthologs that are restricted to the same stage in both species declines, particularly in the maternal-only set (S2 Fig). If the transcript levels of genes that are restricted to a given stage in either of the two species are compared, we find that correlation coefficients for both types of stage-restricted genes are markedly lower than those for all stage 2 and stage 5 genes (S3 Fig). However, if we only compare the transcript levels of genes that are stage-restricted in both species (Fig 3), we see a striking contrast between the maternal-only group, which shows strong transcriptomic divergence as evolutionary distance increases, and the zygotic-only group, which shows remarkable conservation among even the most distant species in our analysis. This suggests that while both groups of stage-restricted genes show rapid evolution relative to non-stage restricted genes present at these stages, the transcript levels of a subset of zygotic-only genes are more tightly regulated across Drosophila clades.
To investigate evolution of transcript pools in early development over longer periods of evolutionary time, we incorporated data from the previously published developmental transcriptomic time series of a basal Dipteran, the mosquito Aedes aegypti, which separated from Drosophila 170 to 250 million years ago [35]. Since Aedes aegypti shares the long-germ band mode of development with Drosophila [36], where the body plan is established simultaneously rather than sequentially [37], early developmental processes are largely conserved. In the Aedes aegypti dataset, all transcripts are maternal at the earliest stage assayed (0–2 hours; the equivalent of Drosophila stage 2), while the approximate equivalent of Drosophila stage 5 is around 10 hours, when Aedes cellularization is complete [38], covered by the 8–12 hour window in this time series. We will refer to these stages as Aedes stage 2 and Aedes stage 5, respectively.
As shown in hierarchical clustering analysis of the Drosophila and Aedes transcriptomes (Fig 4A, S1C Fig), the Aedes transcriptomes from both developmental stages cluster together, rather than clustering with the comparable Drosophila stages, suggesting that evolutionary divergence of this magnitude is more significant than divergence between developmental stages. However, despite this divergence, and in spite of the differences in the experimental protocols that generated the data (the Drosophila transcriptomes were generated from single embryos, while the Aedes time series used pooled embryos), there are still moderately strong and highly significant correlation coefficients between transcript levels of orthologous genes from the respective stages (Fig 4B, S2 Table). We were particularly interested in analyzing the classes of genes that showed the greatest change in representation within Drosophila (maternal-only and zygotic-only) and determining whether we could find conserved orthologs between Drosophila and basal Diptera. In doing so, we found considerable change in stage-specific genes at this evolutionary distance, but also a core set of highly conserved zygotic-only genes likely to have important functions.
Considering first the zygotic-only genes, we examined two sets of genes: 1) genes that displayed any evidence of conservation between Aedes and a Drosophila species, and 2) a subset of these genes that also showed strong evidence of conservation within the Drosophila clade. Looking first for any conservation between Aedes and Drosophila, we found 173 genes (from a set of 4619 orthologs identified across both stages) that were zygotic-only in both Aedes and at least one of the 14 Drosophila species (S4 Table). This set of 173 genes represents approximately 10% of the mean number of zygotic-only genes in a Drosophila species (S3 Table). A given gene in this set was, on average, also zygotic-only in 6.5 other Drosophila species (S4A Fig). By contrast, a random gene from the larger dataset of 4619 orthologs was on average zygotic-only in less than one Drosophila species. To identify a more stringent set of genes that are also conserved within Drosophila, we required genes to be be zygotic-only in Aedes aegypti and two of the earliest-diverging Drosophila species (D. willistoni, the outgroup to Old World Sophophora, and at least one of the two Drosophila subgenus species). With this set of criteria, we identified a core set of 61 genes, which show even greater evolutionary constraint (S4B Fig), and are, on average, zygotic-only in 10 other Drosophila species (a total of 13 of the 15 species). This number represents ~4% of the average number of zygotic-only genes in a Drosophila species (S3 Table).
Our results indicate that although most zygotic-only genes show rapid change in representation over stages, a core set can be identified where the zygotic-only state is highly conserved (Fig 4C, S5 Table). These 61 genes represent key players in D. melanogaster embryonic patterning, with gap genes, pair-rule, segment polarity and homeotic genes in the anterior-posterior pathway represented, in addition to numerous genes in the dorsal-ventral patterning pathway. A gene ontology (GO) analysis using DAVID [39,40] comparing this set to the larger set of stage 5-represented genes shows that it is enriched approximately 8-fold in transcription factors (S4C Fig), with terms related to early development (S4D Fig) strongly overrepresented. These results suggest that a naïve analysis based on conservation of transcriptional state can yield remarkable insight into the set of genes that are functionally significant.
In contrast to the zygotic-only genes, we were unable to identify a significant core set of maternal-only (maternally deposited, completely degraded at the MZT) genes that were conserved between Aedes and Drosophila. Indeed, out of the approximately 4000 orthologs we examined, we only found 3 genes (beaten path Ia, Phospholipase D and Sclp) that were maternal-only in both D. melanogaster and A. aegypti. These results, which are consistent with our findings in Drosophila, suggest that this subset of the maternal genes (or the maternal-only status) may not be essential for key developmental processes that are conserved across Diptera. One possibility is that maternal deposition could be a process with considerable developmental noise, and that degradation might be a method for compensating for non-functional transcripts that are dumped by nurse cells during oogenesis. Alternatively, these transcripts may have clade-specific functions limited to the earliest stage of development.
The previous analysis showed that transcriptomic conservation during early embryonic development is the norm, with RNA levels tightly correlated at up to 60 million years of evolutionary divergence. However, by reconstructing ancestral states (presence or absence of gene transcripts) using the Bayesian phylogenetics package MrBayes [41] (Fig 5; S6 Table), we were also able to identify genes that change in their representation in either the maternal or zygotic transcript pools. For this analysis, a gene was considered represented at a given stage if the FPKM level of its isoforms was at least 1. The ancestral stage 2 and stage 5 states of 7092 genes with one-to-one orthologs in at least 12 of the 14 species were reconstructed (see Methods for more details). Even when using relatively liberal parameters for identifying transitions (see Methods for details), only 245 of these genes, or less than 4%, showed evidence of a stage 2 gain or loss at any node on the phylogeny (Fig 5A). Stage 5 transitions were approximately twice as common, with 499 observed (Fig 5A). By far the greatest number of transitions at both stages were seen in the obscura group (see Fig 5B for examples), in support of the hierarchical clustering results (Fig 4A, S1 Fig) suggesting that the early embryonic transcriptome in species of this group is exceptional. An additional finding of note is that loss of representation does not occur with greater frequency than gain of representation (Fig 5A).
Our phylogenetic analysis allowed us to address two major questions: 1) what are the evolutionary patterns of changes in transcript representation relative to maternal, zygotic, or maternal and zygotic representation; and 2) what types of transcripts undergo gains or losses along the tree? To address the former, we compared patterns of gains and losses along the phylogeny. Observing changes in transcript representation at both developmental stages, we rarely find genes that transition from an entirely maternal (maternal-only) state to an entirely zygotic (zygotic-only) state (S7 Table). Indeed, the only gene in S7 Table to show a simultaneous gain of representation at both stages is quasimodo (qsm), and there is no gene that shows a simultaneous loss at both stages. Primarily, we observe transcripts that are present at both stages losing representation at one stage, or transcripts that are present at one stage gaining representation at the other. This means, for example, that to transition from maternal-only to zygotic-only, the gene would pass through an evolutionarily intermediate state where it is represented at both stages.
Next, we examined which classes of genes change in transcript representation at either stage. We performed gene ontology analysis (see Methods) of genes that change transcript representation when compared with all transcripts present at that stage. In the case of genes that gain stage 5 representation, there is significant enrichment for genes that are involved in transport of small molecules through membranes (ion channel activity, substrate-specific channel activity, ion transport, ion channel complex, channel activity, etc.; see S5 Fig, S8 Table). We might expect transcription of these genes to evolve quickly [42], as their products permit adaptation to new environments on a cellular level. Additionally, some types of these transport mechanisms have been implicated in genetic conflict [43,44] and could thus be expected to evolve quickly. The other categories of transcript abundance changes (stage 5 losses, stage 2 gains, stage 2 losses) had no significant enrichment of gene ontology categories compared to the rest of the genes with transcripts present at that stage.
We highlight a few examples of individual gains of transcript representation in Fig 5B. These particular genes were chosen because they are dramatic examples of changes in transcript representation at different stages and in different species. Tpi (Triose phosphate isomerase) and Nhe2 (Na+/H+ hydrogen exchanger 2), show gains at stage 5 and stage 2 respectively, in the obscura lineage. Tpi functions in glycolysis, and is a classic example of clinal allele frequencies in D. melanogaster populations [45]. Nhe2 is a Na+/H+ exchanger that increases cellular pH, and has been shown to be upregulated in cold-acclimated D. melanogaster [46]. Consistent with this previous evidence of a role in environmental adaptation, both genes showed gains in our cold adapted (temperate) obscura group species.
In most cases, however, it is difficult to interpret the significance of the change in representation. Ect4 (Ectoderm-expressed 4, Fig 5, third panel), was one of number of genes that showed gains of expression in the melanogaster group. In this case, there was a gain in stage 2 representation in this lineage, while stage 5 was also represented in the outgroup. The role of Ect4 has largely been determined relative to its function in axon degeneration for the purposes of repair after injury, which is not directly relevant to its role in the early embryo. Recently, Ect4 was shown to be a target of DPP signaling in the embryo [47], but its precise function is unknown. We also found genes, such as DNAJ-H, that showed highly variable transcript levels across the phylogeny. Transcripts of this gene, which is part of the Heat Shock Protein 40 family of co-chaperones (essential factors in the Hsp70 chaperone cycle), may be represented at both stages or at either of the two stages, with no apparent pattern relative to the phylogeny.
These changes provide evidence that the early embryonic transcriptome is evolutionarily dynamic within Drosophilids. While the mRNA levels of some genes with key roles in development are highly conserved, a substantial subset of genes whose functions are not as well-understood show evidence of lineage-specific changes that are potentially adaptive.
Our results from previous sections suggest relatively rapid evolution of both maternal-only and zygotic-only transcripts (S3 Fig). Transcript levels of a subset of genes where the zygotic-only state is conserved across Drosophila species, however, are highly correlated (Fig 3), and a core set of zygotic-only genes with critical developmental roles are conserved with basal Diptera (Fig 4). To extend our analysis, we looked at two special situations: cases of species-specific representation in the early embryo, and the transcriptional profile of a small group of genes that were previously unannotated.
We were first interested in knowing whether there were instances of unique representation: cases where transcripts of a gene were present at a given stage in a species, while transcripts of one-to-one orthologs from all other species assayed were absent from that stage (S9 Table). We only considered genes with one-to-one orthologs in at least 12 of the 14 species. In order to eliminate cases were the transcript level of a gene happened to be slightly above our threshold of 1 in a single species, we focused on instances where a gene had a transcript level over three times the threshold (FPKM > 3) in a single species but an FPKM of less than 1 in all other Drosophila species (S6 Fig; S9 Table). We found many more cases of species-specific representation at the stage 2 than stage 5, for every species except D. melanogaster (which has low numbers of both). There are an exceptionally high number of cases of unique representation of maternal deposition in both D. virilis and D. erecta, while D. virilis also has an unusually high number of instances of unique representation at the zygotic stage (over twice as many as the next highest species, see S9 Table).
We also examined a different situation: unannotated genes identified during analysis by Cufflinks. These genes were not found in the reference annotations (see Methods for annotations used for each species) provided to the software. The number of unannotated genes ranged from a low of 280 in D. melanogaster, the most thoroughly annotated genome, to a high of 1905 in D. mauritiana (S10 Table). In many cases, unannotated genes may simply reflect the limitations of previously generated annotations. However, as has previously been found for de novo genes [48], unannotated non-D. melanogaster genes show low complexity, having significantly fewer exons (2–3 vs. the mean from all genes of 4.5–5.5; Wilcoxon rank-sum test, p = 7.4 x 10−6), as well as shorter exons (mean of 1240bp vs 2547bp, Wilcoxon test, p = 1.9 x 10−5) and introns (mean of 2105bp vs. 5256bp, Wilcoxon test, p = 5 x 10−8) than annotated genes (S7 Fig; S10 Table). It is possible, therefore, that some of the unannotated genes identified here are de novo genes. We note that these properties of de novo genes are shared with annotated zygotic genes, which also tend to have shorter and fewer exons and few or no introns [9,16,17]. The unnanotated genes expressed by the zygotic genome identified here, and reported in the statistics above, are less complex as compared to annotated zygotic genes (S7 Fig; S10 Table).
We found that unannotated genes were strongly biased towards being zygotic-only (Fig 6). When considering all genes that are transcribed during this early period of development, only 16–25% are zygotic-only, depending on the species (Fig 6, “all genes”). For the unannotated genes represented by transcripts in our dataset, 65–80% are zygotic-only. This is a 3- to 4-fold increase that is highly significant using a Fisher exact test (p<0.0001). The finding that these unannotated genes are more frequently zygotically transcribed than maternally deposited could indicate that early zygotic transcription evolves rapidly for putatively novel genes. This contrasts with our results for conserved genes (S6 Fig), discussed above, which were found to be more likely to evolve species-specific cases of maternal deposition than zygotic representation. In other words, a potentially novel gene of unknown function will often evolve early zygotic transcription, while a gene with conserved orthologs is more likely to evolve a unique instance of maternal representation.
The generation of multiple isoforms for a given gene, through mechanisms such as alternate promoters or alternative splicing, has been recognized as a critical form of genetic regulation [49]. Organisms deploy these isoforms in a context-dependent manner to meet the varied challenges of both development and adult physiology. Little is known, however, about the role of alternative isoforms in the maternal to zygotic transition.
In lineages such as Drosophila, flies that are adapted for rapid development, it has previously been shown that zygotic transcripts are shorter and have fewer introns than maternally deposited ones [9,16,17]. It is reasonable, therefore, to hypothesize that certain isoforms of the same gene may be better suited for maternal deposition or zygotic transcription. Specific examples of such cases are limited, however, and few studies have looked at the extent to which stage-specific isoforms are evolutionarily conserved.
Our data suggest that different isoforms may be used stage-specifically, presenting an additional layer of subtlety when comparing early embryonic transcriptomes across species. Consider, for example, the gene headcase (hdc), whose function has been characterized at later developmental stages, is expressed in all imaginal lineages, and is involved in trachea, head, and neuroblast development [50]. In the early embryo, transcript levels of orthologs of this gene are present at a higher level at stage 5 than stage 2 in both D. simulans and D. sechellia (Fig 7C). However, while both species show zygotic enrichment for the predominant isoform of its ortholog (Fig 7D, labeled “isoform 1” on the figure in both cases), the second most highly expressed isoform of the D. simulans ortholog (labeled “isoform 2”) is higher at the maternal stage, while the second most highly expressed D. sechellia isoform shows no significant difference in transcript levels between stages. A third isoform is also present at low levels in D. sechellia.
To look more closely at the isoforms that are present at each stage, we classified each isoform into one of six categories: maternal (M–only present at stage 2), zygotic (Z–only present at stage 5), predominantly maternal (PM–present at both stages, but at least twice as high at stage 2 than stage 5, with the differences being statistically significant), predominantly zygotic (PZ–present at both stages, but at least twice as high at stage 5 than stage 2, with the differences being statistically significant), maternal-zygotic (present at both stages, but predominant at neither) and not present at either stage (N). When looking across all isoforms, we find that in all species the mean exon number and the mean exonic length are slightly greater for the combined class of M and PM isoforms than for Z and PZ isoforms (S10 Table). The mean intronic length showed much starker differences, and was from 1.5 to 2-fold greater in the M/PM isoforms. This extends the findings from previous studies that zygotic genes are shorter and have fewer introns [9,16,17] to across the genus Drosophila, but our data shows much stronger support for differences in intron length than exon number or length between maternal and zygotic transcripts.
We looked specifically for genes that had at least one isoform that was maternal or predominantly maternal, and another that was zygotic or predominantly zygotic, and referred to these as the “alternative” (ALT) set (S10 and S11 Tables). The number of identified ALT genes varied from a low of 307 (~4% of genes) in D. ananassae to a high of 936 (~ 10% of genes) in D. virilis (Fig 7A, S10 Table), likely partially a consequence of statistical power. Regardless, the identification of hundreds of genes with stage-specific isoforms (Fig 7A) suggests that this could be an important regulatory strategy in early embryonic development.
In contrast to the patterns we observe for all genes, the ALT isoforms do not differ consistently in exonic number and exonic length (S8 Fig). Across isoforms, the number of exons and exonic length are always greater for isoforms of ALT genes than for those of all genes (S10 Table). These results may indicate that ALT gene products are required to be long and potentially complex to produce stage-specific isoforms, and that a lower proportion of intronic/exonic sequence may be selectively favored in isoforms that are zygotically transcribed.
To explore the functions of these genes with stage-specific or stage-enriched isoforms, we performed a GO analysis. When compared to the entire gene set, we found that these genes are moderately enriched in GO categories of genes that regulate development, morphogenesis and cell differentiation, in addition to sexual reproduction and gamete generation, among several other categories (S12 Table). This is consistent with the fact that one of the most celebrated cases of alternative splicing is in Drosophila sex determination, and other the key players in the pathway (Sex-lethal, transformer, doublesex, fruitless) being regulated through alternative splicing.
Regardless of their precise functions, the ALT gene state is frequently conserved over long evolutionary distances (S11 Table; S13 Table), with 67 genes showing stage-specific alternative isoforms in orthologs of at least 2/3 of the species. As one example (Fig 7B), consider the case of cap-n-collar (cnc). This transcription factor, which is essential for viability in D. melanogaster, has one or more long isoforms (11 exons, 6500–7000 bp of exonic sequence in D. melanogaster), one or more medium-length isoforms (5 or 7 exons, approximately 4000 or 5000 bp in D. melanogaster) and one or more short isoforms (3 exons, 3500 to 4000 bp in D. melanogaster). One-to-one orthologs were identified for 12/14 species (S11 Table), and all of these orthologs except D. mauritiana were classified as ALT, with the same pattern of at least one long maternal-only, one medium-length zygotic-only, and one short maternal-only isoform found in each case except D. miranda. The broad conservation of this pattern across 12 species spanning over 55 million years of evolution provides strong evidence for a functional role for both the multiple isoforms and their maternal or zygotic transcription.
The strong conservation of maternal transcript levels across Drosophila demonstrates that the mother’s vast RNA endowment is regulated with a precision that has withstood sixty million years of evolution. This conservation is all the more remarkable given the divergent ecologies and life histories of the species analyzed [19], and the extensive role played by post-transcriptional regulation of maternal transcripts [10–15]. Additional study will be necessary to determine protein abundance in each species, as post-transcriptional regulation may play an additional role in buffering the effects of differential transcript abundance. Since only a small minority of these transcripts are transcription factors, however, it is clear that the function of maternal deposition extends far beyond jumpstarting transcription in development. Theoretical comparisons of selective efficacy notwithstanding [7], maternal genes such as bcd that are recent evolutionary innovations [51,52] should be considered the exception rather than the rule.
From another perspective, however, the conservation of stage 5 transcript levels, while somewhat reduced relative to levels of maternal deposition (stage 2), is arguably even more remarkable. Transcript levels at stage 5 are a function of multiple processes: maternal RNA deposition that occurred during oogenesis, multiple embryonic degradation pathways, which themselves may be activated either maternally or zygotically [53,54], and early zygotic transcription. These processes must be tightly coordinated to generate zygotic levels that are reproducible not only between individuals but also between species, a finding that is all the more impressive given the fact that they are regulated in two separate genomes (that of the mother and the zygote).
Transcript levels of two categories of genes show more rapid divergence (S3 Fig): Maternal genes with transcripts that are entirely degraded by stage 5 (maternal-only), and zygotic genes with no maternal contribution (zygotic-only). These results support the hypothesis that the combination of maternal deposition and zygotic transcription is important for achieving the robust transcript levels that might be generally required for early embryonic development. However, when conducting pairwise comparisons in which only the set of genes where the specific stage-restricted state (maternal-only or zygotic-only) is conserved in both species is considered (Fig 3), there are distinct differences in the evolutionary trajectories of these two gene classes.
Maternal-only genes, those maternal genes that are degraded at the MZT and absent by stage 5 evolve quickly. Interspecific correlation coefficients for transcript levels of maternal-only genes drop off rapidly with evolutionary distance, even when only the genes where the maternal-only state is conserved in both species are considered (Fig 3). The number of shared maternal-only orthologs also decreases rapidly with evolutionary distance (S2 Fig), to the extent that there only three D. melanogaster–Aedes aegypti orthologs that are maternal-only in both clades. How we view these results depends on our interpretation of maternal-only genes. If both the deposition and the degradation are presumed to be functional, these genes would be cases where the transcripts are necessary in the very early, syncytial, embryo (or previously in oogenesis) but are strongly detrimental after cellularization. The finding that an ortholog of such a gene is not maternal-only in a related species would signify that either the gene was no longer necessary in the syncytium or that it was no longer detrimental at later stages. However, our results showing rapid divergence of transcript levels in cases where orthologs are maternal-only in both species, implies that if they have an early function they belong to a class of genes where tight regulation of transcript level may not be necessary. This would distinguish them from the vast majority of genes represented at stage 2, where transcript levels are highly conserved. Alternatively, maternal-only genes may represent developmental noise, with degradation as a method of compensating for the noise. During oogenesis, vast numbers of transcripts and proteins are deposited by the nurse cells in the egg, and it is possible that not all of them are necessary or beneficial to the embryo. Finally, we must consider that the post-transcriptional regulation of maternal transcripts might have a larger impact on the maternal-only transcripts, and that translational control may buffer against any phenotypic consequences of differences in transcript level. This would explain the variation in transcript levels for these genes, but not how quickly the maternal-only status is gained or lost. The sharp decrease in shared orthologs of maternal-only genes as evolutionary distance increases lends weight to the interpretation that maternal deposition may be noisy, and argues against post-transcriptional regulation having a larger role for the maternal-only class of genes as discussed above.
The transcript levels of most genes that are zygotic-only diverge rapidly in pairwise comparisons of interspecific orthologs (S3 Fig). However, if we limit our analysis to the smaller group of shared zygotic-only orthologs (genes that are zygotic-only in both species being compared; S2 Fig), we see a very different pattern (Fig 3), with transcript levels highly correlated, even among distantly related Drosophila species. Looking across much greater evolutionary distances, we were able to identify a core set of genes that are zygotic-only in both Aedes and early-diverging Drosophila species. These genes are strongly biased towards retaining the zygotic-only state across the Drosophila lineage. Our finding that this set is highly enriched in transcription factors with known functions in embryogenesis shows the power of evolutionary transcriptomics to identify key players in development. Functionally, we might expect these genes to be cases where zygotic-only expression is necessary since maternal deposition may be mechanistically deleterious prior to cellularization.
Across species, an overwhelming majority of unannotated genes have zygotic-only transcript representation in the early embryo, while only about a fifth to a quarter are zygotic-only in the annotated set. Many of the unannotated genes we identified display the hallmarks of newly-evolved genes, with relatively few isoforms and low expression levels, possibly suggesting that novel genes are biased towards zygotic expression (more work will need to be carried out on this gene set to determine whether the genes are indeed taxonomically-restricted). The idea that zygotic representation, on the whole, evolves more readily than maternal deposition is also consistent with our phylogenetic analysis, where a strong majority of the gains were stage 5. The maternal-only genes discussed in the previous section are an exception, as they consist of the minority of maternal transcripts that are entirely degraded by stage 5.
Conversely, maternally deposited RNA differs from its zygotically transcribed counterpart in that it can be used during the earliest syncytial stages of embryonic development. If these early stages are highly conserved, the evolution of new genes with a function in this period may rarely be necessary. Maternal deposition may instead evolve to increase the overall robustness of RNA levels during post-syncytial development. Or, perhaps the earliest stages of development where maternal mRNAs act require different gene products to undergo the conserved developmental processes in differing environments [55].
The phylogenetic pattern of evolution of transcript representation at the maternal and zygotic stages speaks to both a regulatory logic and to the relative roles of maternal and zygotic genomes in early development. We found that maternal-only genes hardly ever become zygotic-only (or vice versa) between closely related species. Instead, we see genes transcribed by both the maternal and zygotic genomes losing either maternal deposition or zygotic transcription, or stage-restricted genes gaining transcript representation at the other stage (e.g. a maternal-only gene gains zygotic transcription as well). This pattern can potentially be explained by the logic of regulation, since gaining or losing regulatory binding sites (or regulatory factors) at one stage may be a much more common occurrence than simultaneous evolution to both gain one set of binding sites or factors associated with transcription at one stage and lose binding sites or factors for the other stage. At the same time this provides evidence against compensatory evolution over these two stages as loss at one stage is not associated with gain at the other.
While gene number does not appear to correlate with any measure of organismal complexity, it is often claimed that isoform number might [49,56,57]. Alternative splicing is particularly common in vertebrates, although the record for isoform number is currently held by the Drosophila Dscam1 gene, where isoforms vary across individual neurons [58]. Drosophila also famously uses alternative splicing in sex determination [59].
In addition to tissue-specific and sex-specific alternative isoforms, stage-specific isoforms allow for complex temporal regulation [60]. In the early embryo, where both the mother and zygote provide RNA, the logic of utilizing alternative isoforms stems from the differing constraints of each of these players. For example, the rapidity of transcription is a strong selective pressure in the zygote where cell divisions are extremely rapid, leading to zygotic transcripts with fewer introns [16], while maternally supplied RNA is under no such constraint. Additionally, there are transcripts provided by both maternal and zygotic transcription where the maternal transcripts are to be selectively degraded at the MZT. Then the maternal genome could use isoforms with appropriate motifs to direct degradation in their untranslated regions (e.g. miRNA target sites), and the zygotic isoform, without these motifs, will persist.
Our discovery of hundreds of cases of alternative stage-specific isoforms (ASIs) in the 14 species we examined validates the potential utility for using different isoforms at these different stages. Furthermore, multiple cases of strong conservation of isoform structure (number of exons, overall exonic length) for stage 2 or stage 5 isoforms across 60 million years of evolution suggests functionality for this process. Future research will aim to determine if the localization of transcripts differs between the maternally and zygotically predominant isoforms, how these isoforms are differentially regulated, and whether they are functionally equivalent.
We have demonstrated the remarkable ability of two genomes to collaborate in the regulation of early development, leading to RNA transcript levels in the embryo that are highly stable over tens of millions of years of evolution. Yet, we also find considerable variation in the transcripts present during the earliest stages of development, despite expectations that early development is highly conserved. A large remaining question is how much of this variation is functional. It is plausible that some fraction of it is, and this would imply either that the processes of early development are not as conserved as commonly regarded, or that different complements of transcripts are necessary across different environments and genomes to maintain these conserved early developmental processes. Alternatively, it could be that the processes of early development are remarkably robust, and that considerable variation in transcript abundance or transcript representation may have minimal phenotypic consequences.
Single embryos were collected from 3–8 day old females of each species. Genome lines from the original 12 Genomes study [21] were used for 11 of the species (D. melanogaster, D. simulans, D. sechellia, D. yakuba, D. erecta, D. ananassae, D. pseudoobscura, D. persimilis, D. mojavensis, D. virilis). The lines for the additional species were as follows: D. mauritiana (Dmau/[w1]; 14021–0241.60), D. santomea (STO-CAGO 1402–3; 14021–0271.01), D. miranda (MSH-22). Embryos were dechorionated, and imaged on a Zeiss Axioimager, under halocarbon oil, to determine stage. Since embryos were collected from a large number of mothers, it is unlikely that multiple samples came from the same mother. Stage 2 and late stage 5 embryos were identified based on morphology. Stage 2 embryos were selected based on the vitelline membrane retracting from both the anterior and posterior poles, prior to when pole cells become visible. Late stage 5 embryos were chosen based on having completed cellularization, but not yet having gastrulated. Embryos were then removed from the slide with a brush, cleaned of excess oil, placed into a drop of Trizol reagent (Ambion), and ruptured with a needle, then moved to a tube with more Trizol to be frozen at -80° C until extraction. RNA and DNA were extracted as in the manufacturer’s protocol, with the exception of extracting in an excess of reagent (1 mL was used) compared to expected mRNA and DNA concentration [26–28].
Extracted total RNA from single embryos was treated with the TurboDNA-free kit (Ambion) prior to library construction. Embryo mRNA-Seq libraries were generated for at least 3 replicate individuals per stage and per species, producing a total of 68 stage 2 and 76 stage 5 libraries, or 144 overall. More detail about sampling and replication is available in S14 Table. mRNA-Seq libraries were constructed using TruSeq RNA sample preparation kits (Illumina), using standard protocols, and indexed to pool 12 samples (embryos) per lane. Library concentration was measured using the Qubit fluorometer (Life Technologies) and the qPCR-based Library Quantification kits (KAPA biosystems), and size was measured using the Bioanalyzer (Agilent). Libraries were sequenced on an Illumina HiSeq 2000 DNA Sequencer.
mRNA-Seq libraries were constructed using poly(A) selection. This creates a potential source of bias, as poly(A)-tail length is highly regulated during oogenesis and early embryogenesis, especially for maternally deposited transcripts [10–12,14,15] However, there is previous evidence that the use of oligo(dT)-based poly(A) selection does not bias the transcripts recovered. A study [61] measuring both poly(A) tail length and transcript level during this period of development found that the dynamic changes in poly(A) tail length had minimal impact on the transcript abundance levels measured. To determine if use of oligo(dT)-based poly(A) selection may have biased the transcript level measurements in our experiment, we examined our mRNA-Seq data from D. melanogaster relative to two datasets of poly(A) tail length during early development in the same species [61,62]. In comparing the poly(A)-tail length of sequenced transcripts in our experiment to the total distribution of poly(A)-tail lengths each of these two experiments, we find no difference in distributions (Wilcox test, p = 0.74 compared to [61], p = 0.99 compared to [62]). While we cannot rule out that we are recovering a biased subset of transcripts due to oligo(dT) enrichment, it seems unlikely that this method produces a substantial bias.
Genome and annotation files for the 12 previously sequenced species [21], downloaded from Flybase [63], are listed in S15 Table. The D. mauritiana genome and assembly [22] were accessed from a website maintained by the Christian Schlötterer lab at the University of Veterinary Medicine Vienna. The D. miranda genome assembly (DroMir_2.2) [20,64,65] was downloaded from Pubmed and an annotation file was provided by the Doris Bachtrog lab at the University of California, Berkeley. A draft version of the D. santomea genome (using the non-inbred STO4 line), based on data from David Stern’s lab [66] was provided by Peter Andolfatto (Columbia University). We used an annotation of this genome, generated for us by Kevin Thornton (University of California, Irvine), to help construct the phylogeny of the 14 species (see “Phylogenetic analysis”, below). However, since the D. santomea genome was generated using non-inbred flies, we decided to map our D. santomea reads using the flybase D. yakuba assembly and annotation.
Reads were pre-processed using Cutadapt [67] to remove adapter contamination. Mapping and differential expression analysis were carried out using the Tuxedo suite [68], which allows for the discovery of novel isoforms and genes. Briefly, Tophat2, [69,70], which leverages Bowtie2 [71,72], was used to align reads to the reference and discover new splice junctions for each replicate of each stage and species. An assembly for each stage and replicate was generated using Cufflinks, where upper-quantile normalization was performed using the–N option, and all assemblies for each species were merged with CuffMerge. Using the merged assembly, FPKM levels were calculated with Cuffnorm, and differential expression between stages was assayed using CuffDiff.
With the aid of the output files from CuffDiff and CuffNorm, gene FPKM levels were calculated using the total of the FPKM levels for all isoforms of each gene. Assignment of orthologs relied on an orthology table from Flybase (“gene_orthologs_fb_2014_06_fixed.tsv”) and the D. mauritiana and D. miranda annotations described above. A table was generated (S16 Table) consisting of all genes with one-to-one orthologs in at least 12 of 14 species. If there was no known D. melanogaster ortholog for a gene in a given species, or multiple orthologs, that entry in the table was left blank and was not included in any of the analyses.
Spearman rank sum correlation coefficients were calculated using the R statistical environment [73]. When comparing species, only calculated FPKM values for genes with one-to-one orthologs were considered. Correlation plots and hierarchical clustering were generated using the R heatmap2 package.
Data from an Aedes aegypti transcriptomic time course [35] were compared to our Drosophila results. We used FPKM values from the 0–2 hour time period (the earliest available, and the one which is most likely to represent maternal RNA) and the 8–12 hour period (which corresponds to the completion of cellularization in Aedes [38], approximately equivalent to stage 5 in Drosophila) in our comparison. The data was gleaned from Supplemental S9 Table in Akbari et al. 2013. Using Inparanoid [74] we found a total of 4619 genes from this dataset that had one-to-one orthologs in Drosophila melanogaster. We expanded S16 Table to include orthologs from Aedes, using the data we had gathered from the Akbari et al., 2013 study.
The phylogeny was constructed using 21 loci, listed in S17 Table. These loci were selected from a previously published list [75] of 250 candidate genes for a Drosphila phylogenetic analysis (chosen based on low codon usage bias, availability of one-to-one orthologs, and other criteria). In selecting the loci for our study, we were limited by the quality of the available D. santomea genome and annotation, which was generated using the non-inbred STO4 line. The loci were aligned using MUSCLE, and regions of low quality in the alignment were removed using trimal. A total of 35,829 base pairs were used in generation of the phylogeny, of which 19,176 were informative.
MrBayes3.2 was used to generate the phylogeny and infer ancestral stage 2 and stage 5 states for 8,075 genes with orthologs in 12 of the 14 species. Each gene was assigned a binary of state of 1 (present) if the FPKM level at a given stage was greater than or equal to the chosen threshold and 0 if it was below the threshold. If there was no one-to-one ortholog in a species, the state was considered unknown. Following previous transcriptomic analyses [35], an FPKM threshold of 1 was selected. Our dataset thus consisted of 16,150 states (two per species per gene) in addition to the 35,829 nucleotides of DNA. For the binary data, the frequency of state 1 was 0.715, the frequency of state 0 was 0.219, and the frequency of unknown states was 0.0656. MrBayes was run separately 12 times to reconstruct ancestral states on each internal node of the phylogeny (a MrBayes file for reconstructing the ancestral state for the obscura group is found in S1 Table). For the nucleotide data, we used a GTR model and a gamma distribution to model rate variation across sites. Each chain was run for 200,000 generations with a burn-in fraction of 0.25.
To study changes along the phylogeny, a gain in representation of a gene at a given stage was recorded if an inferred state changed from 0 (with at least 90% posterior probability) in the ancestral node to 1 (was at least 90% posterior probability) in the derived node, while a change from 1 to 0 (with 90% posterior probability in each case) was designated as a loss.
Unannotated genes were categorized as those given numbers by the Tuxedo suite but not found in the reference genome annotations (download from flybase or another source, as described above) that we provided to the software pipeline.
Isoforms were classified as maternal (M) if they were present at stage 2 and absent (below the FPKM threshold of 1) at stage 5, while they were considered zygotic (Z) if they were only present at stage 5. Isoforms that were present at both stages were filtered to select those that showed significant differences between stages (q value less than 0.05 in the CuffDiff output). From this set, those where the level at one stage was at least twice that of the other stage were categorized as primarily maternal (PM) if stage 2 was higher or primarily zygotic (if stage 5) was higher. All other isoforms that were present at the two stages were classified as maternal-zygotic (MZ).
Genes where at least one isoform was primarily maternal and the other was primarily zygotic show evidence of stage-specific isoform usage and were given the ALT classification. Using custom Perl scripts, these genes were identified and the extent of the conservation of the ALT state (across the 14 species) was assessed.
|
10.1371/journal.pgen.1004470 | Regulation of Feto-Maternal Barrier by Matriptase- and PAR-2-Mediated Signaling Is Required for Placental Morphogenesis and Mouse Embryonic Survival | The development of eutherian mammalian embryos is critically dependent on the selective bi-directional transport of molecules across the placenta. Here, we uncover two independent and partially redundant protease signaling pathways that include the membrane-anchored serine proteases, matriptase and prostasin, and the G protein-coupled receptor PAR-2 that mediate the establishment of a functional feto-maternal barrier. Mice with a combined matriptase and PAR-2 deficiency do not survive to term and the survival of matriptase-deficient mice heterozygous for PAR-2 is severely diminished. Embryos with the combined loss of PAR-2 and matriptase or PAR-2 and the matriptase partner protease, prostasin, uniformly die on or before embryonic day 14.5. Despite the extensive co-localization of matriptase, prostasin, and PAR-2 in embryonic epithelia, the overall macroscopic and histological analysis of the double-deficient embryos did not reveal any obvious developmental abnormalities. In agreement with this, the conditional deletion of matriptase from the embryo proper did not affect the prenatal development or survival of PAR-2-deficient mice, indicating that the critical redundant functions of matriptase/prostasin and PAR-2 are limited to extraembryonic tissues. Indeed, placentas of the double-deficient animals showed decreased vascularization, and the ability of placental epithelium to establish a functional feto-maternal barrier was severely diminished. Interestingly, molecular analysis suggested that the barrier defect was associated with a selective deficiency in the expression of the tight junction protein, claudin-1. Our results reveal unexpected complementary roles of matriptase-prostasin- and PAR-2-dependent proteolytic signaling in the establishment of placental epithelial barrier function and overall embryonic survival.
| Development of mammalian embryos is dependent on an efficient exchange of nutrients, oxygen, and waste products between the mother and the embryo. The interface between the two systems is provided by the placenta in a form of a specialized epithelium that both facilitates the transport of molecules between the mother and the embryo and screens the substances that can pass between the maternal and fetal tissues. We now show that two independent signaling pathways that include the serine proteases, matriptase and prostasin, and a G protein-coupled receptor PAR-2, are critical for the establishment of a functional feto-maternal interface by specifically regulating the barrier properties of the placental epithelium. Because aberrant formation of epithelial barriers is an underlying feature of a great variety of human developmental abnormalities, the identification of the two protease-dependent signaling pathways critical for the barrier formation in embryonic tissues may help pinpoint molecular mechanisms involved in the etiology of these conditions.
| The development of eutherian mammals requires an efficient exchange of nutrients, oxygen, ions, hormones, and waste products between the maternal and fetal blood. In humans and mice, a functional feto-maternal interface is established in the placenta by the formation of a complex embryonic vascular tree that is submerged in interstitial space filled with maternal blood [1], [2]. Separating the maternal and fetal circulation is a specialized embryo-derived epithelium that functions both to facilitate the transport of molecules between the mother and the embryo, and as a barrier to screen which substances can pass between the maternal and fetal tissues and which cannot. In mice, the epithelium resides in a histologically distinct region of the placenta, termed the labyrinth, and consists of two layers of differentiated polynuclear syncytiotrophoblasts surrounding the fetal vessels and an underlying layer of mononuclear cytotrophoblasts that are in direct contact with the maternal blood [2]. The thickness of the labyrinth, the degree of vascular branching, and the level of expression of transporter proteins within the labyrinth epithelium are the chief determinants of the efficiency of nutrient transfer to the embryo [3].
The ability of the epithelia to restrict free movement of water, solutes, and larger molecules through the interstitial space between the individual epithelial cells is critical for tissue compartmentalization and protection against chemical damage, infection, dehydration, or heat loss [4], [5]. Establishment of paracellular transport barriers across the epithelial layers is primarily achieved by formation of several types of specialized cell-cell junctions that include desmosomes, adherens, and tight junctions [6]. Disruption of the junctional complexes severely compromises epithelial barrier function and is frequently linked to disease in both mice and humans [5], [7]. In the developing embryo, establishment of a functional feto-maternal barrier is critical to protect the fetus from toxins and infection by blood-born pathogens present in maternal circulation during pregnancy, as well as potential ion imbalance, unchecked diffusion of maternal hormones, or an attack by maternal immune system [8]–[12].
Matriptase is a trypsin-like cell surface-associated serine protease that plays an essential role in the homeostasis of a variety of mouse and human epithelia [13]–[15]. Mice with a complete loss of matriptase function die perinatally due to a defect in epidermal barrier function, leading to a fatal dehydration [16], [17]. Likewise, tissue-specific genetic ablation of matriptase revealed a role in epithelial barrier function in a number of other organs, including oral cavity, salivary gland, small intestine, and colon, suggesting a key role of this protease in the establishment of functional epithelial barriers of both single and multi-layered epithelia [18]. Despite a widespread expression in a number of embryonic and extraembryonic epithelia, matriptase, however, is not required for term development in humans and in most mouse strains ([13], [14], [16] and unpublished data). However, the regulation of matriptase activity is essential for the successful completion of the embryonic development, as documented by severe, matriptase-dependent, defects in placental development, neural tube closure, and overall embryonic survival in mice lacking either of the two endogenous serine protease inhibitors, HAI-1 and HAI-2 [19], [20].
Protease-activated receptors (PARs) are a subfamily of the heptahelical G protein-coupled receptors (GPCRs) activated by a proteolytic cleavage within the N-terminal extracellular region, unmasking new amino-terminal residues that then serve as tethered ligands to activate the receptors [21]–[23]. Both in mice and in humans, the subfamily consists of four members, PAR-1 through PAR-4. Mice carrying loss of function mutations in PAR-1 or PAR-2 genes suffer from partial embryonic lethality although born PAR-1- and PAR-2-deficient animals generally show a normal postnatal development and survival [24], [25]. Furthermore, the combined loss of PAR-1 and PAR-2 leads to more than 60% decrease in pre-term and 90% decrease in pre-weaning survival, and to an array of developmental defects that include neural tube defects, edema, and an enlarged pericardium, indicating the requirement of PAR-mediated signaling for the completion of normal embryonic development [26]. Whereas PAR-1, -3, and -4 are activated in vivo by the secreted serine protease thrombin and appear to primarily function as part of the blood coagulation cascade, PAR-2 is expressed by a majority of epithelial and endothelial cells and can be activated by a number of soluble and membrane-anchored proteases with trypsin-like activity, but is not efficiently activated by thrombin [23]. Matriptase, in particular, has consistently been shown to efficiently stimulate activation of PAR-2 in a variety of cell-based assays. This, together with a highly overlapping pattern of expression in majority of epithelia, suggested that at least some of the functional effects of matriptase in vivo are mediated by PAR-2 and its downstream effectors [26]–[28].
In this study, we report the unexpected observation that matriptase and PAR-2 act during embryogenesis via functionally independent and redundant proteolytic signaling pathways. We show that mouse embryonic development and survival is strongly dependent on the genetic dosage of matriptase and PAR-2 and that the combined loss of both molecules leads to a complete embryonic lethality. Our results reveal unexpected complementary and essential roles of matriptase/prostasin- and PAR-2-dependent signaling in the establishment of placental epithelial tight junctions.
Matriptase has consistently been shown to efficiently stimulate activation of PAR-2 in cell-based assays, leading to a conclusion that at least some of its physiological functions may be mediated by PAR-2 [26]–[28]. However, our recent work tentatively suggested that in the context of mouse embryonic development, matriptase and its partner protease prostasin may act independently of PAR-2, based on the findings that: (i) developmental defects observed in mice lacking endogenous matriptase inhibitor, HAI-2, can be rescued by matriptase or prostasin deficiency but not by PAR-2 deficiency, and (ii) a genetic inactivation of matriptase fails to reproduce phenotypes associated with PAR-2 deficiency in mice lacking PAR-1 [29]. The HAI-2; PAR-2 double-deficient mice (Spint2−/−;F2rl1−/−) used in that study were mostly generated by interbreeding of HAI-2; PAR-2 double-heterozygous and HAI-2; PAR-2; matriptase triple-heterozygous mice (Spint2+/−;F2rl1+/−×Spint2+/−;F2rl1+/−;St14+/−). Unexpectedly, a detailed review of the genotype distribution among the offspring from these breeding pairs at weaning showed that the survival of PAR-2-deficient mice was strongly dependent on the number of active alleles of matriptase. Thus, in offspring carrying two active alleles of matriptase (St14+/+, matriptase wildtype) about 90% of the F2rl1−/− mice survived until weaning (Figure 1A and Table S2), consistent with the previously reported 10–30% pre-term lethality among the F2rl1−/− mice [25], [26]. However, in the offspring that lacked one functional allele of matriptase (St14+/−, matriptase heterozygous), the survival of the F2rl1−/− mice decreased to less than 40% compared to the expected Mendelian distribution (Figure 1A and Table S2). To further analyze this dramatic loss of survival and to test the survival of PAR-2-deficient mice in the complete absence of matriptase, we established a new cohort of breeders double-heterozygous for both matriptase and PAR-2 (F2rl1+/−;St14+/−). Inspection of the allele distribution in newborn offspring (Figure 1B) confirmed the highly significant increase in embryonic lethality (Figure 1C) among F2rl1−/− animals carrying one active allele of matriptase (F2rl1−/−;St14+/−, 40% pre-natal survival), compared to matriptase wildtype animals (F2rl1−/−;St14+/+, 90% pre-natal survival). Furthermore, analysis of the 272 newborn mice from the F2rl1+/−;St14+/− breeding pairs did not identify any F2rl1−/−;St14−/− pups, indicating a complete pre-term lethality of F2rl1−/− mice in the absence of matriptase (Figure 1B and 1C, Table S3). Similarly, embryonic survival of matriptase-deficient mice was strongly dependent on the gene dosage of the PAR-2 gene, as indicated by73%, 35%, and 0% pre-term survival of St14−/− mice carrying two, one, or no active alleles of the PAR-2 gene, respectively (Figure 1B and C, Table S3).
Activation of matriptase during embryonic development was recently shown to be dependent on the activity of the GPI-anchored serine protease prostasin (PRSS8/CAP1), indicating that the two proteases are part of the same proteolytic signaling pathway [29]. Consistent with this finding, analysis of the 247 newborn offspring from the Prss8+/−;F2rl1+/− breeding pairs also showed a diminished survival of mice with low combined dosage of Prss8 and F2rl1 genes. Specifically, the pre-term survival of PAR-2-deficient mice decreased from 93% in prostasin wildtype animals (Prss8+/+;F2rl1−/−) to 62% and 0% in mice carrying one (Prss8+/−;F2rl1−/−) or no (Prss8−/−;F2rl1−/−) active alleles of prostasin (Figure 1D, Table S4). These data show that the pre-term survival of mice is completely dependent on either PAR-2- or matriptase/prostasin-mediated proteolytic signaling.
We next investigated the offspring of interbred F2rl1+/−;St14+/−×F2rl1+/−;St14+/− and F2rl1−/−;St14+/−×F2rl1+/−;St14+/− mice at various stages of embryonic development to further characterize the lack of term survival in mice double-deficient for PAR-2 and matriptase. Analysis of embryos extracted between E10.5–E12.5 revealed normal distribution of F2rl1 and St14 alleles, and nearly all of the F2rl1−/−; St14−/− embryos extracted before E12.5 were alive and appeared normal, indicating that the pre-implantation and the early post-implantation development and survival were not affected by the combined loss of F2rl1 and St14 gene function (Figure 1E). After E12.5, however, the survival of the F2rl1−/−;St14−/− animals dramatically decreased to 59% at E13.5, 14% at E14.5, and no living F2rl1−/−;St14−/− embryos were identified at or after E15.5 (Figure 1E and Table S5). Similarly, analysis of the offspring from F2rl+/−;Prss8+/− breeding pairs found an expected number of living Prss8−/−;St14−/− embryos at or before embryonic day 12.5, followed by a dramatic decrease in survival to 47% at E13.5, 17% at E14.5, and no surviving double-deficient embryos identified at and after E15.5 (Figure 1F and Table S6).
Matriptase, prostasin, and PAR-2 all are membrane-anchored proteins and are believed to function primarily at the surface of the expressing cells. To help us identify embryonic structures potentially affected by the combined loss of matriptase/prostasin- and PAR-2-dependent activities, we looked for cells simultaneously expressing St14, Prss8, and F2rl1 in mid-gestation embryos using the publicly available Eurexpress transcriptome atlas database [30]. All three genes were detected in a number of embryonic tissues at E14.5, showing an extensive co-expression in developing epithelia of skin, oral and nasal cavities, salivary gland, lungs, kidneys, and gastrointestinal tract (Figure 2A–C). Surprisingly, despite the widespread co-expression, a detailed inspection of the living E12.5–14.5 F2rl1−/−;St14−/− or F2rl1−/−;Prss8−/− embryos failed to identify any obvious developmental abnormality. The double-deficient mice did not develop any signs of internal bleeding, edema, or enlarged pericardium typical for many mouse strains exhibiting mid-gestational lethality (Figure 2D), and did not differ in size or total body weight from their wild-type littermate controls (Figure 2D and 2E). Furthermore, histological analysis of living E11.5–14.5 F2rl1−/−;St14−/− embryos did not reveal any obvious defects in the development of any of the epithelia with detectable levels of matriptase, prostasin, and PAR-2 gene expression (see above) or any other major organs, which are typically affected in mouse strains that exhibit mid-gestational lethality, including liver [31] (Figure 2F–I′).
Failure to identify any specific defect associated with a combined loss of matriptase and PAR-2 in the embryo proper prompted us to examine development of other tissues of fetal origin. Consistent with our previous findings, immunohistochemical analysis revealed strong expression of matriptase and prostasin in the epithelial compartment of the placental labyrinth at midgestation (Figure 3A and 3B) [19], [20], [29]. To address the expression of PAR-2 in the absence of suitable antibodies, we employed a previously generated knock-in mouse strain that carries a β-galactosidase reporter construct under the control of the endogenous promoter of F2rl1 gene [32]. Analysis of the distribution of β-galactosidase activity in mouse placental tissues at E12.5 revealed that, similar to matriptase and prostasin, expression of the F2rl1 gene is found predominantly in the epithelium of the chorion and in the syncytiotrophoblast layer lining fetal endothelium within the labyrinth (Figure 3C). These data identify labyrinthine epithelium as the placental population most likely to be affected by a combined loss of matriptase/prostasin and PAR-2.
Initial histological evaluation indicated that all embryo-derived placental structures, including layers of trophoblast giant cells, spongiotrophoblasts, placental labyrinth, and allantoic mesenchyme were all formed in the placentas of F2rl1−/−;St14−/− embryos and there were no signs of placental edema (Figure 3D and 3E). Furthermore, immunohistochemical staining for the endothelial cell marker CD31/PECAM1 demonstrated the presence of a highly branched fetal vasculature in close proximity to maternal blood lacunae within the labyrinth of F2rl1−/−;St14−/− placentas (Figure 3F–3I), indicating that the combined loss of PAR-2 and matriptase does not block any of the major morphogenetic processes involved in placental development. However, a more detailed morphometric analysis revealed noticeable quantitative changes in the development of the labyrinth structure. The overall thickness of the labyrinth, defined as the maximum perpendicular distance between the undifferentiated chorionic epithelium and the labyrinth supporting spongiotrophoblast layer, was reduced by 10–15% in placentas from F2rl1−/−;St14−/− embryos at E12.5 and E13.5 (P<0.05, Figure 3J). This was also reflected in a 15% reduction of total volume of the labyrinth in F2rl1−/−;St14−/− placentas at E13.5, as determined by the Cavalieri stereological technique (Figure S1). More importantly, the complexity of the fetal-derived vasculature within the labyrinth that mediates bidirectional transport of molecules between the mother and the embryo was substantially reduced, as indicated by more than 40% decrease in the number of CD31/PECAM1-positive fetal capillary profiles found within the labyrinth area of double-deficient placentas at E12.5 and E13.5 (Figure 3F–3I and Figure 3K, P<0.01 and 0.001, respectively, Student's t-test, two-tailed). These data indicate that the combined loss of matriptase-prostasin- and PAR-2-dependent signaling interferes with the development of the feto-maternal interface.
To test the possibility that the loss of viability of F2rl1−/−;St14−/− mice at midgestation results from a defect in the development of placental rather than embryonic tissues, we next investigated survival of PAR-2-deficient mice with a specific inactivation of matriptase in the embryo proper. To that end, PAR-2 heterozygous breeders carrying conditional St14 allele (F2rl1+/−;St14fl/fl) were crossed to F2rl1+/−;St14+/− animals expressing Cre recombinase under the control of the endogenous Meox2 promoter. Meox2 expression is ubiquitous in mouse embryonic tissues as early as at E6.5, but is missing from any of the extraembryonic structures, including the placenta [33]. As a result, whereas both the embryos and the placentas of the Meox2-Cre+;F2rl1−/−;St14−/fl offspring would be devoid of PAR-2 expression, only the embryos will also be rendered matriptase-deficient due to the recombination of the remaining floxed St14 allele (Figure 4A). A potential survival of Meox2-Cre+;F2rl1−/−;St14−/fl embryos would therefore demonstrate a critical contribution of placental matriptase to the embryonic survival of PAR-2 and matriptase double-deficient mice, whereas lack of survival would be an indication that embryonically-expressed matriptase determines viability of these animals.
Western blot analysis of protein lysates from the E13.5 embryos and placentas showed a complete loss of expression of matriptase protein in the Meox2-Cre+;F2rl1−/−;St14−/fl mid-gestation embryos, whereas the expression of the protease in the corresponding placentas was easily detectable (Figure 4B), thus confirming the efficient, embryo-specific inactivation of the St14 conditional allele. However, despite being matriptase- and PAR-2-deficient, Meox2-Cre+;F2rl1−/−;St14−/fl embryos survived to term and were detected among the newborn offspring from the F2rl1+/−;St14fl/fl×Meox2-Cre+;F2rl1+/−;St14+/− breeding pairs in the expected ratio (Figure 4C). These mice recapitulated all of the phenotypes previously observed in matriptase knockout mice, including dry skin, and lack of whiskers, and showed no residual expression of matriptase in newborn tissues (Figure 4D–4F). These findings document that the expression of matriptase in the embryo proper is dispensable for the pre-term development and survival of PAR-2; matriptase double-deficient mice.
The labyrinthine epithelium in mice is a principal component of the feto-maternal barrier, serving both to transport and to screen the substances passing between the maternal blood and the fetal tissues. To test whether the loss of matriptase and/or PAR-2 expression interferes with either of these functions, we next investigated the rate of molecular transport across the placental epithelium. Many nutrients, including glucose, are carried across the placenta by facilitated diffusion using an array of cell-surface transporters expressed by placental epithelial cells [34]. The efficiency of glucose uptake by the embryo can therefore be used as a quantitative measure of placental nutrient transport capacity [35]. To that end, radioactively-labeled 3-O-[methyl-14C]-D-glucose was injected into the bloodstream of pregnant females at E12.5 or E13.5, followed by embryo extraction after 2 minutes and measurement of fetal uptake of the [14C] label. Consistent with the notable underdevelopment of placental labyrinth layer in matriptase/PAR-2 double-deficient animals (see above), uptake of glucose by the F2rl1−/−;St14−/− embryos decreased by about 20% compared to wild-type littermate controls (P<0.05, Figure 5A). However, this relatively modest reduction in the transport rate together with a lack of any of the typical indications of embryo malnutrition, such as reduced size, gross underdevelopment, or enlarged pericardium, did not conclusively demonstrate an insufficient placental transport as the primary cause of embryonic lethality in the double knockouts.
In addition to the transport of gases and nutrients, placenta also fulfills a critical function of a barrier protecting the fetus against unchecked diffusion of substances, including ions, hormones, components of maternal immune system, or blood-born pathogens from maternal to fetal tissues. To test the ability of the placental epithelium to act as a barrier against passive diffusion, pregnant females were injected at E12.5 or 13.5 with the radioactively-labeled polysaccharide inulin, which can cross epithelial layers solely by a paracellular route due to lack of specific transporters [35], [36]. Interestingly, measurement of uptake of [14C] label revealed dramatic differences in inulin diffusion into the embryos of different genotypes. Thus, genetic inactivation of either matriptase or PAR-2 alone resulted in 247% and 145% increase in placental permeability (Figure 5B). This defect was further exacerbated in matriptase-deficient animals heterozygous for PAR-2 (F2rl1+/−;St14−/−, 396% increase) or PAR-2-deficient animals heterozygous for matriptase (F2rl1−/−;St14+/−, 351% increase) (Figure 5B). Finally, the highest level of inulin diffusion was observed in double-deficient embryos (F2rl1−/−;St14−/−, 589% increase, Figure 5B). These data indicate a dramatic loss of feto-maternal barrier function in mice with decreased levels of matriptase and/or PAR-2, the extent of which precisely correlates with the overall embryonic survival in these animals.
Loss of placental barrier can, in principle, be a result of either: (i) an indirect effect due to an impairment of terminal differentiation and subsequent barrier acquisition within the labyrinthine epithelium, or (ii) a direct alteration of the properties of the barrier-forming epithelial cell-cell junctions in the terminally differentiated trophoblasts. To distinguish between the two possibilities we first analyzed the expression of two of the principal regulators of labyrinth differentiation, transcription factor Glial Cell Missing (GCM) a and syncytin, a fusogenic retroviral protein that mediates terminal differentiation of placental cytotrophoblasts into multinucleated syncytium [37], [38]. Analysis of placental tissues from E12.5 matriptase/PAR-2 double-deficient animals and their wild-type littermate controls did not reveal any obvious changes in the expression level of either of the two markers (Figure 5C), indicating that the process of labyrinth differentiation is not compromised by a combined loss of the two proteolytic pathways. To assess the formation of barrier-promoting epithelial cell-cell junctions, we next analyzed the levels of claudin-1 and -2, cadherins, and desmogleins 1 and 2, as the major structural components of tight junctions, adherens junctions, and desmosomes, respectively. Whereas the expression of cadherins and desmogleins was not affected by the combined loss of matriptase and PAR-2, the expression of tight junction marker claudin-1 but not claudin-2 was severely diminished (Figure 5D and 5E). Matriptase or PAR-2 single-deficient placentas did not exhibit significant changes in claudin-1 expression (Figure S2). These findings suggest that a combined activity of matriptase/prostasin and PAR-2 may regulate permeability of placental labyrinth by specifically altering properties of epithelial tight junctions.
Inadequate placental development may lead not only to miscarriage, but also to poor health and increased risk of chronic disease in born individuals [39]. The current study for the first time reveals that the matriptase-prostasin system plays an important role in placental morphogenesis. Thus, mice with placental ablation of matriptase displayed impaired placental barrier formation, which was synergistically exacerbated by haploinsufficiency for or by complete deficiency of PAR-2. A role for the matriptase-prostasin system in placental morphogenesis would be anticipated because of the temporally and spatially coordinated expression of both membrane-anchored serine proteases and their cognate inhibitors, HAI-1 and HAI-2, in the developing placenta [19], [20], [29], [40] and the need for strict posttranslational regulation of both proteases for normal placentation to occur [19], [20], [29]. However, both matriptase-deficient humans and mice have been reported to complete embryonic development [13], [14], [16], and the role of this proteolytic cascade in development therefore has remained unclear. Because of the coordinated expression of matriptase, prostasin, HAI-1 and HAI-2 in several other developing epithelia, it now appears likely that matriptase/prostasin-initiated cellular signaling may also contribute to the development of several other organ systems in the embryo proper, acting in a redundant manner with other proteolytic systems that provide developmental backups.
Numerous studies have demonstrated a role of PAR-2 signaling, triggered by proteases of the coagulation cascade and by other trypsin-like serine proteases, in promoting both epithelial and endothelial barrier leakage by increasing paracellular permeability [41]–[45]. Our study is the first to identify a context in which PAR-2 signaling promotes epithelial barrier formation, rather than barrier leakage, thus adding another property to this multifunctional receptor. The observed specific deficiency in the expression of tight junction protein claudin-1 could suggest that the two proteolytic pathways directly promote barrier formation of terminally differentiated placental epithelium rather than doing it indirectly by providing morphogenetic signaling enabling the terminal differentiation of placental epithelium and the subsequent paracellular barrier acquisition between terminally differentiated epithelial cells. However, it should be noted that it is not yet clear whether the demise of the matriptase/PAR-2 double-deficient animals can be attributed to the observed defect in the expression of claudin-1, especially considering that claudin-1-deficient mice can complete embryonic development, at least in some mouse genetic backgrounds [46]. A previous report by Buzza and coworkers [47] suggested that matriptase regulates intestinal barrier function in a claudin-2-dependent manner. Our current analysis did not reveal any differences in claudin-2 expression in matriptase/PAR-2 double-deficient placental tissues, suggesting that different mechanisms of proteolysis-mediated regulation of tight junction function may be employed by different tissues and/or at different stages of mouse development.
Previous studies have shown that the matriptase-prostasin system is a potent activator of PAR-2 in a variety of cell-based assays [26]–[28], and PAR-2, matriptase, prostasin, HAI-1 and HAI-2 all are co-expressed in multiple developing and adult epithelia. Our genetic demonstration that the matriptase-prostasin system and PAR-2 are components of two independent and functionally redundant proteolytic pathways during placental morphogenesis therefore is unexpected and raises two unanswered questions: First, what is the identity of the placental protease(s) that activates PAR-2 in the absence of matriptase or prostasin? Second, what is the identity of the non-PAR-2 substrate(s) that is cleaved by the matriptase-prostasin cascade to promote placental morphogenesis? Regarding the former, PARs are activated by members of the S1 family of trypsin-like serine proteases, but not by other serine proteases or by other protease classes [48]–[50]. Previous transcript analysis of developing mouse embryos (in the context of neural tube closure) revealed the expression of a surprisingly large number of secreted and membrane-anchored serine proteases [20], [26]. The identification of the protease(s) that activates PAR-2 during placental morphogenesis by using a candidate epistasis analysis approach therefore may be a challenging undertaking. Regarding the latter question, the matriptase-prostasin cascade has been proposed to execute the cleavage of a wide variety of substrates, including other proteases (urokinase plasminogen activator, kallikrein-5 and -7, stromelysin-1), growth factors (hepatocyte growth factor, platelet-derived growth factor, macrophage-stimulating protein) tyrosine kinases (angiopoietin receptor, subtractive immunization M(+)HEp3 associated 135 kDa protein/CUB domain-containing protein-1, epidermal growth factor receptor) and more [15], [28], [51]–[62], reviewed in [18]. Identification of the specific proteolytic targets for the matriptase-prostasin cascade in the placenta therefore may prove difficult.
In summary, our study is the first to demonstrate a function of matriptase/prostasin- and PAR-2-dependent protease signaling in placental morphogenesis and to show that the two pathways act in an independent and functionally redundant manner to promote formation of the placental barrier.
All experiments were performed in an Association for Assessment and Accreditation of Laboratory Animal Care International-accredited vivarium following Standard Operating Procedures. The studies were approved by the NIDCR Institutional Animal Care and Use Committee. All studies were littermate controlled. Matriptase-deficient (St14−/−), PAR-2-deficient (F2rl1−/−), and Meox2-Cre transgenic mice have been described previously [16], [18], [33], [63]. Prostasin-deficient (Prss8−/−) mice were generated by standard blastocyst injection of C57BL/6J-derived embryonic stem cells carrying a gene trap insertion in the Prss8 gene (clone IST10122F12, Texas A&M Institute for Genomic Research, College Station, TX). Details on mouse generation will be published separately. Ear or tail clips of newborn or two week old mice were subjected to genomic DNA extraction and genotyped by PCR (see Table S1 for primer sequences).
In situ hybridization data were retrieved from the Eurexpress transcriptome atlas database [30]. Gene names St14, Prss8, and F2rl1, respectively, were individually searched to determine expression at embryonic day 14.5.
Breeding females were checked for vaginal plugs in the morning and the day on which the plug was found was defined as the first day of pregnancy (E0.5). Pregnant females were euthanized in the mid-day at designated time points by CO2 asphyxiation. Embryos were extracted by Caesarian section and the individual embryos and placentas were dissected and processed. Only living embryos with detectable heartbeat were used for further analysis. Tail clips or yolk sacs of individual embryos were washed twice in phosphate buffered saline, subjected to genomic DNA extraction, and genotyped by PCR (Table S1). Newborn pups were euthanized by CO2 inhalation at 0°C. For histological analysis, the embryos and newborn pups were fixed for 24 h in 4% paraformaldehyde (PFA), processed into paraffin, sectioned, and stained with hematoxylin and eosin (H&E), or used for immunohistochemistry as described below. To evaluate the development of embryonic and placental tissues, H&E-stained midline sagittal embryonic and the midline cross sections from at least three F2rl1−/−;St14−/− and three corresponding littermate controls were inspected by a certified pathologist for possible abnormalities.
Placental tissues from living E12.5 and E13.5 embryos were extracted and processed into paraffin as described above. To analyze the overall thickness of the placental labyrinth, a single midline cross-section at the level of the umbilical cord was stained with hematoxylin & eosin and the maximum perpendicular distance between the undifferentiated chorionic epithelium and the labyrinth-supporting spongiotrophoblast layer was measured under a light microscope by an investigator blinded to the genotypes of individual embryos. The volume of the placental labyrinth was estimated using Cavalieri's principle. Briefly, PFA-fixed placental tissue was embedded in paraffin, cross sections 200 µm apart covering the entire volume of the tissue were stained with h&e, scanned using ScanScope system (Aperio Technologies, Vista, CA), and the area of the labyrinth was measured on each section using Image J 1.46r software (National Institutes of Health, MD). The total volume of the labyrinth was then estimated as the sum of partial volumes calculated as area of the labyrinth on each section multiplied by 200 µm.
To evaluate branching of the fetal vasculature in the E12.5 and E13.5 placentas, a single midline cross section of each placenta at the level of the umbilical cord was immunostained with an anti-CD31 antibody (see below), followed by the manual counting of individual profiles of CD31-stained vessels within the placental labyrinth.
Antigens from 5 µm paraffin sections were retrieved by incubation for 10 min at 100°C with 1 mM EDTA, pH 8.0 for CD31 staining, or by incubation for 20 min at 100°C in 0.01 M sodium citrate buffer, pH 6.0, for all other antigens. The sections were blocked with 2% bovine serum albumin (fraction V, MP Biomedicals, Solon, OH) in phosphate-buffered saline (PBS), and incubated overnight at 4°C with 2 ug/ml rabbit anti-human CD31 (Santa Cruz Biotechnology, Santa Cruz, CA), or sheep anti-human matriptase (R&D Systems, Minneapolis, MN), primary antibodies. Bound antibodies were visualized using biotin-conjugated anti-rabbit, or -sheep secondary antibodies (Vector Laboratories, Burlingame, CA) and a Vectastain ABC kit (Vector Laboratories) using 3,3′-diaminobenzidine as the substrate (Sigma-Aldrich, St. Louis, MO). All microscopic images were acquired on an Olympus BX40 microscope using an Olympus DP70 digital camera system (Olympus, Melville, NY).
Placentas were extracted from embryos at E12.5 or E13.5. The embryonic portion of each placenta was manually separated from maternal decidua using a dissection microscope. The tissues were then homogenized in ice-cold 62.5 mM Tris/HCl, pH 6.8; 2% SDS; 10% glycerol buffer supplemented with 1× protease inhibitor cocktail (Sigma-Aldrich) and incubated on ice for 10 min. The lysates were centrifuged for 10 min at 20,000 g at 4°C to remove tissue debris and the supernatant was used for further analysis as described below.
The protein concentration in cleared lysates from embryonic, placental, and newborn tissues was determined by BCA assay (Pierce, Rockford, IL). 80 µg of total protein was loaded on 4–12% reducing SDS-PAGE and analyzed by Western blotting using a polyclonal sheep anti-human matriptase (R&D Systems), mouse anti-cow desmoglein 1 and 2 (Fitzgerald Industries International, Acton, MA), mouse anti-human Gcm1 (Abcam, Cambridge, MA), rabbit anti-human syncytinA (SantaCruz Biotechnology), rabbit anti-human pan-cadherin, rabbit anti-human GAPDH (both Cell Signaling Technology, Danvers, MA), rabbit anti-human claudin-1, and mouse anti-human claudin-2 (both Invitrogen, Carlsbad, CA) primary antibodies, and goat anti-mouse, mouse anti-rabbit (both DakoCytomation), or donkey anti-sheep (Sigma-Aldrich) secondary antibodies conjugated to alkaline phosphatase. Alkaline phosphatase activity was then visualized using nitro-blue tetrazolium and 5-bromo-4-chloro-3′-indolyphosphate substrates. Where indicated, protein signal quantification was performed using ImageJ 1.46r software.
Pregnant female mice from breeding pairs heterozygous β-galactosidase-tagged F2rl1 knock-in mice were euthanized at E12.5 by cervical dislocation. Embryos were extracted by Caesarian section, and the individual embryos and placentas were dissected and placed in 2% PFA with 0.2% glutaraldehyde. Fixed tissues were stained overnight at room temperature in 1 mg/ml X-gal, 5 mM potassium ferricyanide, 5 mM potassium ferrocyanide, 2 mM magnesium chloride and 0.02% NP-40 in PBS. The tissues were post-fixed for 16 h in 4% PFA, embedded in paraffin, and sectioned. The sections were counterstained with nuclear fast red. All microscopic images were acquired on a Zeiss AxioImager Z1 light microscope (Carl Zeiss AG, Gottingen, Germany) using a Qicam FAST1394 digital camera from Qimaging.
100 µl of PBS containing 1 uCi of 3-O-[methyl-14C]-D-glucose or [carboxyl-14C]-inulin (both Perkin-Elmer, Hanover, MD) was injected into the tail vein of pregnant female mice from St14;F2rl1 double-heterozygous breeding pairs at 12.5 or 13.5 days of gestation. The mice were euthanized two min later by CO2 inhalation, the embryos were extracted as described above, weighted, and lysed overnight at 60°C in 500 µl of 2% potassium hydroxide solution. The lysate was then neutralized with 50 µl of 14 M hydrochloric acid, mixed with 10 ml of ScintiSafe scintillation liquid (Fisher Scientific, Pittsburg, PA), and [14C] activity was measured using Beckman LS6000IC scintillation counter (Beckman Coulter, Brea, CA). To evaluate the differences in placental function between the individual embryos, the activity was normalized to gram of embryonic tissue.
The survival of newborn and pre-weaning mice from PAR-2/matriptase double heterozygous or PAR-2/prostasin double heterozygous breeding pairs was statistically evaluated by using chi-square analysis of the observed versus the expected distribution of mice wildtype, heterozygous, and deficient for PAR-2 (F2rl1+/+; F2rl1+/−; and F2rl1−/−, respectively) among living offspring carrying two (St14+/+ or Prss8+/+), one (St14+/− or Prss8+/−), or no functional alleles (St14−/− or Prss8−/−) of the gene encoding the corresponding protease.
To evaluate the effect of matriptase or prostasin deficiency on the embryonic survival of PAR-2-deficient mice, chi-square analysis was performed on the observed versus the expected distribution of protease-expressing (St14+/+ and St14+/−, or Prss8+/+ and Prss8+/−, respectively) and protease-deficient (St14−/− or Prss8−/−) animals among the PAR-2-deficient (F2rl1−/−) embryos extracted at different embryonic stages, as indicated in the text.
Morphometric parameters of the placental development were analyzed using tissues from at least five control and five F2rl1−/−;St14−/− double-deficient animals and the observed values were statistically evaluated using a two-sample Student's t-test, two-tailed.
|
10.1371/journal.pntd.0001349 | Factors Affecting Infestation by Triatoma infestans in a Rural Area of the Humid Chaco in Argentina: A Multi-Model Inference Approach | Transmission of Trypanosoma cruzi by Triatoma infestans remains a major public health problem in the Gran Chaco ecoregion, where understanding of the determinants of house infestation is limited. We conducted a cross-sectional study to model factors affecting bug presence and abundance at sites within house compounds in a well-defined rural area in the humid Argentine Chaco.
Triatoma infestans bugs were found in 45.9% of 327 inhabited house compounds but only in 7.4% of the 2,584 sites inspected systematically on these compounds, even though the last insecticide spraying campaign was conducted 12 years before. Infested sites were significantly aggregated at distances of 0.8–2.5 km. The most frequently infested ecotopes were domiciles, kitchens, storerooms, chicken coops and nests; corrals were rarely infested. Domiciles with mud walls and roofs of thatch or corrugated tarred cardboard were more often infested (32.2%) than domiciles with brick-and-cement walls and corrugated metal-sheet roofs (15.1%). A multi-model inference approach using Akaike's information criterion was applied to assess the relative importance of each variable by running all possible (17,406) models resulting from all combinations of variables. Availability of refuges for bugs, construction with tarred cardboard, and host abundance (humans, dogs, cats, and poultry) per site were positively associated with infestation and abundance, whereas reported insecticide use showed a negative association. Ethnic background (Creole or Toba) adjusted for other factors showed little or no association.
Promotion and effective implementation of housing improvement (including key peridomestic structures) combined with appropriate insecticide use and host management practices are needed to eliminate infestations. Fewer refuges are likely to result in fewer residual foci after insecticide spraying, and will facilitate community-based vector surveillance. A more integrated perspective that considers simultaneously social, economic and biological processes at local and regional scales is needed to attain effective, sustainable vector and disease control.
| Vector-borne transmission of Chagas disease remains a major public health problem in parts of Latin America. Triatoma infestans is the main vector in the countries located in the South American Cone, particularly in the Gran Chaco ecoregion where residual insecticide control has achieved only a moderate, irregular impact. To contribute to improved control strategies, we analyzed the factors associated with the presence and abundance of T. infestans in 327 inhabited houses in a well-defined rural area with no recent vector control interventions in the humid Argentine Chaco. Bugs were found mainly in domiciles, kitchens, storerooms, and chicken coops and nests, particularly where adequate refuge and animal hosts (humans, dogs, cats or poultry) were available. Domiciles constructed from mud were the most often infested, but brick-and-cement domiciles, even in good conditions, were also found infested. Availability of refuge and hosts for T. infestans are key targets for vector control. Ten-fold variations in domestic infestation observed across neighboring villages, and differences in the relevant factors for T. infestans presence with respect to other areas of the Gran Chaco region suggest that host management, building techniques and insecticide use need to be tailored to the local environment, socio-economic characteristics, and climatic conditions.
| The transmission of Trypanosoma cruzi, the etiologic agent of Chagas disease, by hematophagous triatomine bugs remains a major public health problem in many rural and some periurban communities in Latin America [1]. For decades vector control actions have relied mostly on residual insecticide spraying; implementation has been heterogeneous geographically and impacts on T. cruzi transmission variable. In addition to particular regional characteristics and the inherent difficulties encountered in controlling the vectors, T. cruzi transmission usually occurs in settings of marginalized human populations [2].
Triatoma infestans is not an exception. This species mainly infests poor rural dwellings and the surrounding human-made structures. Although the distribution range of T. infestans has been greatly reduced and parasite transmission interrupted in Brazil, Chile and Uruguay [3], residual insecticide spraying has had a moderate, irregular impact in the core of its range where vector elimination has encountered serious obstacles [1], [4], [5]. This area corresponds mainly to the Gran Chaco ecoregion, a 1.1 million km2 semiarid plain covering large parts of northern and central Argentina, southeast Bolivia, and central and western Paraguay [6]. The local human population consists mostly of Creoles and several indigenous groups sparsely distributed in rural areas, with scarce access to a very poor infrastructure, including weak health services and institutions under inconsistent provincial and federal policies [7]. These factors modify the effectiveness of vector control operations in an area where T. infestans thrives and where environmental conditions make pyrethroid insecticides less effective in some of the most heavily infested peridomestic structures surrounding human dwellings [8], [9]. Peridomestic structures usually play a fundamental role in maintaining abundant triatomine bug populations of various species close to domiciles [8], [10]–[16].
Understanding the factors associated with infestation may help to elucidate how the social dimensions of the problem become materialized; identify possible intervention targets, and predict parasite transmission risks. Broadly speaking, T. infestans depends on the availability of three basic biophysical characteristics to establish successful colonies: i) warm-blooded hosts, ii) suitable habitats and climatic conditions, and iii) not being exposed to effective insecticides. Humans, dogs, cats and chickens have repeatedly been identified as important domestic hosts [17]. Suitable bug habitat has often been associated with structural characteristics of houses and peridomestic structures: poor building conditions, wall cracks, mud-and-thatch houses and certain peridomestic structures such as goat or pig corrals and chicken coops [10], [12], [18]–[20]. The application of low-concentration insecticides by householders has been also found negatively associated with infestation [21], [22]. Seldom has the relationship between these various biophysical variables and house infestation been addressed simultaneously in a large number of houses with standardized bug collection methods, detailed information at biologically meaningful units of habitat, and robust analytical methods that reduce biases in variable selection. A multi-model inference approach serves that purpose [23] and has not been applied to Chagas disease vectors.
Despite the wide distribution area of T. infestans, variables associated with house infestation by T. infestans have been thoroughly assessed in few areas: in the Brazilian Cerrado [10], [24] and in the dry (western) section of the Gran Chaco [8], [21], [22]. However, the ecology of T. infestans and parasite transmission patterns may vary between ecoregions depending on climate, environment, human practices and ethnic background. Detailed descriptions of the key elements providing adequate habitats for triatomine bugs have rarely been given. As part of an ongoing multi-site research program on the eco-epidemiology and control of T. infestans in the Gran Chaco, the present study aims at modeling the variables associated with the presence and abundance of T. infestans in a well-defined rural area in the Argentine Chaco 12 years after the last residual insecticide spraying campaign.
The study was conducted in a section (450 km2) of the municipality of Pampa del Indio (25°55′S 56°58′W), Province of Chaco, Argentina (Figure 1A, inset), located in the humid (east) Chaco, close to the transition to the dry (west) Chaco. Based on data collected by the Chagas disease control program of Chaco, which indicated high infestation levels, we selected the municipality of Pampa del Indio as this project's location. Based on an exploratory survey throughout the rural area of Pampa del Indio, in which we inspected for triatomine infestation a systematic sample (11%) of the district's houses, we selected a well-defined section with slightly higher infestation than the rest and more than 300 adjacent houses isolated ≥1 km from the nearest villages outside of the selected area.
The study area included 353 houses and several public buildings of 13 neighboring rural villages (Figure 1). The climate is continental, warm, with rain mainly in summer. Annual mean temperature is 22.8°C (mean minimum and maximum, 16.9 and 29.3°C). Annual rainfall historically has been 954 mm, although in 2008 and 2009 a severe drought affected the region. The landscape is flat and comprises mainly a mosaic of patches of crops mixed with native dry forest that has undergone various degrees of degradation, and with occassional water bodies and marshes.
The two main ethnic groups are Creole and Toba. Tobas represent 24% of the 1,187 inhabitants of the study area and occupy 16% of the houses. Creoles are of European descent and usually had a high degree of mixing with indigenous people generations ago. Most Creoles migrated to the area during the last 50–100 years from nearby provinces or from Europe. Tobas –the only indigenous group in the area– were traditionally nomadic or semi-nomadic hunter-gatherers. Following local colonization in the 1920s they began to rely increasingly on agriculture, temporary informal jobs, and state-run welfare programs a few decades ago [25]. Local authorities reported having approximately 5,000 beneficiaries of welfare programs among the ∼14,000 inhabitants in the whole municipality, both Toba and Creoles. Rural residents live mostly on a subsistence economy, and may grow cotton, corn, pumpkins and water-melons or raise livestock (mainly goats, but also cows and occassionally sheep). The nearest hospital is in Pampa del Indio town (∼5,000 inhabitants), 10–45 km away from the study area by dirt roads. The last community-wide insecticide spraying campaign conducted by vector control personnel was carried out in 1996, except for a few houses treated by villagers or hospital staff in 2006, and no specific or systematic educational campaigns regarding Chagas disease or its transmission were performed in the area.
The study area encompassed 327 inhabited house compounds, including all of its domestic and peridomestic sites. There were 26 uninhabited or abandoned houses and 37 public buildings (including 11 schools, five primary health care centers, and several temples and community centers). A house compound encompassed the domicile and all sites within the peridomestic area (i.e., peridomicile) –usually a latrine, a storeroom, a kitchen, an oven, one or more corrals, and one or more sites for chickens and other poultry (trees, coops, nests) (Figure 2A,B). A site was any individual structure built and/or given a defined use by householders which might provide refuge for bugs. Ecotope (a categorical variable) was defined as a site characterized by some typical structure and use (e.g., domicile, storeroom, chicken coop, etc.). Nests were frequently found within a distinctive structure called ‘nidero’ (from the Spanish for nest, ‘nido’, Figure 2C), consisting of an elevated shelf made of wood or sometimes bricks where chickens and, occasionally, turkeys or ducks nested. Most domiciles were mud-and-thatch huts with corrugated metal or tarred-cardboard roof.
We conducted a cross-sectional survey of all houses in the study area accompanied by local personnel (health-care agent, vector control personnel) to explain to householders the goals of the research in a clear and simple language and to request access to their premises in September 4–15 and October 24–November 17, 2007, following protocol approved by IRB No. 00001678 (NIH registered). The full name of the head of each household and numbers of resident people aged 0–4 and 5–14 years old were recorded at each house. A sketch map of the spatial setting of the different structures in the house compound was drawn. Each structure was given an individual code according to its use to unequivocally identify it in follow-up surveys and associate it with the collected triatomines. The location of each site was georeferenced with a GPS receiver (Trimble GeoXM or Garmin Legend). After seeking householders' permission, an aluminum numbered plate was placed by the main entrance of the domicile and in public buildings. A labeled self-sealing plastic bag was provided to each household to keep any triatomine bug they may capture in domestic or peridomestic sites after insecticide spraying. Householders were asked to keep the bugs until our next visit and were instructed on how to collect the bugs without incurring a contamination risk.
Building materials used in roofs and walls, presence of wall plaster, condition of wall surface, plaster material, ecotope and physical characteristics of each site were recorded. The availability of suitable refuges for T. infestans was assessed visually by an experienced bug collector of the research group, and rated from 1 (no refuge at all) to 5 (abundant refuges). The head of each household (or some other adult person) was asked the number and type of domestic animal they owned, their resting sites, and domestic use of insecticides (type, frequency, mode and date of last application). Data on reported insecticide use, ecotope, construction materials, refuge availability, numbers of people and hosts and their resting places were collected in April, August and December 2008. Because animal host numbers varied seasonally, data collected in December 2008 were used to describe host numbers for the initial survey conducted in October 2007. This may introduce some inaccuracies, particularly in sites seldom used by one or very few individual hosts, such as some chickens or other poultry nesting in a domicile or storeroom, as this could vary from year to year. For all other variables with no seasonal variations, data collected on April 2008 were used; when values were missing, data collected in the following survey were used.
To evaluate whether the visual assessment of refuge availability of a site varied between different observers, after a preliminary standardization of criteria in two houses six similarly experienced bug collectors familiar with the study area independently scored refuge availability in 145 sites from 45 houses. These bug collectors were grouped in three different teams that scored the sites inspected for bugs during one day of survey.
Simultaneously with the household survey, all sites within each house compound were searched for triatomine bugs by timed manual collections conducted by two skilled bug collectors from the national or provincial vector control programs using 0.2% tetramethrin (Espacial, Argentina) as a flushing-out agent. Domiciles were inspected by one person for 20 min whereas each peridomestic site was searched by one person for 15 min. In practice, most searches lasted less because the site was completely inspected before the stipulated time. Therefore, the duration of searches in domiciles and in other ecotopes was approximately similar and therefore, the search effort was rounded to 0.25 person-hour per site. Public buildings were inspected in the same fashion. In several houses bugs were also collected after the stipulated search time (after-manual collections), during insecticide applications, or by householders a few days after timed manual collections or insecticide spraying. These additional bug collections were used solely as a qualitative measure of infestation. Householders' bug collections were only considered when accompanied with date and place of collection. Immediately after the entomological survey, all sites from every house were sprayed with suspension concentrate deltamethrin (K-Othrin, Bayer) at standard dose (25 mg/m2) by vector control personnel. The collected triatomine bugs were transported to the field laboratory in plastic bags labeled with bug collection site, identified taxonomically and counted according to species, stage or sex [26].
All results are reported solely for inhabited houses unless otherwise noted; uninhabited houses and public buildings were not found to be infested. Prevalence of infestation and colonization by T. infestans was calculated both for site or house compound levels. A site was considered infested when at least one live T. infestans nymph or adult was found by timed-manual collections. If at least a live nymph of T. infestans was found, the site was additionally considered colonized. Bug abundance was computed as the number of live T. infestans collected in a specific site by timed-manual collection per unit effort. The prevalence of site-specific infestation (or colonization) was calculated as the total number of infested (or colonized) sites divided by the total number of sites inspected for infestation. House-level infestation (or colonization) prevalence was calculated as the total number of infested (or colonized) house compounds divided by the total number of house compounds inspected for infestation.
The relationship between study factors and site-specific infestation or bug abundance per unit effort was assessed by means of logistic and negative binomial regression models, respectively. Negative binomial regression was chosen instead of Poisson regression because bug abundance per site was overdispersed. This was reflected in extremely larger values of Akaike's information criterion (AIC) for Poisson models than for negative binomial models (data not shown). Analysis was restricted to the most frequently infested ecotopes (i.e., kitchens or storerooms, domiciles, chicken coops, and ‘nideros’) because maximum likelihood procedures did not converge when other rarely infested ecotopes were included in negative binomial models. Regression analyses were performed for site level (i.e., each individual observation in the regressions represented a site).
Two separate sets of analyses were performed according to the typical characteristics of ecotopes to address the fact that the physical structure of a site affects bug detectability and therefore, comparability of results. One set of analyses included ecotopes typically made of proper walls and roofs (i.e., domiciles, kitchens and storerooms); the second set included small (peridomestic) ecotopes used by chickens (i.e., chicken coops and ‘nideros’). In both cases the response variables were site-specific infestation for logistic regressions and bug abundance per unit effort for negative binomial regressions. For domiciles, kitchens and storerooms, predictors included in the model were ethnic group, householders' reports of insecticide use, ecotope (with two categories, taking domiciles as the reference category), building materials (each material was taken as a separate variable since several of them could be present within the same site), refuge availability, numbers of people, dogs or cats, poultry (mostly chickens), and fledglings. For chicken coops and ‘nideros’ the same predictors were included, except ethnic group (due to the extremely few such sites among Tobas); householders' insecticide use (virtually restricted to domiciles, kitchens and storerooms), and numbers of people, dogs and cats (not reported to rest in chicken coops or ‘nideros’).
Pair-wise correlation coefficients between predictors were low (−0.2<r<0.2) for all pairs of variables except for: mud and refuge availability (r = 0.38), mud and thatch (r = 0.39), metal and brick (r = 0.32), domiciles and wood (r = −0.27); all these correlation coefficients were statistically significant (P<0.05). However, due to the possibility of linear dependencies involving more than two variables (i.e., multicollinearity) and its eventual detrimental effects on parameter estimation, the condition numbers of the predictors matrix were calculated [27]. The largest condition number was 8, indicating weak dependencies among variables that should not imply estimation problems [27].
A multi-model inference approach based on AIC was used to assess the relative importance (RI) of each variable according to Burnham and Anderson [23]. This procedure provides a quantitative ranking of the relative contribution of each variable to model fitting given the variables and models considered, while helping to reduce overfitting. All models considering every possible combination of the study variables were run: 14 variables gave 16,383 models for domiciles, kitchens and storerooms, and 10 variables gave 1,023 models for chicken coops and ‘nideros’. We considered all variables in equal terms since there was virtually no evidence on their relative importance when considered simultaneously. Therefore, the models considered were not alternative sets that evolved over time. The relative likelihood (i.e., Akaike weights) of each model was calculated as the quotient of the log-likelihood of the particular model divided by the total sum of the log-likelihoods of all considered models. The RI of each variable was calculated as the sum of the Akaike weights over all models in the set where the respective variable was present. The maximum value RI can take is 1, representing maximum importance relative to the set of variables considered, whereas RI = 0 represents no importance at all relative to the set of variables considered. Parameter estimates for each variable resulted from averaging the parameter value in each model where the variable was present weighted by the Akaike weight of the respective model. The overall quality of the fitted logistic regression models was assessed by means of the Hosmer and Lemeshow goodness-of-fit test using the averaged coefficients and grouping the data in 10 equal-sized groups; the area under the receiver operating characteristic (ROC) curve; sensitivity and specificity, using as cutoff values the observed prevalence of infestation for each data set (i.e., 17.7% for domiciles, kitchens and storerooms; 14.8% for chicken coops and ‘nideros’). Analyses and calculations were performed in R 2.7.0 [28]; the scripts with the commands for running the models and the respective calculations were prepared using a Visual Basic macro on Microsoft Excel (Text S1 and Script S1); model fitting was assessed in Stata 9.0 [29].
The spatial distribution of infested sites was assessed using the random labeling method with the pair-correlation function implemented in Programita [30]. The pair-correlation function g12(r) [31] evaluates if the number of points of pattern 2 within a ring of radius r and a given width, centered at each point of pattern 1, corresponds on average to a random process (i.e., a homogeneous Poisson process; g12(r) = 1), aggregation of 2 around 1 (g12(r)>1), or regularity of 2 with respect to 1 (g12(r)<1). The random labeling method assesses the spatial distribution of points belonging to a given pattern –relative to other points belonging to either that or other pattern– taking into account the spatial distribution of all points. Several site-specific characteristics affect the probability of being infested, regardless of the vicinity of other infested sites. Therefore, in the present case the null model for reassigning the membership to either pattern (and thus generating the expected distribution) took into account the probability of being infested according to the logistic regression model with all the variables weighted according to their RI. The spatial analysis only included sites pertaining to the four ecotopes most frequently infested (i.e., domiciles, kitchens and storerooms, chicken coops and ‘nideros’). The grid size for analysis was 100 m; ring width, 400 m; maximum radius, 5 km; 999 simulations were performed, and the upper and lower 25th simulations were used as a 95% confidence envelope. A goodness-of-fit test [30] was used to evaluate the overall fit of the observed pattern to the expected distribution.
To analyze the consistency among different observers of the visual estimation of refuge availability, the kappa index of agreement was calculated using Stata 9.0 [29]. This measure of agreement reaches the value of 1 when there is complete agreement and zero if observed agreement is equal to random agreement. Values>0.60 may be considered substantial to perfect agreement; values<0.40 represent poor agreement beyond chance, whereas values between 0.40 and 0.60 may be interpreted as moderate agreement beyond chance [32].
A total of 2,584 domestic or peridomestic sites was georeferenced, mapped and inspected for infestation. T. infestans was found by timed manual collections in 39.8% of the 327 inhabited house compounds and in only 7.4% of all sites inspected; bug colonies occurred in 30.8% of inhabited house compounds. Infestation with T. infestans was detected in 25.9% of domiciles and in 26.2% of peridomiciles. Triatoma sordida was found in 18.3% of houses, in 1.2% of domiciles (only adults) and in 17.6% of peridomiciles. A total of 2,062 T. infestans and 331 T. sordida were collected by timed manual collections. House compound infestation with T. infestans reached 45.9% when results from all collection methods were pooled together.
The most frequently infested ecotopes were domiciles (25.8%), kitchens or storerooms (15.4%), chicken coops (13.2%), and ‘nideros’ (18.2%), whereas <1.0% of corrals (including goat, pig and cow corrals), latrines, trees used by chickens and other ecotopes (ovens, chapels, abandoned cars, piles, etc.) were infested (Figure 3). The relative abundance of T. infestans tended to peak in ‘nideros’ followed by kitchens or storerooms, and was highly variable within and between ecotopes (Figure 4). Among frequently infested ecotopes, infestation in chicken coops was the most variable. Goat corrals and other types of corrals were rarely infested. In corrals, the mean bug abundance was widely variable but this was due to the very large number of bugs (80) collected at one of the four infested sites.
Most domiciles were built, partly or wholly, with mud (55.2%) and corrugated metal-sheet roofs (66.2%) (Table 1). Of domiciles with metal-sheet roofs, a substantial fraction had part of the roof made of thatch (14%) or corrugated tarred-cardboard sheets (7%) and ceilings made of wood (14%) or cane (2%), all of which provided suitable refuges for bugs. Mud, mixed with straw, was mainly applied onto a wooden frame for building walls (Figure S1A,G), and less frequently it was used for laying bricks instead of cement (Figure S1F). Deteriorated mud walls tended to have big cracks as the mud separated from the logs of the wooden frame of walls (Figure S1B); minor cracks were frequently observed in the mud plaster. In many heavily infested sites, no fecal streaks of triatomine bugs were found on wall surfaces but such streaks were abundant within cracks along the wooden frame of walls. Metal sheets were mostly used for roofs and, in a few cases, walls. Domiciles made solely of brick-and-cement walls and with metal-sheet roofs (comprising 36.4% of all domiciles) were also found infested but had substantially lower infestation (15.1%), colonization (6.2%) and refuge availability index (mean ± SD = 2.9±1.0) than domiciles with mud walls and roofs made of thatch or sheets of corrugated tarred-cardboard (32.2%, 21.6% and 3.8±0.9, respectively).
Kitchens and storerooms had similar construction features as domiciles, though brick walls were less frequent (Table 1). These ecotopes frequently exhibited a remarkable patchy distribution of building materials (Figure 2A,B), reflecting the changing access to building materials over time. Chicken coops were largely made of wood (80.2%); <17% of chicken coops was built with mud, brick, cardboard (either as tarred-cardboard sheets or as cardboard boxes) or thatch. ‘Nideros’ were similarly made as chicken coops though mud and bricks were more frequent and were usually used for building the small walls that separated one nest from the next (Figure 2C). Corrals were largely made of wood. Goat corrals were fenced with logs, boards or wire, and rarely had a roofed part for kids to shelter; when they did, the roof was typically made with corrugated metal sheets or wood boards and, more rarely, with tarred-cardboard sheets or thatch. Site-specific infestation by T. infestans and the index of refuge availability showed a close, positive association (Figure 5). The most frequently infested ecotopes also had higher refuge availability indices (Table 1, last column). Inter-observer agreement on the refuge availability index ranged between moderate and substantial and was significantly different from agreement by chance alone (kappa index = 0.56, z = 11.67, P<0.0001). Ratings between observers coincided in 68% of the rated sites, and differed by only one unit in 28% of the sites.
Dogs and cats were the most frequent domestic animal hosts resting in domiciles, kitchens or storerooms as reported by householders (Table S1). Dogs were reported to sleep indoors rarely, though they usually rested against the external walls of domiciles or in verandas. Chickens and other poultry were reported to rest at night mainly on trees and secondarily in chicken coops. Fledglings occurred in all peridomestic ecotopes. Corrals, latrines and ovens were rarely used by chickens, dogs or cats. Goats, sheep, pigs and cows occupied almost exclusively their respective corrals (having few refuges) where T. infestans was hardly ever found.
Toba households were larger and younger than Creole households, had more dogs, and substantially fewer chickens, goats, pigs, cows and equines than Creoles (Table 2). On average, Toba households had fewer sites per house compound (5.8) than Creole's (7.9), mainly due to fewer kitchens or storerooms, corrals, and peridomestic structures housing chickens (Table S2). The prevalence of house infestation with T. infestans (as determined by any collection method) was 58.8% for Tobas and 43.5% for Creoles (χ2 = 4.1, d.f. = 1, P = 0.04). Infestation was higher in Tobas' domiciles, kitchens or storerooms, and lower in Tobas' chicken coops and ‘nideros’ than in those owned by Creoles (Figure 3). The median abundance of T. infestans was similar between ethnic groups in domiciles, much higher in chicken coops and ‘nideros’ among Creoles, and slightly higher among Tobas in kitchens or storerooms (due to two sites with >12 bugs collected).
Reported use of some kind of insecticide during the previous year was higher among Creoles (64%) than Tobas (41%). Among families reporting insecticide use, 81.3% of Tobas and 77.5% of Creoles applied low-concentration pyrethroid sprays more than once a month, and 6.3% and 5.6% applied them between three and 12 times a year, respectively. Only 6.3% of Tobas and 8.1% of Creoles reported applying high-concentration pyrethroid or carbamate insecticides with manual compression sprayers (used for agriculture) once or twice a month, whereas 6.3% and 8.7% reported applications once or twice a year, respectively. Among households reporting insecticide use, infestation prevalence was lower for Creoles than Tobas in domiciles (19.7% vs. 28.3%), kitchens or storerooms (11.0% vs. 18.4%), and chicken coops (9.7% vs. 14.3%), respectively. Domestic infestation prevalence varied ten-fold across villages, whereas peridomestic infestation tended to be more similar between most villages (Figure 1B). Of the two villages with very high domestic infestations (>60%), one was inhabited only by Creole families whereas in the other village half of the households were Toba.
The multivariate analyses of factors associated with site-specific infestation and abundance of T. infestans in the most frequently infested ecotopes are shown in Table 3. In domiciles, kitchens and storerooms, refuge availability showed maximum RI (1.00) for infestation and bug abundance. Insecticide use (RI = 0.99), numbers of people and dogs or cats (RI = 0.95–0.96), and having a tarred-cardboard roof (RI = 0.86) all had higher RI regarding infestation than for bug abundance. Insecticide use was the only factor that had a negative association with infestation or bug abundance. Ethnic group evidenced a low RI regarding infestation and a moderately high RI in relation to bug abundance. In chicken coops and ‘nideros’, the numbers of chickens were important for both infestation and bug abundance whereas cardboard, thatch and mud were more important in relation to infestation than bug abundance, all with large positive effects. Refuge availability showed no clear importance in chicken coops and ‘nideros’.
The averaged logistic model for infestation in domiciles, kitchens and storerooms had a marginally significant fit (Hosmer-Lemeshow χ2 = 14.9, 8 d.f., P = 0.06). The area under the ROC curve was 0.742; sensitivity was poor (49%) and specificity higher (82%), with 76% of observations correctly classified. The averaged model for chicken coops and ‘nideros’ had a good fit (Hosmer-Lemeshow χ2 = 4.5, 8 d.f., P = 0.81); the area under the ROC curve was 0.759; sensitivity of 65% and specificity of 71%, with 70% of observations correctly classified.
The averaged negative binomial models in both groups of ecotopes exhibited significant correlations between observed and predicted bug abundances (for domiciles, kitchens and storerooms, r = 0.23; for chicken coops and ‘nideros’, r = 0.61). However, these models considerably underestimated the frequency of zero bugs per site and overestimated higher bug catches (data not shown).
The spatial distribution of sites infested with T. infestans (as determined by any collection method) was assessed only among the most frequently infested ecotopes. Infested sites were significantly aggregated at distances between 0.8–2.5 km (Figure 6). This aggregation remained significant even after taking into account the probability of finding each infested site according to a logistic regression model that included insecticide use, ecotope, refuge availability, building materials, and number and type of animal hosts.
Through a multi-model inference approach we showed that availability of appropriate refuges for T. infestans bugs, use of cardboard as a building material and domestic host abundance (humans, dogs, cats, and chickens) were strongly and positively associated with bug infestation and abundance, whereas the association with reported insecticide use by householders was negative in a high-risk rural area in the Argentine Chaco. It is noteworthy that ethnic background, when adjusted for other factors, had little or no association with infestation.
Refuge availability was a key variable for describing the site-specific presence and abundance of T. infestans in domiciles, kitchens and storerooms both in univariate and multivariate analysis, having higher relative importance than building materials or host abundance. Refuge availability was also moderately important for bug abundance in chicken coops and ‘nideros’. Refuge availability was demonstrated experimentally to be essential for the development of T. infestans colonies [20], but most field studies addressed its effects on domestic infestation only indirectly by assessing the types of wall and roof materials, and only a few considered the degree of wall cracking explicitly [21], [22], [33]. However, wall cracks represent only a fraction of the possible refuges in sites with proper walls, such as domiciles, kitchens and storerooms. The refuge index probably captured part of the effects of mud as a building material as both factors were moderately correlated (r = 0.38). Use of mud has been consistently found to be an important factor associated with infestation throughout the Gran Chaco and in other regions [34] because it frequently cracks (Figure S1A) and therefore provides adequate refuges for triatomine bugs, unless well maintained as in Figure S1G [26]. However, in our multiple regression analysis the inclusion of refuge availability, mud and thatch were moderately associated only with infestation in chicken coops and ‘nideros’, and not at all with infestation in domiciles, kitchens or storerooms. Even though estimates of refuge availability may appear to be a subjective and imprecise measure, concordance among several experienced bug collectors familiar with the study area was moderate to substantial. Therefore, refuge availability may be used as an indicator of habitat suitability for T. infestans.
A striking demonstration of the effects of refuge availability on infestation with T. infestans is provided by local goat corrals. Because of the type of construction, most goat corrals in Pampa del Indio had very few or no refuges for bugs and were rarely infested (Figure S1I,J). In contrast, goat corrals in the dry Chaco were typically constructed with fences made of piled thorny shrubs and a thatched or earth-made roofed area for shelter, and were very often heavily infested before and soon after residual insecticide spraying [8], [35].
Cardboard material (either as boxes or as tarred-cardboard sheets) used in ‘nideros’, chicken coops and corrals was associated positively and directly with infestation; it provides excellent refuges for T. infestans and is also easy to inspect for bugs. In contrast, in domiciles, kitchens and storerooms (where cardboard as a building material is only present in roofs, Figure 2A) no bugs were collected from cardboard material itself. This is hardly surprising because roofs are more difficult to search for bugs, especially high roofs. Corrugated cardboard-made roofs may also be linked to infestation indirectly since they characterize households with fewer resources. Rural villagers reported that cardboard sheets were much cheaper than metal sheets and easier to get than appropriate thatch for roofing (Spartina densiflora, locally known as ‘espartillo’). Cardboard use exemplifies the variable and complex relationships that an apparently simple building material may have with several other relevant factors.
Insecticide use by householders had high RI for infestation in domiciles, kitchens and storerooms. The frequent domestic use of insecticides in Pampa del Indio may explain, at least in part, why the observed prevalence of house infestation with T. infestans was much lower than expected, considering that the most recent residual spraying with insecticides conducted by the vector control program occurred 12 years earlier, and other areas in the dry Chaco experience fast reinfestation after insecticide spraying campaigns [4], [8], [9], [36]. Insecticide use may also reflect to some extent the economic status of householders and/or their concern for and response to the presence of bugs and other domestic pests [22]. Additionally, our study area exhibited a relatively high proportion of brick-and-cement houses with corrugated metal roofs (36.4%). Although such construction materials do not fully impede house invasion and colonization by triatomine bugs, they probably present much more of an obstacle than mud-and-thatch huts. Local residents frequently recalled that T. infestans had been much more abundant a few decades ago and attributed the decline to earlier vector control campaigns (traditionally lacking an educational component and not followed by a vector surveillance and response system), domestic insecticide use, declining use of thatch for roofs and mud for walls, and deforestation. Villagers usually considered the surrounding forest as the main source of T. infestans. Although sylvatic foci of T. infestans have not been detected in the study area so far (Alvarado-Otegui et al., unpublished data), they have been reported elsewhere in the Gran Chaco [37]–[39]. Taken together, householders' reports and the mixture of building materials recorded (Figure 2B) indicate a slow, gradual trend toward improvement of rural housing in this area of Chaco.
The abundance of certain host species at site level was an important factor in describing variations in infestation and bug abundance. Dogs, cats and people were associated positively with variations in infestation in domiciles, kitchens and storerooms as in the dry Chaco [21], [22] and elsewhere for other triatomine species [40]–[42]. The number of domestic chickens (and some other poultry) was highly relevant for infestation in chicken coops and ‘nideros’, as it was elsewhere in the Argentine dry Chaco [17], [22], [43]. However, such expected effects were not detected in domiciles, kitchens and storerooms, probably due to the relative imprecision of householders' reports on host presence at a given site combined with the lower frequency of chickens nesting in domiciles in comparison to previous study areas.
Differences in health indicators between ethnic groups abound in the literature [44], but we were unable to find any comparisons of levels of house infestation with triatomine bugs between Creole and indigenous populations within the same area. Our results did show that Toba households, which had several demographic differences and used insecticides less frequently than Creole households, had higher infestation in domiciles, kitchens and storerooms. These differences are consistent with the much higher seroprevalence of T. cruzi among Tobas (69.9%) compared with Creoles (40.4%) residing close to our study area [45], whereas 54% of Tobas were seropositive in an unspecified section of Pampa del Indio [46]. However, in our multiple regression analysis, when ethnic group was considered simultaneously with other relevant variables, its relationship with infestation was not important. Ethnic group could be relevant in understanding the distribution of T. infestans insofar as it is a surrogate of factors directly linked to infestation (e.g., insecticide use). An excessive focus on ethnic background, construed mainly as cultural differences, may obscure other substantial variations within and between Tobas and Creoles. These differences may be related to resources, insecticide use, types and numbers of peridomestic structures and demographic features. Indeed, we documented large differences in house infestation between and within study villages, even among those only having Creole households. Whether differences in infestation between and within Tobas and Creoles reflect average differences in social and economic position between and within groups [47] rather than cultural or behavioral differences that modify infestation merits further inquiry.
A major finding of our study is the marked heterogeneity in the distribution of infestation, housing construction patterns (Figure S1) and several risk factors within an apparently homogeneous rural area. Domestic infestation varied ten-fold among villages, and there was a significant spatial aggregation of infestation. While some houses were heavily infested, 54% of houses were apparently free of T. infestans despite the absence of recent vector control actions. Moreover, the presence of the vector species was widespread throughout the study area. Both of these findings imply that T. infestans already had sufficient time and opportunities to reach the most suitable sites for colonization [4]. The significant spatial aggregation of infested sites within a range of 0.8–2.5 km suggests the contribution of factors related to infestation that were not explicitly included in our models. Such putative factors that may contribute or underlie the aggregated occurrence of T. infestans are the economic status of householders (e.g., by affecting housing quality or maintenance, and access to insecticides), extended family clusters (which usually live in nearby houses and probably share similar housing characteristics, domestic and productive practices), cultural groupings (such as Tobas, who tend to live in clusters) and productivity (as it strongly affects economic status), all of which may be spatially clustered and eventually interrelated. This underscores the need for identification of relevant socioeconomic variables for an improved understanding of system structure and dynamics.
We found substantial differences regarding host availability and local characteristics of adequate habitats for triatomine bugs in comparison to other study areas in the Argentine dry Chaco [9], [22], [35]. The local use and relevance for infestation of tarred-cardboard and metal sheets as building materials contrast with its rarity elsewhere. ‘Nideros’, highly infested and widespread in Pampa del Indio (Figure 2C), are virtually absent in the dry Chaco. Goat corrals were rarely infested in Pampa del Indio and often heavily infested in the dry Chaco for reasons explained above. Domiciles in Pampa del Indio are roofed mostly with corrugated metal or tarred-cardboard sheets and occasionally thatch (Figure S1C,G), verandas are rare and small (Figure 2B), and most bugs are collected from walls and stored goods. In contrast, in the dry Chaco most domiciles have thatch-and-earth roofs (where T. infestans is very common) and large verandas where people sleep during the hot season [21], [22], [48], [49]. These contrasting patterns likely derive from differences in climate (i.e., earth-and-thatch roofs deteriorate fast with more rainfall in the humid Chaco); domestic animal management practices (influenced by traditions, local needs and possibilities and other factors); local political practices (corrugated metal and cardboard sheets were frequently provided pro bonus by political candidates in Pampa del Indio), and more intense activities of charity organizations. Taken together with the rich environmental, cultural and climatic diversity within the Gran Chaco, our findings emphasize the need of considering local specificities and processes (including householders' practices) as these may vary widely within the same region and modify the likelihood for house infestation in unforeseen ways.
Cross-sectional, observational field studies such as ours have inherent limitations and strengths. Timed-manual searches for bugs with a dislodging spray is the standard method used for detecting infestations but it has limited sensitivity and precision [50], [51]. Its sensitivity for detecting domestic T. infestans infestations reached 70–77% in the dry Chaco [52], but likely declines at very low bug densities. At the expense of increased cost, repeated searches on the same site may be used to estimate a probability of bug detection that later can be incorporated in statistical modeling [53]. Information on the presence and abundance of animal hosts and domestic use of insecticides (both varying seasonally) based on householders' reports is subject to recall bias and therefore is rather imprecise. These issues probably underlie the limited fit of logistic and negative binomial models. Multiple regression analyses only show that the variability registered in the explanatory variables is not related strongly enough to the variability registered in the outcome variables (infestation or bug abundance). A statistically insignificant result may reflect that such explanatory and outcome variables are truly uncorrelated, or that there is not enough variability registered in them so as to detect any correlation. On the flip side, major strengths of this research effort are the detailed information at site level in a sizable number of house compounds from a well-defined rural area, analyzed with a multi-model inference approach that reduces overfitting.
Housing improvement is usually mentioned as the ultimate solution for domestic infestation [1], but in practice residual insecticide spraying remains virtually the only tool used for suppressing house infestation with triatomine bugs. Our results have implications in both directions: i) Housing improvement (including key peridomestic structures) should be promoted more widely and conducted more effectively to reduce refuge availability and bug infestations, rather than focusing solely or mainly on replacing some building materials (e.g., thatch by corrugated metal sheets). An urban-like house with brick-and-cement walls and corrugated-metal roof does not guarantee the absence of bugs, as shown in Figure S1B,H, since adequate refuges may still exist in walls, beds, furniture and stored goods. Conversely, mud houses may be kept free of T. infestans provided they are adequately maintained (Figure S1G), the presence of animal hosts in domiciles is minimized, and insecticides are used when needed [22], [26]. Inexpensive modifications of key peridomestic ecotopes, such as reducing refuge availability in ‘nideros’ or renovating them more often, may strongly reduce peridomestic populations of T. infestans; ii) Better housing with fewer refuges for bugs will likely reduce the chance of residual foci after insecticide spraying [8], [33] and facilitate community-based vector surveillance, and iii) Local and regional heterogeneities in the distribution of infestation and relevant factors (including human practices and socio-economic characteristics) need to be considered in order to prioritize and allocate insecticide spraying operations and vector surveillance more efficiently. The preference of T. infestans for certain hosts and types of refuge is strictly biological, but their occurrence in a house is also a social phenomenon that needs to be addressed more broadly. A more integral perspective that simultaneously considers social, economic and biological processes at local and regional scales, and the needs, possibilities and expectations of local populations is required for attaining effective and sustainable vector and disease control [47], [54]–[56].
|
10.1371/journal.ppat.1002661 | AMP-Activated Kinase Restricts Rift Valley Fever Virus Infection by Inhibiting Fatty Acid Synthesis | The cell intrinsic innate immune responses provide a first line of defense against viral infection, and often function by targeting cellular pathways usurped by the virus during infection. In particular, many viruses manipulate cellular lipids to form complex structures required for viral replication, many of which are dependent on de novo fatty acid synthesis. We found that the energy regulator AMPK, which potently inhibits fatty acid synthesis, restricts infection of the Bunyavirus, Rift Valley Fever Virus (RVFV), an important re-emerging arthropod-borne human pathogen for which there are no effective vaccines or therapeutics. We show restriction of RVFV both by AMPK and its upstream activator LKB1, indicating an antiviral role for this signaling pathway. Furthermore, we found that AMPK is activated during RVFV infection, leading to the phosphorylation and inhibition of acetyl-CoA carboxylase, the first rate-limiting enzyme in fatty acid synthesis. Activating AMPK pharmacologically both restricted infection and reduced lipid levels. This restriction could be bypassed by treatment with the fatty acid palmitate, demonstrating that AMPK restricts RVFV infection through its inhibition of fatty acid biosynthesis. Lastly, we found that this pathway plays a broad role in antiviral defense since additional viruses from disparate families were also restricted by AMPK and LKB1. Therefore, AMPK is an important component of the cell intrinsic immune response that restricts infection through a novel mechanism involving the inhibition of fatty acid metabolism.
| RNA viruses represent an important worldwide source of infection and disease in both humans and animals. While individual viruses have unique characteristics, some stages of infection are conserved and must be completed in order to successfully infect and grow. Viruses must undergo genome replication, protein synthesis, and assembly of new virus particles. In particular, numerous RNA viruses manipulate cellular membranes to create new complex structures required for viral replication in a process that is often dependent on fatty acid biosynthesis. This is a process that is tightly regulated by the energy sensor AMPK. We have found that energy-mediated activation of AMPK restricts infection of the Bunyavirus Rift Valley fever virus by decreasing levels of fatty acid synthesis. Furthermore, several additional RNA viruses from disparate families that share this dependence of membrane modification and fatty acid synthesis are also restricted by AMPK. Thus AMPK likely represents a novel component of the cell intrinsic immune response to RNA viruses, and may be a good target for the development of antiviral therapeutics against a range of medically important viruses.
| Emerging and re-emerging arthropod-borne viral pathogens have lead to significant world-wide morbidity and mortality in humans and domestic animals, and are of medical and agricultural concern. Bunyaviruses are an important group of insect-borne RNA viruses that include disease causing members such as Sin Nombre, Hantavirus, Crimean-Congo hemorrhagic fever virus, and Rift Valley Fever Virus (RVFV). RVFV is a mosquito borne Category A agent initially endemic to sub-Saharan Africa. However, outbreaks of RVFV have recently occurred in Egypt and the Arabian Peninsula, indicating the potential of this virus to spread to new geographical areas [1]. RVFV has particular importance as an agricultural pathogen, where infection of livestock can lead to significant morbidity and mortality among young animals, and cause catastrophic abortion rates [1]. Most humans infected with RVFV develop self-limited febrile illness, although approximately 1–3% die from the disease due to hemorrhagic symptoms [2]–[5]. No effective vaccines or antiviral therapies have yet been developed against RVFV.
All viruses undergo sequential steps to complete their replication cycles. Bunyaviruses and other RNA viruses compartmentalize their RNA replication machinery on cellular membranes. An essential feature of these infections is the ability of viruses to rearrange and proliferate internal cellular membranes into distinct structures compartmentalizing the viral replication complex and supporting viral genome replication [6]. Depending on the virus, these membrane modifications can be derived from distinct cellular sources, including ER, Golgi, endosomal, and mitochondrial membranes, and may have complex biogenesis pathways derived from multiple intracellular origins [7]–[14]. Bunyamwera virus, a member of the Bunyavirus family related to RVFV, induces the formation of a new Golgi membrane-derived tubular structure with a globular head that harbors the viral replication complex [14], [15]. Disrupting the formation of this structure is associated with decreased levels of virus replication [15]. While different families of viruses use membranes derived from different cellular sources, and create membranous structures with distinct morphologies, there are some similarities in these structures, suggesting that commonalities exist in the mechanisms by which disparate viruses depend upon lipid metabolism or trafficking [16]. One clear point of overlap includes a requirement for cellular lipid biogenesis pathways and the generation of newly synthesized lipids [6]. Furthermore, enveloped viruses, which include Bunyaviruses, require incorporation of cellular membranes into their lipid envelopes during virus assembly, in a process that may also involve lipid modifications [17].
AMP-activated Kinase (AMPK) is a heterotrimeric complex that is the core energy sensor of the cells [18]. Thus AMPK activity is important for survival during periods of stress, and also has implications in type II diabetes, obesity, metabolic syndrome, longevity, and cancer [19]–[25]. The AMPK complex consists of a catalytic alpha subunit, and regulatory beta and gamma subunits [26]. Activation is triggered through binding of AMP or ADP to the Bateman domains of the gamma subunit, leading to increased phosphorylation of the threonine 172 on the alpha subunit by inducing allosteric activation and inhibiting dephosphorylation [27]–[30]. The canonical upstream activator that catalyzes this phosphorylation event is the constitutively active tumor suppressor LKB1, but additional activators such as CaMKKβ have been identified [31]–[35]. Under conditions of energetic stress, AMPK signals the cell to stop anabolic pathways and activate a catabolic state by inducing oxidative pathways that generate energy while inhibiting synthesis and growth pathways, thereby returning the cell to a state of energy homeostasis [26]. To achieve this regulation, AMPK targets a number of downstream pathways including those involved in lipid metabolism.
As a potent regulator of lipid metabolism, AMPK activity inhibits both sterol and fatty acid synthesis, while promoting fatty acid degradation [18]. AMPK directly phosphorylates acetyl-CoA carboxylase (ACC) and HMG-CoA Reductase (HMGCR), thereby inactivating these rate limiting enzymes in the metabolism of fatty acids and sterols respectively [36], [37]. In particular, ACC catalyzes the irreversible conversion of acetyl-CoA to malonyl-CoA, a key metabolite that plays multiple roles in fatty acid metabolism. First, malonyl-CoA is the substrate for fatty acid biogenesis, which drives de novo production of the fatty acid palmitate [38]. Second, malonyl-CoA is a co-substrate for chain lengthening of endogenously synthesized and dietary-derived essential fatty acids into higher polyunsaturated fatty acids [39]. Third, malonyl-CoA binding inhibits carnitine palmitoyltransferase I (CPT-1), an essential factor in the transport of fatty acids to the mitochondria for beta oxidation [38]. Thus malonyl-CoA production by ACC promotes fatty acid synthesis, while inhibiting fatty acid oxidation. Mammalian systems encode two non-redundant ACC isoforms, ACC1 and ACC2, which are both inactivated by AMPK-mediated phosphorylation. Studies suggest that malonyl-CoA produced by ACC2 is involved in fatty acid oxidation, while ACC1 contributes to fatty acid biogenesis [18]. Therefore, activation of AMPK through stress or low energy conditions induces fatty acid oxidation through ACC2, while inhibiting fatty acid synthesis through ACC1, with a net result of lipid breakdown.
We found that AMPK is potently antiviral against RVFV, and this restriction is dependent on the upstream activator LKB1. Furthermore, pharmacological activation of AMPK inhibited viral infection. AMPK was activated by RVFV infection, and in particular we observed striking changes in ACC activity dependent on AMPK, leading us to discover that AMPK is antiviral through its role in fatty acid metabolism. Cells lacking AMPK had increased global lipid levels, while pharmacological activation of AMPK led to decreased cellular lipids, consistent with AMPK control of lipid availability as a restriction point for viral replication. Importantly, we could bypass the antiviral effects of AMPK by feeding cells palmitate, the first fatty acid produced downstream of ACC. Since palmitate treatment restored RVFV infection, we demonstrate that AMPK specifically restricts infection through its role in inhibiting fatty acid biosynthesis. Since many viruses are dependent upon fatty acid biosynthesis for their replication, we tested whether AMPK restricted additional RNA viruses. We found that indeed, AMPK has antiviral activity against multiple arboviruses from disparate families including: the Flavivirus Kunjin virus, the Togavirus Sindbis virus, and the Rhabdovirus Vesicular stomatitis virus. Taken together, our data suggest that AMPK activation is broadly anti-viral, and may provide a novel antiviral therapeutic target.
We previously reported that AMPK was required for efficient vaccinia infection through its role in macropinocytosis [40]. This led us to investigate the role of AMPK in other virus infections; we were particularly interested in RVFV as it is a virus that is medically important, but little is known about the mechanisms by which it establishes a productive infection. For our studies we used the lab adapted strain MP12 that has 11 amino acid differences from the wild type strain, since the wild type strain must be used in high containment facilities [41]. In order to test the role of AMPK in RVFV infection, we took advantage of mouse embryonic fibroblasts (MEF) that are genetically altered and null for both of the catalytic α subunits, AMPKα1 and AMPKα2 (AMPKα1/AMPKα2−/−) [42]–[44]. We challenged either the AMPKα1/AMPKα2−/− MEFs or their sibling control wild type MEFs with RVFV and measured infection by plaque assay (Figure 1A). We found an increase in titer from 5×105 pfu/ml to 3×106 pfu/ml, indicating a 6-fold increase in the number of plaques formed in AMPKα1/AMPKα2−/− MEFs compared to wild type (Figure 1B), concomitant with a 4-fold increase in average plaque area in AMPKα1/AMPKα2−/− MEFs (Figure 1C). Moreover, RVFV infection was also increased in AMPKα1/AMPKα2−/− MEFs as measured by an immunofluorescence assay that detects production of the RVFV N protein produced during viral replication (Figure 1D, quantified in Figure 1E), indicating that RVFV is able to infect and spread more efficiently in the absence of AMPK. Consistent with a role for AMPK both in early events during viral replication and in spread as measured by plaque assay Figure 1A), we observed an increase in viral infection at early time points before virus spread, as well as increased spread in cells lacking AMPK by monitoring the production of RVFV N protein over time by microscopy (Figure S1A–B).
This increased spread, indicated by the increase in plaque size (Figure 1C), as well as the immunofluorescence assay (Figure S1A–B), could result from increased production of infectious virus or increased infectivity of the virions produced in cells lacking AMPK. We measured the amount of infectious virus produced in wild type and AMPKα1/AMPKα2−/− MEFs over time in a one-step growth curve. Medium from infected cells was collected at various times after infection, and virus was tittered on wild type BHK cells. Little virus (less than 1×104 pfu/ml) was detected at 2–4 hpi, indicating that input virus was not detected in this assay (Figure 1F). Virus release began at 8 hpi, where we already observed an 8-fold increase in titer in the AMPK deficient MEFs (1.6×105 pfu/ml versus 1.3×106) (Figure 1F). This increase in titer was also observed at 12 hpi. Therefore, the increase in RVFV spread is likely due to increased virus production in AMPKα1/AMPKα2−/− MEFs.
AMPK is activated through phosphorylation of a threonine residue on the catalytic alpha subunit [45]. Since AMPK deficiency increased RVFV infection, we hypothesized that AMPK activation would inhibit infection. Therefore, we tested whether RVFV was sensitive to pharmacological treatments that activate AMPK. First, we tested drugs that activate AMPK by reducing the levels of cellular energy using an independent cell line, the human osteosarcoma cell line (U2OS). We tested the glucose analog 2-deoxyglucose (2DG), and the ATP synthase inhibitor oligomycin, and found that both treatments significantly decreased infection by RVFV compared to vehicle controls (Figure 2A). In contrast, the AMPK inhibitor Compound C significantly, albeit modestly, increased RVFV infection (Figure S2A). Since 2DG and oligomycin activate AMPK indirectly by reducing cellular energy levels, and thus likely have other effects that may also contribute to viral infection, we tested whether these treatments affected vaccinia virus infection, which is not restricted by AMPK, but rather requires AMPK, independent of the energy sensing pathway for efficient viral infection [40]. Vaccinia virus infection was not affected by these treatments (Figure 2B), indicating that the compound-treated cells remain healthy enough to support viral infection, and the reduced infection levels were specific to RVFV. Moreover, we found that none of these drug treatments reduced cell number by greater that 20%, and therefore were not cytotoxic (Figure S2B).
Next, we took advantage of a recently developed thienopyridone compound A769662 that activates AMPK directly, independently of the energy status of the cell [46], [47]. This drug mimics both allosteric activation of AMPK and inhibition of dephosphorylation without affecting binding of AMP to the gamma subunit [48]. We found that RVFV infection of U2OS cells was significantly reduced in the presence of this compound (Figure 2C), and that both 2DG and A769662 inhibit RVFV in a dose-dependent manner (Figure S3A–B), indicating that AMPK activation restricts RVFV infection independently of the pleiotropic effects of reduced cellular energy levels. Moreover, we also found that the AMPK activating drugs 2DG and A769662 significantly inhibit RVFV infection in MEFs (Figure 2D). To determine if the effects of these drugs was specific for AMPK we treated AMPKα1/AMPKα2−/− MEFs with the direct AMPK activator A769662. Treatment with this drug inhibited RVFV less than 2 fold in AMPKα1/AMPKα2−/− MEFs and was not significant, whereas infection was inhibited greater than 5-fold in the wild type cells (Figure S4A) with no toxicity in either cell type (Figure S4B), indicating that the major action of this drug was through AMPK as previously published [46], [47]. Taken together, these studies suggest that AMPK activation has antiviral activity against RVFV in multiple cell types.
Since pharmacological activation of AMPK restricted RVFV infection, we were interested in investigating which pathway upstream of AMPK was responsible for this restriction. The classic activator of AMPK is the tumor suppressor LKB1, which phosphorylates AMPK in response to a variety of stimuli that cause a reduction in cellular energy levels, such as glucose starvation or hypoxia [26]. In order to determine if LKB1 signaling was important for AMPK-mediated RVFV restriction, we tested whether LKB1 also restricted RVFV. We challenged MEFs that are null for LKB1 and complemented with either vector alone (LKB1−/−; Vec), or an LKB1 cDNA (LKB1−/−; LKB1) [40] and found increased RVFV infection in MEFs lacking LKB1 by plaque assay (Figure 3A). Quantification revealed a 2-fold increase in the number of plaques (increase in average virus titer from 7.8×105 to 1.5×106 pfu/ml in LKB1 null MEFs) (Figure 3B) and a 5-fold increase in plaque area in LKB1−/−; Vec MEFs compared to MEFs complemented with LKB1 (Figure 3C). Moreover, we observed increased infection in the LKB1−/−; Vec MEFs compared to those complemented with LKB1 by immunofluorescence (Figure 3D, quantified in 3E). Finally, we measured RVFV infection over time in cells lacking LKB1 and found increased infection in the absence of LKB1 at early and late times after infection, indicating increased initial infection as well as spread (Figure 3F). Since AMPK activation downstream of LKB1 is dependent on a decrease in cellular energy, we measured cellular ATP levels during RVFV infection using a luciferase assay. While 2DG significantly reduced cellular ATP levels, neither A769662 nor RVFV had any impact on ATP levels as measured by this assay (Figure S5). While infection with RVFV did not induce global changes in cellular ATP, this does not rule out localized changes in cellular energy that could influence AMPK.
In addition to LKB1 other upstream activators of AMPK have been identified. Notably, calcium-calmodulin kinase kinase (CaMKK) has been shown to activate AMPK in response to an increase in intercellular calcium [33], [34], [49]. Since LKB1 did not restrict RVFV as strongly as AMPK did (Figure 3), we investigated if other upstream activators, such as CaMKK could also contribute to RVFV restriction. To this end, we treated U2OS cells with the CaMKK inhibitor STO609 prior to infection, and found no increase in RVFV infection in response to this drug, although at very high concentrations there was a decrease in infection (Figure S3C). This decrease was likely due to additional kinases that are inhibited at these concentrations [50]. This finding is consistent with previous reports that changes in intercellular calcium levels are not induced by RVFV infection [51]. We next investigated if LKB1 and CaMKK function redundantly to restrict RVFV infection. We tested whether simultaneously inhibiting both LKB1 and CaMKK would lead to a greater increase in RVFV infection than LKB1 deficiency alone. To this end, prior to infection, we treated LKB1 null MEFs or those complemented with LKB1 with STO609 and monitored RVFV infection. Consistent with our previous findings, we observed a 3-fold increase in the percentage of infected cells in LKB1 null cells compared to those complemented with LKB1; however pretreatment with STO609 had no effect on infection level in either cell type (Figure 3G). In contrast, and as expected, we found that pretreatment with the AMPK activating compound A769662 significantly inhibited RVFV in both LKB1 null and complemented MEFs (Figure 3G). Taken together, these data suggest that LKB1 is the major upstream activator responsible for AMPK-mediated restriction of RVFV.
To dissect the mechanism by which AMPK restricts RVFV infection, we first determined which early step in the viral replication cycle is restricted by AMPK. We observed decreased protein production, as measured by immunofluorescence (Figure 1D–E and Figure S1) in addition to decreased production of infectious progeny virus (Figure 1F) in the presence of AMPK. This suggests that AMPK may inhibit a step in the viral replication cycle at, or prior to, protein production. To determine if viral RNA replication was affected by AMPK, we monitored both viral genomic RNA replication and viral mRNA production in the presence or absence of AMPK. We found an increase in both viral mRNA (N) and genomic RNA (S segment) in AMPK deficient MEFs both early in infection and upon virus spread (Figure 4A–C). At 4 hpi, a time point prior to RVFV release, we observed a 3-fold increase in viral mRNA production in AMPK deficient MEFs compared to wild type, which continued to increase over time (Figure 4A–B). Likewise, genomic RNA production was increased prior to virus release and spread (Figure 4A and C). These data suggest that the increased N protein production observed by immunofluorescence at early time points (Figure S1A) may be due to increased N mRNA production.
Next, we investigated whether entry, a step upstream of RNA replication, was inhibited by AMPK. First, we tested whether RVFV binding was more efficient in the absence of AMPK. To this end, MEFs were pre-bound with RVFV for an hour at 4°C, unbound virus was removed and RVFV binding was measured by quantitative RT-PCR to detect genomic RVFV S segment within virions. We observed no difference in virus binding in wild type or AMPK deficient cells (Figure 4D). Moreover, the majority of virus was removed by trypsin treatment in both wild type and AMPK deficient MEFs, indicating these virions had bound to the cell surface, but not entered (Figure 4D).
Since AMPK did not impede virus binding, we next performed a time of addition assay to test whether AMPK-activating drugs restricted entry. Since Bunyaviruses such as RVFV enter cells through a pH-dependent route of endocytosis [51]–[53], we used the lysosomotropic agent ammonium chloride, which raises the pH of lysosomal compartments, to define the timing of virus entry. Ammonium chloride inhibited infection strongly (to 20% of the 4 hpi addition) when added 1 hour prior to infection or with infection (t = 0); however, by 1 hpi, more than 70% of infection had returned, indicating that the majority of RVFV had entered by this time point (Figure 4E). Thus we compared each treatment to the post entry level of RVFV infection (ammonium chloride added at 4 hpi). AMPK activating drugs 2DG, and A769662 significantly inhibited infection when added at post entry stages (Figure 4E); however, since one of the AMPK activating drugs, A769662, had a significantly greater impact on RVFV when added prior to or with infection, we cannot rule out that AMPK also inhibits RVFV entry. Taken together these data suggest that AMPK restricts RVFV during initial stages of replication post entry, likely at the step of RNA replication. This reduction in viral RNA and protein production likely leads to a reduction in release of infectious virus and spread observed at later stages of infection.
The classical cell-mediated response to viral infection is the type I interferon system [54], [55]. Therefore, we investigated whether AMPK impacts the expression of interferon beta (IFNβ) or its downstream effector 2′-5′-oligoadenylate synthetase 1 (OAS1) by qRT-PCR. We found that RVFV infection induced both IFNβ and OAS1 in both wild type and AMPK deficient cells although the basal levels and induction of these genes were higher in cells lacking AMPK (Figure S6A–B). This result was opposite to what would have been predicted, if IFNβ induction was responsible for the antiviral phenotype. In addition, we tested whether IFNβ treatment induced AMPK or ACC phosphorylation and found that it did not (Figure S6C, quantified in D). Altogether, these data indicate that AMPK has antiviral activity independent of the classical type I IFN response.
Since AMPK activation has antiviral activity against RVFV, we examined whether AMPK is activated by RVFV infection. To this end, we measured AMPK phosphorylation at Thr172 by immunoblot. AMPK phosphorylation was increased at 4 and 8 hours after infection compared to uninfected controls (Figure 5A, quantified in Figure S7A), indicating that RVFV infection induced AMPK activation. Furthermore, we found that UV-irradiated virus, incapable of replication (Figure S8), also induced AMPK phosphorylation at 4 and 8 hours after treatment (Figure 5C), suggesting that activation was triggered by incoming virus particles and viral replication was not required. Finally, we confirmed that LKB1 was required for RVFV-dependent activation of AMPK (Figure S9).
AMPK regulates several downstream pathways that could be important for viral infection, in particular protein translation and lipid synthesis [56]. Thus, we examined the activation status of two classical downstream effectors of AMPK involved in translation and lipid biosynthesis which are inactivated by AMPK-mediated phosphorylation [26]. Elongation Factor 2 (eEF2) is an important regulator of translation elongation, and Acetyl-CoA Carboxylase (ACC) consists of two enzymes involved in fatty acid metabolism (ACC1 and ACC2) [38], [57]. Both eEF2 and ACC had increased levels of phosphorylation at 4 and 8 hours after infection with RVFV compared to uninfected controls, consistent with the activation status of AMPK (Figure 5A, quantified in Figure S7B–D). Little difference in total protein levels of AMPK, ACC or eEF2 was observed during infection. Taken together, these data suggest that RVFV infection leads to increased AMPK signaling.
To explore the mechanism by which AMPK restricts RVFV replication, we examined the impact of AMPK on translation and lipid biogenesis, both of which contribute to important steps in virus infection. In particular, AMPK inhibits translation initiation by inactivating mTORC1, and translation elongation by inactivating eEF2 [58]–[60]. Inactivation of mTORC1 by AMPK leads to decreased translation initiation as well as increased autophagy, both of which could have anti-viral effects [58]. Since AMPK activation inhibits mTORC1 activity, we hypothesized that mTORC1, and thus protein synthesis, would be overactive in AMPK deficient cells, perhaps allowing for increased viral protein production and replication. We tested the requirement for mTORC1 signaling in RVFV infection using the mTORC1 inhibitor Rapamycin, and found no significant difference in RVFV infection in cells treated with Rapamycin compared to vehicle controls in either wild type or AMPKα1/AMPKα2−/− MEFs (Figure S10A). This finding suggests that the antiviral activity of AMPK is independent of mTORC1 signaling. Furthermore, since AMPK activation can increase autophagy, which has been shown to have antiviral effects in some models [61], we tested whether inhibition of autophagy impacted RVFV infection by plaque assay, and found no significant difference in MEFs expressing a ATG5 hairpin, which knocks down ATG5, compared to control MEFs (Figure S10B–C).
Next we investigated whether reduced translation elongation through eEF2 inactivation could be responsible for AMPK's antiviral activity against RVFV (Figure 5A). Since eEF2 is regulated by multiple upstream pathways in addition to AMPK, we first determined the sensitivity of eEF2 to AMPK regulation. In wild type MEFs, treatment with the AMPK activating drugs 2DG, oligomycin, or A769662 led to increased phosphorylation of AMPK, as well as downstream effectors eEF2 and ACC (Figure 5B, quantified in Figure S7E–H), as expected. As a control, we found that AMPK deficient MEFs did not express phosphorylated AMPK or total AMPK under any treatment condition. Interestingly, we observed an increase in phosphorylated eEF2 in response to all three drugs in AMPKα1/AMPKα2−/− MEFs (Figure 5B, quantified in Figure S7H). In contrast, while we observed an increase in ACC phosphorylation in response to drug treatments in wild type MEFs, phosphorylated ACC was undetectable in AMPK deficient MEFs both basally and in response to treatment with AMPK activating compounds (Figure 5B). These phenotypes were not due to changes in total protein levels as they remained unchanged under all treatment conditions; although the AMPK deficient MEFs had a slightly lower basal level of ACC (Figure 5B). These findings suggest signaling pathways other than AMPK are important in regulating eEF2 phosphorylation, while ACC phosphorylation is exquisitely regulated by AMPK.
Given this observation, we pursued ACC as a potential regulator of antiviral defense. ACC is the first rate-limiting enzyme and master regulator of fatty acid metabolism, both by inhibiting fatty acid biosynthesis and activating fatty acid catabolism through beta-oxidation [18], [38]. Fatty acid biosynthesis is an important component of viral infection since numerous RNA viruses, including Bunyaviruses, proliferate cellular membrane structures for proper formation of the viral replication complex, in addition to using cellular membranes for their lipid coats [6], [14], [15], [17]. In order to assess the importance of fatty acid synthesis in RVFV infection, we tested the ability of RVFV to replicate within cells pretreated with the fatty acid synthase inhibitors. Fatty acid synthase is the next enzyme in fatty acid metabolism, using the product of ACC to generate palmitate, and thus is required for all fatty acid biosynthesis [62]. We observed a 5-fold decrease in RVFV infection in the presence of fatty acid synthase inhibitors cerulenin and C75 by immunofluorescence, similar to the decrease observed in cells pretreated with the AMPK activator A769662 (Figure 5D), indicating that de novo fatty acid synthesis is an important step early in RVFV infection.
ACC is the enzyme that converts acetyl-CoA into malonyl-CoA, a precursor in the synthesis of palmitate, the first product of de novo fatty acid biosynthesis. Since AMPK activation inhibits de novo fatty acid synthesis by inactivating ACC, we tested whether altered levels of AMPK activation or expression affected cellular lipid levels. To this end, we stained MEFs with the lipophilic BODIPY fluorescent dye. We found that treatment with the AMPK activator A769662 led to a decrease in BODIPY staining compared to untreated MEFs (Figure 5E, quantified in F), consistent with decreased fatty acid synthesis during AMPK activation. In contrast, MEFs lacking AMPK had increased BODIPY staining compared to wild type cells (Figure 5G, quantified in H). These findings are consistent with previous reports that AMPK activating drugs, such as A769662 increase levels of beta-oxidation while decreasing fatty acid synthesis [46], [63], [64], and suggest that the absence of AMPK leads to overproduction of cellular lipids, while AMPK activation globally reduces cellular lipid levels.
If AMPK activation restricts RVFV infection by reducing levels of fatty acid synthesis, exogenous addition of fatty acids should restore infection. Therefore, we tested whether we could bypass the requirement for AMPK-regulated fatty acid synthesis by pretreating cells with palmitate, the first product of fatty acid biosynthesis. We treated U2OS cells with palmitate overnight, and then added A769662 1 hour prior to infection with RVFV to activate AMPK. After 10 hours of infection, cells were fixed and stained for RVFV to measure percent infection in an immunofluorescence assay that monitors the initial round of infection. In cells treated with the AMPK activator A769662 alone, we found a 5-fold decrease in RVFV infection, consistent with our previous findings (Figure 6A, quantified in 6B). However, addition of palmitate prior to treatment with A769662 was able to restore infection to levels seen in untreated cells (Figure 6A, quantified in 6B). We observed a 5-fold increase in RVFV infection in cells treated with A769662 and palmitate compared to those treated with A769962 alone (Figure 6B), while addition of palmitate alone had little effect on infection (Figure 6A–B). Since chronic exposure to high concentrations of palmitate has previously been reported to inhibit AMPK activation, we confirmed by immunoblot that AMPK phosphorylation was not inhibited by the concentrations of palmitate used in our assay (Figure S11). Together, these data suggest that AMPK restricts RVFV infection primarily through inhibiting fatty acid biosynthesis.
A dependence on lipid biosynthesis and virally induced membrane modifications is not unique to Bunyaviruses; many RNA viruses require extensive membrane modifications and proliferations to support their replication complex [6], [65]. Therefore, we tested whether AMPK restricts additional arboviruses. To this end we tested the ability of the Flavivirus Kunjin virus (KUNV), the Togavirus Sindbis virus (SINV), and the Rhabdovirus Vesicular stomatitis virus (VSV) to grow in wild type and AMPKα1/AMPKα2−/− MEFs by immunofluorescence. KUNV (Figure 7A–B), SINV (Figure 7E–F) and VSV (Figure 7I–J) had increased infections in AMPKα1/AMPKα2−/− MEFs compared to wild type MEFs. Moreover, KUNV (Figure 7C–D), SINV (Figure 7G–H), and VSV (Figure 7K–L) infections were also increased in LKB1−/−; Vec compared to MEFs expressing LKB1, indicating that both AMPK and its canonical upstream activator LKB1 restrict additional arboviruses. Moreover, we have found that KUNV is also sensitive to the AMPK activator A769662, and can be partially rescued by palmitate addition (Figure S12A–B), although palmitate treatment itself decreased KUNV infection (Figure S12C). These data suggest that AMPK may restrict multiple RNA viruses by limiting fatty acids. Taken together our data suggest that AMPK is broadly anti-viral across disparate virus families, and may represent a novel cellular target for anti-viral therapeutics.
Arboviruses represent a group of emerging pathogens of both medical and agricultural importance for which there are few therapies. RVFV is a particularly important member of this group that causes disease both in humans and livestock, and is considered a Category A pathogen due to its high pathogenesis and potential for geographical spread. Here, we identified AMPK as a novel antiviral factor that restricts RVFV infection independent of the type I IFN system. This restriction is dependent on the canonical upstream activator LKB1. Furthermore, we found that AMPK is activated by RVFV infection, and this activation restricts infection at the level of RNA replication likely by reducing fatty acid biosynthesis, an essential process in RVFV infection. We extended these studies by demonstrating that additional arboviruses, known to require lipid biosynthesis, were also restricted by this pathway. Since treatment with drugs that activate AMPK restricted infection, this could represent a novel therapeutic strategy toward the control of many RNA viruses.
AMPK is a central regulator of cellular energy that regulates a number of cellular pathways that could influence viral replication, including protein and lipid biosynthesis [56]. AMPK activation inhibits protein translation through two major downstream pathways. First, AMPK activation inhibits translation initiation by inhibiting mTORC1 activity. Second, AMPK inhibits translation elongation through inactivation of eEF2. We explored these two targets as potentially regulating RVFV infection. However, we found RVFV was insensitive to treatment with the mTORC1 inhibitor, Rapamycin, regardless of AMPK status. Furthermore, eEF2 phosphorylation induced by drugs that alter the energy status of the cell was not affected in the absence of AMPK, indicating additional upstream regulators are contributing to eEF2 activity. Therefore, we explored lipid biosynthesis as a potential target for AMPK-dependent anti-viral activity.
AMPK controls fatty acid metabolism through ACC, and may be the only physiologically relevant kinase that controls ACC activity [18]. This is consistent with our findings that ACC phosphorylation was exquisitely dependent on AMPK, in contrast to eEF2, which was phosphorylated during energy depletion even in the absence of AMPK. ACC is the enzyme responsible for the conversion of acetyl-CoA to malonyl-CoA [38]. Malonyl-CoA production impacts lipid metabolism in at least three ways. Malonyl-CoA is a substrate driving de novo palmitate production, and is also important in converting simple essential fatty acids into more complex polyunsaturated fatty acids that can be used to build triglycerides and other cellular lipids [39]. Finally, malonyl-CoA inhibits transport of fatty acids to the mitochondria, thus inhibiting fatty acid oxidation [38]. In addition to its role in fatty acid metabolism, AMPK is also an important regulator of HMG-CoA reductase (HMGCR), the rate limiting enzyme in the synthesis of isoprenoids and sterols, including cholesterol. Cholesterol is known to contribute to infection of multiple viruses, and therefore could also be targeted in AMPK-mediated virus restriction.
Since we found that fatty acid biosynthesis was required for RVFV infection, and changes to AMPK expression and activation status led to global changes in cellular lipid levels, we hypothesized that inhibiting fatty acid synthesis downstream of ACC was responsible for AMPK-mediated restriction of RVFV. This was supported by our finding that we could bypass the requirement for malonyl-CoA production by introducing exogenous palmitate. Since the addition of palmitate rescued RVFV overcoming the restriction mediated by AMPK activation (Figure 6), the ability of AMPK to inhibit fatty acid biosynthesis is likely the most important determinant of AMPK-mediated RVFV restriction. Palmitate is a substrate for the biosynthesis of a number of lipid moieties that could contribute to RVFV infection. Palmitate undergoes chain elongation and additional modifications in the ER to produce saturated fatty acids as well as triglycerides, phospholipids, and cholesterol esters [66], [67]. It is also a substrate for sphingolipid biosynthesis in the Golgi. Sphingolipids become incorporated into cellular membranes and participate in signaling events that could contribute to RVFV infection. Finally, palmitate addition is a form of post-translational modification of some proteins [68].
There are several stages during the course of RVFV infection where cellular lipids are utilized. Many RNA viruses induce the formation of novel membranous structures derived from various organelles within the cell to support the viral replication complex [6]. Notably, formation of these structures is often dependent on de novo fatty acid synthesis [69]–[72]. While RVFV-induced membrane alterations have not been well characterized, a related Bunyavirus, Bunyamwera virus, was reported to induce Golgi-derived tubular structures with globular heads in association with the viral replication complex, suggesting that other Bunyaviruses could likewise induce membrane changes [14], [15]. In addition to RNA replication, enveloped viruses bud from cellular membranes, thereby incorporating those lipids into the viral particle [17]. RVFV assembly occurs on Golgi membranes, with virus particles ultimately budding into the Golgi for transport and release at the plasma membrane [73]. Cellular lipids derived from de novo palmitate production downstream of ACC could contribute to each of these steps, although our findings that viral RNA synthesis is inhibited by AMPK suggests that RNA replication is a key target.
In addition to RVFV, we found that three additional viruses including the Togavirus SINV, the Flavivirus KUNV, and the Rhabdovirus VSV are restricted by AMPK and LKB1 (Figure 7). Importantly, this group includes members of the three major families of arboviruses that contribute to human disease. Members of the Togavirus family including Semliki Forest virus and Rubella virus have been described to induce characteristic modified endosomal and lysosomal structures termed cytopathic vacuoles that support the viral replication complex [10], [11], [74], [75]. Furthermore, a number of Flaviviruses have been shown to have important lipid dependencies. KUNV, a strain of West Nile virus, has been described as forming two distinct membrane structures that include double membrane spherical vesicles that are the sites of viral replication, as well as arrays of convoluted membranes that are the sites of viral polyprotein processing [76]–[81]. Moreover, both fatty acid synthesis and oxidation have been shown to be essential for another Flavivirus, Dengue virus (DENV). Infection is characterized by virally-induced increases in cellular fatty acid synthesis and a redistribution of the enzyme fatty acid synthase to sites of DENV replication [70]. Free fatty acids are also derived through autophagosomal processing of triglycerides, and exogenous addition of the fatty acid oleate was able to rescue DENV infection when autophagy is inhibited [82]. Furthermore, induction of ER-derived lipid droplet formation is necessary for DENV particle formation [83]. Therefore DENV and perhaps many other viruses require complex and unique interactions with cellular lipid metabolism through both synthesis and degradation pathways. In addition, Hepatitis C Virus (HCV), a distantly related Flavivirus, induces formation of a membranous web derived from intracellular vesicles, whose formation requires fatty acid synthesis for replication [8], [9]. Interestingly, AMPK has been implicated to play a role in HCV infections. AMPK-activating drugs inhibited the replication of HCV replicons concomitant with a decrease in cellular lipid levels, while knock down of the upstream activator LKB1 led to increased replication, [84], consistent with our findings with RVFV, KUNV, SINV, and VSV. Importantly, KUNV could be partially rescued from AMPK-mediated restriction by the addition of the fatty acid palmitate. Thus, AMPK may restrict multiple families of viruses through this mechanism. Since all positive strand RNA viruses are thought to induce membrane modifications for viral RNA replication, and include a large number of medically significant groups (e.g., Picornaviruses, and Coronaviruses) [16], [80], [85], [86], it will be important to determine the full scope of viruses restricted by AMPK as well as the mechanism of restriction.
Since many disparate viruses are restricted by AMPK, it is interesting to speculate how AMPK could be activated in response to these viral infections. We have found that both live virus and UV-inactivated replication incompetent RVFV is capable of activating AMPK via LKB1. This suggests that the energy sensing pathway is responsible for this activation yet we were unable to detect global changes in cellular energy levels during the period in infection when AMPK becomes phosphorylated. Thus, we hypothesize that RVFV infection induces a localized drop in cellular energy to activate AMPK. Since this is independent of viral replication and can restrict a large panel of disparate viruses that have the commonality of entering cells via endocytic routes and fusing within these compartments, we postulate that a local energy drop may occur during these steps. Since endocytosis is a highly energetic process usurped by many viruses, it is possible that increased levels could themselves could provide the trigger for this rapidly inducible antiviral response. We have previously reported that receptor-mediated endocytosis, employed by many viruses including KUNV, SINV and VSV for entry is intact in AMPK deficient cells [40]. Therefore at least some routes of endocytic entry used by viruses are unaffected by AMPK, and may provide a trigger for activation rather than a point of restriction. This would allow broad activation of AMPK by many types of viruses internalized by such routes and provide a rapid response to restrict virus infection by inhibiting fatty acid synthesis.
Since AMPK activators are currently in the clinic to treat metabolic disorders such as type II diabetes [87], and restrict RVFV and KUNV replication in cell culture, they may prove to be useful antiviral therapeutics. Several AMPK activating drugs have been shown to reduce morbidity and mortality during lethal influenza infection in mice [88]. In addition, treatment of AMPK-activating drugs inhibited infection of HCMV and HIV in cells, and the addition of AMPK-activating drugs such as Metformin to current HCV treatment regimens had promising, albeit modest, effects on reducing patient viral loads [84], [89]–[92]. Infections with HCMV, HIV, and HCV have also been shown to inhibit AMPK activity [56], [84], [89], [92]. AMPK may have multiple effects on these infections since different downstream mechanisms have been implicated [56], [84], [89], [92], [93]; however, this suggests the possibility that some viruses have developed mechanisms of immune evasion that target AMPK. Taken together, AMPK plays a broad role in cellular innate immunity through potent inhibition of fatty acid synthesis, which is broadly utilized by viruses, suggesting that AMPK and perhaps other modulators of lipid biosynthesis are potential targets for broad pan-antiviral therapeutics.
MEFs, BHK and U2OS cells were maintained at 37°C in DMEM supplemented with 10% FBS (Sigma), 100 µg/ml penicillin/streptomycin, 2 mM L-glutamine, and 10 mM Hepes. LKB1−/− MEFs [94] were complemented with MIGR (Vector) or FLAG-LKB1-MIGR (LKB1 cDNA) retrovirus and sorted on GFP+ cells by FACS as previously described [40]. Rift Valley fever virus MP-12 was grown in Vero-E6 cells supplemented with 10% FBS [51]. RVFV was UV-inactivated in a Stratalinker. KUNV (gift from M. Diamond) was grown in BHK cells. VSV-GFP [95] was grown in BHK cells as described [96]. SINV-GFP virus [97] was grown in C636 cells [98]. All viruses were tittered by plaque assay in BHK cells. Antibodies were obtained from the following sources: anti-RVFV ID8 (gift from C. Schmaljohn USAMRIID), anti-KUNV 9NS1 (gift from R. Doms), anti-tubulin (Sigma), and anti-P-AMPK, t-AMPK, P-ACC, t-ACC, P-eEF2, t-eEF2 (Cell Signaling Technology). Fluorescently labeled secondary antibodies and BODIPY-TR were obtained from Invitrogen. HRP-conjugated antibodies were obtained from Amersham. A769662 was obtained from Santa Cruz. Other chemicals were obtained from Sigma.
Viruses were plaqued on MEFs as indicated. Confluent monolayers were treated with serial dilutions of virus for two hours, after which the viral inoculums were removed, and cells were overlayed with 0.75% agarose in MEM, and incubated at 37°C for 48 hours. Cells were fixed in 10% formaldehyde, and stained with crystal violet. Plaque number was determined manually, and plaque diameter was measured using MetaXpress software and used to calculate areas.
For all infections, washes and media changes were performed in the control untreated wells, as well as those infected with virus. Viral immunofluorescence experiments were performed in 96 well plates as previously described [99]. Briefly, cells were grown overnight in 96 wells plates, media was removed and fresh media was added. When appropriate, drug was added at the indicated concentration in 5 µl PBS, and cells were incubated at 37°C for 1 hour before addition of virus. Cells were infected with the indicated MOI of virus in complete media and spinoculated for 1 hour at 1200 RPM, and incubated at 37°C. Cells were fixed and processed for immunofluorescence as previously described 10 hours post infection for RVFV, SINV, and VSV, and 24 hours post infection for KUNV unless otherwise indicated [100]. Briefly, cells were fixed in 4% formaldehyde/PBS, washed twice in PBS/0.1% TritonX-100 (PBST), and blocked in 2% BSA/PBST. Primary antibodies were diluted in block, added to cells, and incubated overnight at 4°C. RVFV was stained with anti-RVFV ID8; KUNV was stained with anti-KUNV 9NS1. VSV and SINV expressed GFP, and did not require antibody staining. Cells were washed three times in PBST, and incubated in secondary antibody with Hoescht33342 (Sigma) counterstain for one hour at room temperature. Plates were imaged at 10× using an automated microscope (ImageXpress Micro), capturing four images per well per wavelength, and quantification was performed using MetaXpress image analysis software. Significance was determined using a Student's T-test. For immunofluorescence assays, a minimum of three wells per condition was imaged, with four images taken per well. To control for variability in baseline level of infection, a Student's T-test was performed on both the raw percent infection data in each individual experiment, and across a minimum of three replicate experiments where the untreated control had been normalized. Significance was determined if p<0.05 in all tests.
MEFs were infected with RVFV MOI 1 in 6 well dishes and incubated at 37°C. Two hours post infection, inoculums was removed, and fresh medium was added. At indicated time point, medium was removed from infected cells and tittered on BHK cells by plaque assay.
MEFs were grown overnight in a 6 well dish. Medium was replaced with 1 mL of fresh complete medium and cells were chilled to 4°C for 10 minutes. RVFV (MOI 10) was added on ice, and cells were incubated at 4°C for 1 hour to allow virus binding. Cells were washed in PBS, then treated with either PBS or 0.25% trypsin to remove bound virus as previously described [101]. Cells were pelleted, then washed again, and lysed in Trizol to extract total RNA. Samples were then prepared for quantitative RT-PCR. cDNA was prepared from total RNA using M-MLV reverse transcriptase (Invitrogen) random primers, and transcripts were amplified by quantitative PCR. ΔΔCT was calculated for RVFV S segment using GAPDH as a cellular loading control.
Time of addition experiments were performed as previously described [51]. U2OS cells were grown overnight, and the media was replaced. Cells were infected with RVFV (MOI 1), spun at 1200 rpm for 1 hour, and subsequently incubated at 37°C. 12 mM 2DG, 200 µM A769662, or 12 mM Ammonium Chloride were added either 1 hour prior to infection (−1), with infection (0), or 1, 2 or 4 hours after infection. 10 hours post infection cells were fixed in 4% formaldehyde in PBS and processed for immunofluorescence. Significance was determined using a Student's T test.
MEFs were infected with RVFV MOI 1 in 6 well dishes (∼50% infection) and incubated at 37°C for indicated time point. For protein analysis, cells were washed briefly in cold PBS and lysed in NP40 lysis buffer supplemented with protease (Boehringer) and phosphatase (Sigma) inhibitor cocktails. Samples were separated by SDS-PAGE and blotted as described [69]. HRP-conjugated secondary antibodies and Western Lightening Chemiluminescence Reagent were used for visualization. To analyze downstream effectors of AMPK, MEFs were treated with 12 mM 2DG, 10 µM oligomycin, or 100 uM A769662 for 4 hours, lysed and blotted as above.
For RNA analysis, cells were lysed in Trizol buffer, and RNA was purified as previously described [100]. To detect viral mRNA, total RNA from infected cells was separated on a 1% agarose/formaldehyde gel and blotted with the indicated probes as previously described [100]. Samples were quantified and normalized against controls using ImageQuant software.
Cellular lipids were stained as previously described [82], [84]. MEFs were grown to confluence overnight, and then treated with PBS vehicle or 100 µM A769662 for 10 hours. Cells were fixed in 4% formaldehyde for 10 minutes and washed three times in PBS. Staining was performed with 10 µg/ml BODIPY-TR and counterstained with Hoescht33342 in 100 mM glycine in PBS overnight. Cells were washed three times in PBS and imaged using the ImageXpress Micro automated microscope. Integrated intensity of BODIPY signal per cell area was calculated using MetaXpress image analysis software. Significance was determined using a Student's T test.
Exogenous palmitate addition was performed as previously described [102]. Delipidated Fetal Calf Serum and Albumin-bound palmitate were prepared as described [102] and obtained as a kind gift from Robert Rawson. U2OS cells were set up on day 0 in 96 well plates and grown over night in normal growth medium. On day 1 medium was removed and cells were washed briefly in PBS. Cells were treated with low glucose DMEM supplemented with 5% delipidated Fetal Calf Serum with or without 100 µM Albumin-bound palmitate, and incubated overnight. On day 2 cells were treated with 100 µM A769662 or PBS vehicle for 1 hour, and infected with RVFV for 10 hours. Cells were fixed, processed for immunofluorescence, and imaged at 10× using the automated microscope ImageXpress Micro, as described above. Quantification was performed using MetaXpress image analysis software. Significance was determined using a Student's T-test.
|
10.1371/journal.pntd.0003338 | Influence of Household Rat Infestation on Leptospira Transmission in the Urban Slum Environment | The Norway rat (Rattus norvegicus) is the principal reservoir for leptospirosis in many urban settings. Few studies have identified markers for rat infestation in slum environments while none have evaluated the association between household rat infestation and Leptospira infection in humans or the use of infestation markers as a predictive model to stratify risk for leptospirosis.
We enrolled a cohort of 2,003 urban slum residents from Salvador, Brazil in 2004, and followed the cohort during four annual serosurveys to identify serologic evidence for Leptospira infection. In 2007, we performed rodent infestation and environmental surveys of 80 case households, in which resided at least one individual with Leptospira infection, and 109 control households. In the case-control study, signs of rodent infestation were identified in 78% and 42% of the households, respectively. Regression modeling identified the presence of R. norvegicus feces (OR, 4.95; 95% CI, 2.13–11.47), rodent burrows (2.80; 1.06–7.36), access to water (2.79; 1.28–6.09), and un-plastered walls (2.71; 1.21–6.04) as independent risk factors associated with Leptospira infection in a household. We developed a predictive model for infection, based on assigning scores to each of the rodent infestation risk factors. Receiver operating characteristic curve analysis found that the prediction score produced a good/excellent fit based on an area under the curve of 0.78 (0.71–0.84).
Our study found that a high proportion of slum households were infested with R. norvegicus and that rat infestation was significantly associated with the risk of Leptospira infection, indicating that high level transmission occurs among slum households. We developed an easily applicable prediction score based on rat infestation markers, which identified households with highest infection risk. The use of the prediction score in community-based screening may therefore be an effective risk stratification strategy for targeting control measures in slum settings of high leptospirosis transmission.
| The Norway rat is an important reservoir for urban leptospirosis, a life-threatening zoonotic disease. In urban settings, leptospirosis transmission occurs primarily in the peri-domiciliary environment of the slums. Rodent control is one of the most frequent strategies to prevent leptospirosis, but the identification of domiciles at higher risk of transmission is challenging. We compared households where an individual with evidence of recent leptospirosis infection resided and households where none of the residents had evidence for infection. Houses with evidence of leptospirosis transmission had higher levels of rodent infestation and environmental conditions related to rodents. We propose a new methodology to easily characterize slum households, based on environmental characteristics, at different levels of risk for leptospirosis transmission. The findings of this study indicate that evaluation for rodent infestation intensity and environmental features may be a feasible strategy for targeting augmented control measures for leptospirosis.
| In developing countries, leptospirosis is an emerging health problem affecting urban slum communities [1]–[4]. Annual epidemics of the disease typically occur during periods of seasonal rainfall [1], [5]–[8]. Lack of sanitation infrastructure such as open sewage systems and poor refuse collection services provide conditions for proliferation of rats, which are the main reservoir for leptospirosis in urban settings [1], [9]–[11]
Pathogenic Leptospira infection produces a broad spectrum of clinical manifestations with case fatality exceeding 10% and 50% for Weil's disease and severe pulmonary hemorrhage syndrome, respectively [2], [12], [13]. Currently, there are no effective interventions which can be easily implemented in slum communities to prevent leptospirosis transmission. Rat-control programs are commonly implemented as a control measure for leptospirosis in many cities, such as those in Brazil, but their effectiveness is questionable and has not been systematically explored.
In this setting, two rat species, the Norway rat (Rattus norvegicus) and black rat (Rattus rattus), are the main reservoirs for this bacterium and contaminate environments via urinary shedding, providing conditions for transmission to humans [2]. Prior studies in urban areas have shown that Leptospira carriage ranges between 7–82% for R. norvegicus [14], [15] and between 7–34% for R. rattus [16], [17]. However, the Norway rat is far more common within the urban slum environments: nearly 100% of rats trapped in the city of Salvador, Brazil comprised of this species [14], [15], [18]. Leptospira strains isolated from Norway rats were genotypically identical with strains obtained from human patients based on PCR-based typing methods [19]. Additionally, epidemiological studies have found that peri-domiciliary resident reporting of rat sightings and living in proximity to open sewers placed residents at increased risk for leptospiral transmission in slum areas [5], [10]. These findings support the role of urban peri-domestic transmission due to contact with water contaminated with rat urine.
Rodent control programs based on environmental application of a chemical rodenticide [20] as an strategy to reduce the incidence of leptospirosis are costly and their effectiveness has not been evaluated [20], [21]. Programs implemented in Brazil [20] are based on the Centers for Disease Control and Prevention (CDC) approach for pest management [22] which includes an environmental form to assess rodent infestation levels and infrastructural deficiencies in peridomestic areas. Nevertheless, the CDC survey form has not been validated for application in slum areas of developing countries. Furthermore, no studies have systematically examined whether indicators of rodent infestation assessed during rodent surveys can be used to predict leptospirosis risk and, therefore, guide targeted interventions specifically implemented among high risk households. Herein, by using a community-based cohort study aimed to evaluate Leptospira infection, we describe an environmental and rodent survey instrument, adapted from CDC guidelines, for use in a tropical slum area. In addition, we developed and evaluated the accuracy of a scoring system to predict Leptospira transmission using data easily produced by this instrument.
This study was conducted in Pau da Lima, a slum community situated in the periphery of Salvador, a city of 2.7 million inhabitants [23] in Northeast Brazil. This site was selected for epidemiological studies on leptospirosis based on the high annual incidence of severe leptospirosis (35.4 cases per 100,000 pop.) identified in this community by active surveillance during1996 to 2002. The study site is a four valley area of 0.46 km2, characterized by the absence of basic sanitation and high levels of rat infestation [10] (Figure 1A–B). In 2003, we conducted a census in the area and identified 14,122 residents living at 3,689 households. The socioeconomic profile in this site was similar to other slum populations in Brazil: subjects were squatters (85%) who did not complete primary school (77%) and had a median household per capita income of $1.30 per day.
We performed a case-control study of households in the Pau da Lima community in order to evaluate the association between household environmental characteristics and rodent infestation on household-level Leptospira infection risk. Households were selected among those which participated in a prospective community-based cohort study designed to characterize the burden of Leptospira infection [11]. This cohort investigation, performed between 2003 and 2007, comprised in part of four annual serosurveys of 2,003 cohort subjects who were greater than five years of age and resided in 684 (18.5% of 3,689) randomly-selected households at the study site [11].
The microscopic agglutination test (MAT) was performed on serum samples from the baseline survey (initiated in 2003 and completed in 2004) and follow-up surveys (2004/2005, 2005/2006 and 2006/2007) to identify subjects with serologic evidence of a recent Leptospira infection. As previously described [10], [11], MAT evaluations were performed with a panel of five reference strains and two clinical isolates [1], which included L. interrogans serovars Autumnalis, Canicola and Copenhageni; L. borgspetersenii serovar Ballum, and L. kirschneri serovar Grippotyphosa. All sera were screened at dilutions of 1∶25, 1∶50 and 1∶100. Positive samples at a dilution of 1∶100 were titrated to determine the endpoint agglutination titer. A recent leptospiral infection was defined as seroconversion during which the MAT titer increased from negative at the baseline survey to a titer ≥1∶50 during the follow-up survey or as a four-fold rise in MAT titer in a participant with a titer of ≥1∶25 during the baseline survey [11].
All cohort members provided written informed consent before enrollment. Minors (<18 years of age) provided assent to participate, in addition to informed consent from their legal guardian. Ethical clearance for this study was provided by the Ethical Committee in Research of the Oswaldo Cruz Foundation and IRB committees of Weill Medical College of Cornell University and the Yale School of Public Health.
A case household was defined by occurrence of at least one Leptospira infection event among subject residents of the household during the 3 years of follow-up. Identification of case households were performed after completion of the three years of cohort follow-up. A random number table was used to select control households (1∶1 case∶control ratio) among those at the Pau da Lima site which fulfilled the following criteria: a) absence of Leptospira infection event among cohort subjects who were members of the household, b) the presence of at least one household member who participated for the duration of the cohort study, and c) households situated ≥30 m from the nearest case household. The criterion of 30 m was selected to minimize the possibility of overlapping rat infestations as the typical home range of R. norvegicus varies between 30–50 m in urban areas [24], [25]. Control households were selected after completion of the three years of cohort follow-up, just after identification of case households.
Environmental surveys of case and control households were conducted by a team of rodent control specialists from the municipal Zoonosis Control Center (ZCC) in Salvador. Study houses and peri-domestic areas (10 m around each household) were surveyed between October and November of 2007. The survey team used a modified exterior inspection form, adapted from the CDC manual (Table S1) [22]. The form included the following six groups of variables: a) 7 variables on premise type; b) 5 variables on food sources for rodents; c) 3 variables on water sources for rodents; d) 11 variables on harborage for rodents; e) 5 variables on entry/access for rodents and f) 6 variables on signs of rodent infestation (Table S1, available in English and Portuguese; Figure 1C–F). Team performing the surveys was blinded regarding household case status.
In addition to the environmental survey, we administered a standardized questionnaire to the head-of-the-household, which included 4 questions on demographics and 4 on ownership and number of domestic animals. Because of the time lag between occurrence of Leptospira infection and the household rodent survey, we asked the head-of-household if domicile or peridomicile structure changes had occurred (i.e. rebuilding or expansion), if nearby open sewers were closed/created and if refuse deposits were removed/created since the year of Leptospira infection. Year of infection for case households was available to the field team in order to perform the questionnaire, and to maintain blinding, we generated random years to serve as sham infection dates for surveys of control households.
During September 2007, an environmental inspection was performed in the entire study area. Location of study households, in addition to site and size of open refuse deposits, open sewage and rainwater drainage systems, were geocoded and entered in a Geographic Information System (GIS) mapping database, as described previously [10], [11].
Epidemiological and laboratory data were double-entered and validated using Epi-Info for Windows software (Centers for Disease Control and Prevention, Atlanta, GA). There were no missing values for any of the analyzed variables. We used proportions and medians with interquartile range to characterize signs of rodent infestation in case and control households.
Concordance between specific markers of current Rattus novergicus infestation (rat feces) and variables indicating present or prior rodent infestation (burrows and runs) was assessed by the kappa index statistic. We used Chi-square and Wilcoxon rank sum tests to compare socio-demographic and environmental characteristics of case and control households for categorical and continuous data, respectively. We used the same tests to compare households with and without domicile or peridomicile structural changes in the period between serological and environmental evaluations. These last analyses were performed for the groups of case and control households separately.
Environmental variables with p<0.1 in the bivariate analysis were included in a multivariate logistic regression analysis to identify independent predictors of risk for Leptospira transmission. As some of the exposure variables studied was correlated at two conceptual levels of risk, firstly, environmental variables that influence the rodent infestation level and secondly, variables that measure proxies for rodent infestation, we used a hierarchical approach [26], [27] to identify independent predictors of risk for Leptospira transmission. We built three multivariate logistic regression models using backward elimination. The first model included environmental variables previously described as associated with rodent infestation, such as refusal deposit, sewage and vegetation. The second model included rodent infestation variables, such as rodent runs, feces and burrows. The third and final model included variables retained (P<.05) from the first and second models. We used SAS software for Windows for these analyses [28].
To develop a practical prognostic risk score for each household, we weighted independent variables identified by logistic regression proportionally to their β regression coefficient values, as previously described [29]. The use of mutually adjusted weights per predictor is the standard methodology to develop a prognostic risk score [30]. A risk score was calculated for each household. We assessed the discriminative power of the score by using receiver operating characteristic (ROC) curves of sensitivity and specificity. Sensitivity and specificity measured the proportion of case and control households, respectively, which were correctly identified as such by the risk score. Score predictive ability (C-statistics) was classified as excellent (>0.80), good (0.70–0.79), fair (0.60–0.69), and poor (0.50–0.59).
As previously described, we enrolled 2,003 participants from 684 households in a community cohort study designed to measure risk factors and infection rates for leptospirosis [11]. Of these, 1,585 (79%), 1,324 (66%) and 1,394 (70%) participants completed the first, second and third year of follow-up, respectively. We identified 104 Leptospira infections in 97 participants residing in 80 households during the three year study period. We identified fifty-one infections (49%) in the first year, 26 (25%) in the second and 27 (26%) in the third year. In all but four cases (96%), L. interrogans serogroup Icterohaemorrhagiae was identified as the presumptive infectious serogroup based on agglutination titers. A majority of the case households (63, 79%) had a single participant with evidence of infection, while 17 (21%) case households had two participants. Seven participants had serologic evidence of two exposures. In addition to the 80 households defined as case-households we also included 115 households, out of a possible 186, meeting inclusion criteria as controls.
We performed environmental surveys in 189 (97%) of the 195 households (80 case and 109 control households). Six control households could not be inspected because no persons were present during at least three attempted visits. The final number of case and control households included in this study was 80 and 109, respectively. We observed that case households had a higher number of inhabitants and number of subjects enrolled in the cohort study than control subjects (4 [IQR: 4–6] vs. 4 [3]–[5], respectively, Table 1; and 4 [IQR: 2–5] vs. 3 [2]–[4], respectively, P<0.05). Other characteristics regarding environment and rodent signs, of case and control households with comparative bivariate values of p<0.1 (inclusion criterion for logistic regression analyses), are shown in Table 1.
We detected rodent infestation (presence of at least one rodent sign) in 63 (78%) of the case and 46 (42%) of control households. Rat burrows were frequent signs of rodent infestation (65% and 29% for case and control households, respectively). A total of 101 among the 189 households surveyed, had fecal droppings, 77% were from R. norvegicus, 10% from M. musculus and 4% from R. rattus; feces from a non-identified species were present at 9% households. Overall, the presence of R. norvegicus fecal droppings had good concordance with presence of any rodent burrow (kappa = 0.61) and a moderate concordance with any rodent run (kappa = 0.51).
Eighty (42%) households, 35 cases and 45 controls, had structural, sewage or trash modifications between the year of Leptospira infection and date of environmental survey. Case households with modifications were compared with case households lacking modifications. The same analysis was performed for control households. There were no significant differences among the study variables between groups (data not shown) and all 189 households were considered for further analyses.
In bivariate analyses, we identified 13 environmental variables which were associated (P<0.05) with case households (Table 1). A larger percentage of case households (P<0.01) showed signs of rodent infestation related to R. norvegicus. Additionally, the presence of a case of Leptospira infection in a household was associated with low socioeconomic status as per capita income and number of inhabitants in the house. Residents of a large number of case and control households were squatters, 91% and 86% respectively. We did not identify significant differences between case and control households with respect to the presence or number of dogs, cats or chickens (data not shown). Residents also reported ownership of other species of animals as ducks, small birds, rabbits, hamsters, monkeys and turtles, but the presence/number of these animals was not associated with differences in Leptospira infection among residents of case and control households.
The first multivariate logistic regression model, including variables related to household environment, retained the following characteristics: rodent access to water, domicile built on a slope and un-plastered exterior wall surfaces. The second model, including rodent infestation variables, retained the presence of R. norvegicus fecal droppings and rodent burrows. The final model retained four variables: two household environmental variables and two rodent infestation factors (Table 2). R. norvegicus fecal droppings had the strongest association with case households in the final model followed by rodent burrows, rodent access to water and un-plastered exterior wall surface.
To build a risk score, we assigned numerical scores to each of the four independent variables from the final model proportionate to the regression coefficient for each variable (Table 2). The sum of the number was used to classify each household into ten categories ranging from 0 to 9. None of the households received 1 or 8 points. Five percent of the case households and 30% of the control households had a score value of 0. Because score values were not normally distributed within case and control households, we used Wilcoxon rank-sum tests to compare the scores by case status. The median risk score for case households was 7, statistically different from the value of 2 for control households (p<0.001). Receiver operator curve (ROC) analysis yielded a very good to excellent c statistic of 0.78 (95 percent confidence interval: 0.71–0.84) (Figure 2). Table S2 presents the sensitivity, specificity and the estimated proportion of the case and cohort households for each score level.
High levels of rodent infestation and the predominance of Rattus norvegicus are frequent features within urban slum areas, in Brazil and around the world. Efforts to implement and improve rodent management interventions to reduce urban leptospirosis have been hampered by the lack of readily available information and epidemiologically-based markers that allow identification and monitoring of households at increased risk for infection. Our study demonstrates not only the large proportion of houses in a typical Brazilian slum at risk of acquiring Leptospira transmission, but also that the risk is significantly associated with four markers of rodent infestation and environmental factors (R. norvegicus feces, rodent burrows, access to water, un-plastered walls). The risk score system we developed, by weighting and combining values for each of these features, accurately classified households into risk groups for Leptospira infection with high precision. Of note, similar household-based markers of rodent infestation and infection risk have been described in association with Lassa fever [31] and hantavirus pulmonary syndrome [32]. We propose that targeting households at higher predicted risk for leptospirosis for augmented chemical, environmental and educational interventions would result in the greatest reduction of Leptospira transmission.
Our summary score value for high risk environments is easy to use and increases accuracy in identifying high risk households. The score performed well in discriminating case from control households with an accuracy of 0.78 and a sensitivity and specificity at a point value of 3 of 80% and 60%, respectively. These findings are encouraging and suggest that this tool could help inform more aggressive rat control to household locations with similar risk profiles without increasing the current workload of zoonotic control assessment teams.
Our scoring method could benefit other cities in Brazil and other countries where rodent control programs are the principal strategy to decrease leptospirosis incidence [20]. Rodent control programs are time-consuming and expensive, as they require large numbers of trained persons. In Salvador alone, the ZCC programs prioritize target areas with 20,000–60,000 households in locations where a high incidence of leptospiral disease (21.4 cases per 100,000 pop. in 2008 [unpublished data]) has been detected. Considering the coverage of the ZCC program in cities such as Salvador in Brazil and the leptospirosis incidence in these urban centers, it may be feasible to use changes of leptospiral disease incidence to evaluate the efficacy of enhanced targeted rat-control strategies which include the proposed score.
Although our study focused on a single area within Salvador, it would be useful to assess the utility of our measure to predict risk of leptospiral infection in other slum areas. However, this will require some tailoring of the survey instrument to some other locations, as every slum has unique characteristics and significant socioeconomic and environmental heterogeneities [33] within the broad definition for slum settlement as proposed by the United Nations [34]. Additionally, information regarding the incidence of leptospiral infection as determined by annual serosurvey is costly and labor-intensive, consequently such studies are a rarity. In the few reports where geocoded data for leptospirosis are available [35], spatial information has been restricted to outcomes of hospitalized cases of severe leptospirosis. The absence of other prospective studies to evaluate Leptospira infection prevented us from performing an external validation to evaluate the accuracy of our score in other settings. However, even with these limitations it may be possible to test the external validity of our survey methods and subsequent risk-scoring within a large area of Salvador and within other Brazilian cities where up to 33% of the urban population has equal or greater levels of poverty as found in our study community [23].
We identified two environmental risk factors, access to water and un-plastered walls, which were associated with an increased risk of Leptospira infection. Our finding that the presence of standing water, including sewer water, increases the risk of acquiring Leptospira infection, builds on our group's prior findings that household proximity to open sewers is associated with both Leptospira infection and severe disease in humans [5], [10], [11]. This may reflect the need of rats to a ready access of water as suggested by previous studies [36]–[38], and the capacity of open sewers to serve as environmental features contributing to the risk of leptospiral transmission to humans [9]. Infrastructural deficiencies, such as the presence of un-plastered walls in the home, were significantly associated with case households, and are a characteristic presumed to increase the probability of rat ingress into residences. However, Rattus norvergicus, which was the predominant rodent in the study area, typically resides outdoors and consequently it is more probable that un-plastered walls serve as a proxy for socioeconomic status than a proxy of rodent infestation. In conjunction, these specific household environmental and rodent infestation characteristics showed to be objective markers of Leptospira transmission.
This study provides further evidence of the importance of rats in urban leptospirosis transmission. The high household infestation rate, elevated Leptospira prevalence [15] and long term carriage [39] making R. norvegicus a major reservoir host for leptospires. We did not evaluate Leptospira carriage in domestic animals such as dogs. It is possible that dogs, which are often found in poor urban communities and may be infected with serovar Copenhageni [40], could contribute to the transmission cycle. However we think this is unlikely in our study area because this and other studies did not identify an epidemiological link between dogs and human Leptospira infection or leptospirosis [5], [10], [11].
We successfully identified environmental features associated with higher infection risk, but our study was limited by the time lag between occurrence of Leptospira infection and assessment of housing rodent survey. However, we showed that the four major causes of temporal modification in a household or environs (i.e. rebuilding or expansion; closing or creation of a nearby open sewer and removal or creation of refuse deposits since the year of Leptospira infection) did not influence environmental or infestation measures among case or control households. We did not evaluate the characteristics of the places where participants with Leptospira infection worked. However previous studies have found strong associations between risk of leptospiral transmission and environmental conditions around the case household, irrespective of work location [5], [10], [11], so we believe our findings are plausible and consistent.
We used serologically confirmed cases of subclinical Leptospira infection to define case households. Notwithstanding, it is likely that mild or subclinical Leptospira infection and clinical disease share the same environmental risk exposures. A previous study showed that members of households living with an index case of clinical leptospirosis were more likely to have serologic evidence for a prior infection than members of other households in the same communities [41]. Additionally, environmental deficiencies such as presence of open sewer near of the household and sighting rats in the peridomiciliary environment were independent risk factors for both severe leptospirosis [5] and Leptospira infection [10], [11].
In conclusion, we developed a risk score based on four variables related to objective signs of rodent infestation and environmental features that predict risk of Leptospira infection among persons living in households located within urban slums of Salvador. These findings have the potential to better inform policymakers and rodent management programs by identifying high-risk households and areas for frequent interventions and reducing effort directed at lower risk households and neighborhoods.
|
10.1371/journal.ppat.1000484 | Vector Transmission of Leishmania Abrogates Vaccine-Induced Protective Immunity | Numerous experimental vaccines have been developed to protect against the cutaneous and visceral forms of leishmaniasis caused by infection with the obligate intracellular protozoan Leishmania, but a human vaccine still does not exist. Remarkably, the efficacy of anti-Leishmania vaccines has never been fully evaluated under experimental conditions following natural vector transmission by infected sand fly bite. The only immunization strategy known to protect humans against natural exposure is “leishmanization,” in which viable L. major parasites are intentionally inoculated into a selected site in the skin. We employed mice with healed L. major infections to mimic leishmanization, and found tissue-seeking, cytokine-producing CD4+ T cells specific for Leishmania at the site of challenge by infected sand fly bite within 24 hours, and these mice were highly resistant to sand fly transmitted infection. In contrast, mice vaccinated with a killed vaccine comprised of autoclaved L. major antigen (ALM)+CpG oligodeoxynucleotides that protected against needle inoculation of parasites, showed delayed expression of protective immunity and failed to protect against infected sand fly challenge. Two-photon intra-vital microscopy and flow cytometric analysis revealed that sand fly, but not needle challenge, resulted in the maintenance of a localized neutrophilic response at the inoculation site, and removal of neutrophils following vector transmission led to increased parasite-specific immune responses and promoted the efficacy of the killed vaccine. These observations identify the critical immunological factors influencing vaccine efficacy following natural transmission of Leishmania.
| The generation of vaccines that protect against intracellular pathogens such as malaria, human immunodeficiency virus and leishmaniasis have met with limited success. A perplexing aspect of this failure as it relates to leishmaniasis is the knowledge that individuals typically get the disease only once, and that individuals who are experimentally infected with cultured parasites are protected against sand fly transmitted infection, thereby providing a “gold standard” for vaccine design. Many engineered, non-living vaccines have been developed to mimic the immune response observed in protected individuals and some of these have been shown to provide excellent protection against needle inoculation of Leishmania parasites in mice. However, very similar vaccine formulations adapted for use in people have failed to protect against natural exposure to infected sand fly bites. In the present study, we attempt to reconcile these long-standing differences, and to provide the critical correlates of immunity that will predict vaccination success against natural exposure.
| Leishmania are obligate-intracellular protozoan parasites that establish infection in mammalian hosts following transmission to the skin by the bite of an infected Phlebotomine sand fly [1]. Different Leishmania species are associated with a spectrum of clinical outcomes in humans, including fatal, disseminated infection of the spleen and liver following infection with L. donovani, and self-curing cutaneous lesions associated with L. major and other cutaneous strains. Healed cutaneous lesions often result in a permanent scar that has been shown to harbor low numbers of parasites over the long term [2]. While this chronic, sub-clinical state can serve as a long-term reservoir for disease, it also maintains powerful protective immunity for the host, as individuals with healed primary lesions are highly resistant to re-infection, and complete elimination of a primary infection in animal models results in susceptibility to reinfection [3],[4]. Deliberate needle inoculation with viable parasites in a selected site, referred to as “leishmanization,” has been employed extensively as a live “vaccine” in people for generations, and is highly effective against natural exposure [5],[6],[7],[8]. However, due to reports of adverse reactions at the site of inoculation, quality control issues, and concerns over causing serious disease in immuno-compromised individuals, leishmanization has fallen out of favor [8],[9]. Employing the mouse model of L. major infection, numerous non-living [10],[11],[12],[13],[14],[15] and live-attenuated [13],[16],[17], or DNA-based [10],[18] vaccine formulations have been developed as alternatives to leishmanization, which in many cases have conferred relatively long-term protection against experimental needle challenge [10],[11],[12],[18]. In contrast, non-living vaccines, including formulations similar to those shown to work effectively in mice against needle challenge [11],[13], have yet to confer significant protection against natural exposure in people, despite the generation of measurable cell-mediated immunity [9],[19],[20],[21],[22],[23],[24],[25],[26],[27],[28]. This contradiction between the results in humans and animal trials suggests that the correlates of vaccine efficacy developed mainly from the mouse model, namely the generation of Th1 responses and the reduction of lesion size and/or parasite number following needle challenge, may not adequately define the requirements for protection against natural transmission. Observations by Rogers et al. [29], in which vaccination with soluble leishmanial antigen plus IL-12 delayed the onset of progressive lesions following needle, but not infected sand fly challenge in BALB/c mice, support this suggestion.
In addition to the delivery of infectious stage parasites into the dermis, sand flies also deposit pharmacologically active saliva, which aids in blood feeding, and egest parasite-released glycoconjugates, which accumulate behind the mouthparts in infected flies and form a promastigote secretory gel (PSG). These molecules have been shown to enhance the severity of disease when co-administered with infectious stage parasites [30],[31],[32],[33]. We have recently reported that sand fly transmission induces a qualitatively unique inflammatory response at the localized bite site that includes a dynamic recruitment of neutrophils, and that these neutrophils markedly enhance the ability of parasites to establish primary infection [34]. Thus, an analysis of the influence of sand fly transmission on vaccine efficacy is likely to be highly relevant to the generation of a Leishmania vaccine that is effective in people.
Healed primary L. major infection initiated by needle inoculation of mice has been extensively employed as a model that mimics the clinical practice of leishmanization. Mice with resolved primary lesions harbor L. major specific CD4 T cells that simultaneously produce IFN-γ, TNF-α, and IL-2 effector cytokines and mount powerful protective immunity at a site of needle re-challenge, resulting in the rapid control of parasite growth [13],[35]. In order to characterize the protective immune response following natural transmission, 4 P. duboscqi sand flies, infected with L. major (L.m.-SF), were allowed to feed on the ears of C57BL/6 mice with a healed primary lesion in the footpad. Under these conditions, a median of 2 flies will show evidence of blood engorgement, thereby ensuring parasite transmission to a sufficient number of ears to conduct the experiment, while at the same time more faithfully replicating natural transmission, which likely occurs following exposure to a single infected fly. At 1 and 3 days following exposure to the infected flies, a slight but significant increase in infiltrating CD4 T cells was found in the ears of healed mice relative to fly challenged, naïve, age-matched controls (AMC) (Figure 1A). At 7 days post-challenge, the number of infiltrating CD4 cells in the healed mice was dramatically increased relative to controls. In order to determine if parasite antigen was required to mediate this recruitment, healed mice were also exposed to uninfected sand fly bites (SF). Both infected or uninfected bites recruited equivalent numbers of T cells at day 3 post-bite, however, parasite antigen appeared necessary for the dramatic increase observed on day 7 (Figure 1A). Remarkably, Ag re-stimulation of dermal derived cells revealed Leishmania-specific IFN-γ producing CD4+ T cells at the challenge site within 24 hours, a response that gradually increased to 17% of the total CD4 T cell population at 7 days (Figure 1B), correlating with a >100 fold reduction in parasite numbers in the skin (Figure 1C). Antigen re-stimulation of T cells from the ears of healed mice exposed to uninfected sand fly bites also revealed the presence of L.m.-specific IFN-γ producing CD4+ T cells (Figure 1B), suggesting that a functional property of these effector cells is their ability to rapidly migrate to sites of tissue inflammation whether antigen is present or not.
Vaccination with autoclaved L. major (ALM), or a recombinant leishmania protein, plus CpG oligodeoxynucleotides (ODN) has been shown to effectively protect against needle challenge with L. major in mice [11],[13]. We therefore employed ALM+CpG to test the efficacy of a non-living vaccine against natural transmission. Mice vaccinated with ALM+CpG three times s.c. in the footpad at two week intervals, along with age-matched naïve controls and mice with healed primary lesions, were exposed to the bites of 4 infected sand flies twelve weeks following the last vaccine injection. Four weeks following infected sand fly exposure or needle inoculation, coincident with the time of peak parasitic load in naïve mice, parasite burden in the ear dermis was assessed. Mice with healed primary lesions again dramatically controlled parasite growth following exposure to the bites of infected sand flies (Figure 2A). In contrast, ALM+CpG vaccination conferred no protection against transmission by sand fly bite, despite conferring strong protection against needle inoculation. Ear lesion measurements obtained 4 weeks after infection also revealed a compromised benefit of the ALM+CpG vaccine against sand fly challenge (Figure S1). Note that despite the comparable parasitic loads in naïve mice following sand fly or needle challenge, the pathology associated with transmission by bite was far more severe.
The respective doses of the fly versus needle inocula did not appear to be a factor in the different outcomes of infection in the ALM+CpG vaccine as naive mice infected via needle or sand fly bite contained similar numbers of parasites in the challenge sites at 4 wks post-infection. In order to address the issue of dose more directly, and to determine if sand fly-derived parasites might be more virulent than those obtained from culture, ALM+CpG vaccinated and healed mice were challenged by infected sand fly bite or by inoculation with a five-fold higher dose of metacyclic promastigotes purified from the midguts of sand flies harboring 14 d, mature infections. Based on previous observations [36], 5×103 sand fly derived parasites are within the projected upper range of the variable doses transmitted following exposure to 4 infected sand flies. Mice with healed primary lesions were again powerfully protected against both needle and sand fly challenge (Figure 2B), and the ALM+CpG vaccinated mice maintained their immunity against the higher dose, sand fly-derived, needle inoculum, demonstrating a 100-fold decrease in parasite load. Importantly, these mice again failed to demonstrate any protection against sand fly transmitted infection as measured by either parasite load (Figure 2B), or a significant reduction in lesion scores (Figure S2). These results strongly suggest that transmission of L. major by sand fly bite, rather than an inherent difference in the dose or virulence of sand fly derived parasites, is responsible for the inability of ALM+CpG vaccinated mice to protect against natural challenge.
Kinetic analysis of the immune response among groups of mice challenged by the bite of infected sand flies in Figure 2A revealed that healed mice mounted a rapid and robust L.m.-specific response, while ALM+CpG vaccinated mice mounted a much weaker response, as determined by intracellular staining of dermal- derived CD3+ T cells for IFN-γ (Figure 2C) or TNF-α (Figure 2D). These different effector cell frequencies were reflected in the levels of IFN-γ secreted by ear-derived cells, as detected by ELISA (Figure 2E). Previous observations suggest that CD4+ T cells capable of producing multiple cytokines in response to antigen stimulation are more effective at protecting against disease [13]. In agreement with these studies, we found that a large proportion of L.m.-specific T cells in the healed mice produced IFN-γ and TNF-α simultaneously at day 7 (Figure 2F). These results emphasize the correlation between an early response and parasite clearance following sand fly transmission, and explain why ALM+CpG vaccinated mice were unable to control sand fly transmitted infection as compared to healed mice.
We were also interested to understand why the delayed appearance of the Th1 effector response in ALM+CpG vaccinated mice was sufficient to protect against needle challenge but not sand fly challenge. At 4 wks post-infection, despite enhanced numbers of lymphocytes (Figure S3) and increased levels of parasite antigen (Figure 2A) at the site of infected sand fly bite versus needle inoculation in ALM+CpG vaccinated mice, we observed a decrease in the frequency of both IFN-γ+ (6.8% versus 11%) and TNF-α+ (3.7% versus 9.3%) L.m.-specific T cells (Figure 2, C and D), as well as reduced levels of secreted IFN-γ (Figure 2E). Thus, conditions in the bite site appear to compromise the activation and/or effector function of the memory response generated by the killed vaccine.
We have recently demonstrated that the early host response to sand fly bites is associated with a unique and prolonged recruitment of neutrophils into the localized bite site, resulting in the formation of a “neutrophil plug”, and that the presence of neutrophils during the initiation of infection promotes the establishment of sand fly transmitted disease [34]. In order to explore the possibility that the host inflammatory response to sand fly bite is responsible for the failure of ALM+CpG vaccination to protect against sand fly transmitted infection, we first compared the inflammation induced by sand fly versus needle inoculation of L. major. When ear dermal cells from naive (Figure 3A) and ALM+CpG vaccinated mice (Figure 3B) were analyzed for the presence of neutrophils over the first week of infection, both needle and sand fly inoculated ears revealed a significant recruitment at 24 hours, although sand fly bitten ears had significantly greater numbers (p = 0.001). Importantly, only sand fly bitten ears maintained the neutrophilic infiltrate at the inoculation site at 3 and 8 days post-inoculation (Figure 3, A and B). Of note, a very transient neutrophilic response was also observed in the ears of the sham-transmitted mice, elicited by manipulation of the ear dermis during exposure to the transmission apparatus. Examination of cells derived from the ears of the needle or sand fly challenged mice shown in Figure 2A, revealed that only sand fly inoculation maintained recruitment of large numbers of neutrophils at the inoculation site at 1 and even 4 wks post-infection (Figure 3, C and D), which at least in the case of the naïve mice, was not explained by differences in the parasitic load. Analysis of all CD11b and Ly-6G/C (Gr-1) expressing cells reveals that increased numbers of neutrophils were also associated with large numbers of CD11b+Gr-1int macrophages/monocytes (Figure 3E).
In order to visualize neutrophil recruitment and maintenance at individual sites of L.m. inoculation over time, we employed 2-photon intra-vital microscopy (2P-IVM) in conjunction with a red fluorescent protein-expressing strain of L.m (L.m.-RFP) [36], and naïve mice expressing enhanced green fluorescent protein (eGFP) under the control of the endogenous lysozyme M promotor (LYS-eGFP mice) [37]. As previously reported, the GFPhi cells in these mice are neutrophils [34],[37], and accumulate within both needle and sand fly inoculation sites shortly after infection (Figure 3F, 2 hours; and Video S1 and S2). The sand fly inoculation site is distinguished by an especially tight co-localization of RFP+ parasites and GFPhi neutrophils, which form a plug delineating the site of proboscis penetration. (Figure 3F, 2 hours; Video S1 and S3). While neutrophils were maintained at the site of parasite deposition by sand fly bite (Figure 3F, Video S4) this co-localization was rapidly lost at the site of needle inoculation, and the majority of neutrophils present in the field of view at later times were within blood vessels (Figure 3F and Video S5).
We explored the possibility that neutrophil depletion might rescue the ability of the killed vaccine to confer protection against sand fly transmitted infection. As neutrophils are important for the early establishment of sand fly transmitted infections, their depletion at the time of challenge would, as previously shown [34], promote early resistance and compromise infection even in the naïve mice. Thus, the mice were left untreated for the first 3.5 days following sand fly transmission, then treated on days 3.5, 9, and 14, with a neutrophil depleting Ab [38],[39] or control IgG to mimic the loss of neutrophils observed following needle inoculation, but not sand fly transmission. Analysis of CD11b+Ly-6G+F4/80− neutrophils and CD11b+Ly-6G−F4/80+ macrophages/monocytes at the site of infection 6 days post-transmission revealed that the neutrophil depletion was both specific and efficient (Figure 4, A and B). At 2 weeks post-transmission, the neutrophil depletion promoted stronger Ag-specific IFN-γ and TNF-α responses in the ALM+CpG vaccinated mice (Figure 4C). More importantly, the neutrophil depletion enhanced the efficacy of the killed vaccine. Analysis of extensive data pooled from three independent experiments revealed that on day 28 post-transmission, the neutrophil depleted, ALM+CpG-vaccinated mice showed a highly significant reduction in parasite load compared with neutrophil depleted, naïve mice (p<0.0001), as well as control treated, ALM+CpG vaccinated mice (p = 0.002), and indistinguishable from that in healed animals (Figure 4D). The enhanced parasite clearance in neutrophil depleted, ALM+CpG vaccinated mice was associated with a significant reduction in lesion size compared with neutrophil depleted, naïve mice (p<0.0001) and control treated, ALM+CpG vaccinated mice (p = 0.01) (Figure 4E). Importantly, the neutrophil-depleted, naïve controls did not exhibit lower parasite loads compared with their control treated counterparts, suggesting the effect of neutrophil depletion after the initial establishment of infection, and during the extended period of neutrophil recruitment following transmission by bite, was specific to the vaccine setting.
The generation of a safe, non-living, prophylactic vaccine against leishmaniasis has been largely unsuccessful, a failure that is not explained by the lack of available target antigens with the potential to confer a protective response [40]. Failed human trials reported in the 1990s employing ALM+BCG were particularly perplexing as the same or a similar vaccine has been shown to work well as immunotherapy to hasten cure in patients with active disease [41],[42]. Furthermore, it elicits detectable parasite-specific IFN-γ production and leishmanin skin-test conversion in at least a proportion of recipients [20],[21],[22],[23],[24],[25],[26],[27],[28], and similar vaccine formulations, including the ALM+CpG vaccine employed in this study, have been shown to be highly effective against needle challenge in mouse models [11],[13]. The results reported here suggest that the killed vaccines failed in people because, while generating some correlates of immunity that may provide adequate defense against a needle inoculum, failed to generate and/or maintain the rapid, robust response at the site of secondary challenge induced by leishmanization that is required to prevent disease following delivery of parasites by sand fly bite. The protective response in healed mice is likely associated with the speed with which effector cells appear at a site of tissue damage, irrespective of the presence of parasites (see Figure 1). Following encounter with antigen in the inoculation site, these cells might then provide an immediate burst of effector cytokines, and counteract early on the down-modulatory environment created by the highly localized, neutrophil-dominated, response to sand fly bite. In contrast, and despite the ability of the killed vaccine plus CpG to generate multi-functional, effector T cells protective against needle challenge [13] (see also Figure 4), these cells are not present in adequate numbers and at sufficiently early time points to protect against sand fly transmission. This point is emphasized by the presence of similar numbers of neutrophils in both ALM+CpG vaccinated and healed mice one week following exposure to infected sand flies (Figure 3E), yet only healed mice were protected. Both the rapidity of the effector response in healed mice, as well as the fact that these cells were recruited by uninfected sand fly bites, suggests that these cells are not derived from a “central” memory population, that would require antigen encounter and several rounds of division in the DLN before gaining effector function [43]. More likely, the rapid appearance of these cells in the challenge site reflects a pre-existent, tissue-seeking effector population, undergoing constant renewal by the persistence of viable organisms in the healed mice [3],[4]. Further understanding of the effector population maintained by persistent infection, including the role of CD8+ cells, is likely to be highly informative to strategies of successful vaccination [44].
A critical question remains the mechanism by which neutrophil persistence following sand fly transmission inhibits parasite elimination in ALM+CpG vaccinated mice. Phagocyte clearance of apoptotic neutrophils during the resolution of inflammation has a known inhibitory effect on macrophage functions [45], and DC functions are similarly impaired following uptake of apoptotic neutrophils [46]. Thus, infected macrophages and DC persistently exposed to apoptotic neutrophils at the site of sand fly bite are likely to be refractory to activation signals, inhibiting both the killing and APC functions of these cells. This is especially relevant to sand fly transmission where the association between neutrophils and macrophage/monocytes, as well as dendritic cells, is highly localized at the sand fly bite site. Apoptosis of infected neutrophils has been readily captured by 2P-IVM [34]. The maintenance of neutrophils at sand fly bite sites is likely the result of conditions leading to their protracted recruitment, as opposed to prolongation of their life span in the skin [47]. Thus, while the initial recruitment of neutrophils may be driven primarily by tissue injury, their continued presence is likely influenced by PSG or salivary components, that are themselves chemotactic or that initiate the inflammatory cascade [33].
The results reported here represent the first determination, so far as we are aware, of the factors influencing the efficacy of protective immunity generated by different vaccine formulations against sand fly challenge, and may be relevant to the conditions that modulate vaccine induced immunity to other vector borne pathogens. Beyond emphasizing the somewhat obvious importance of using natural challenge models to evaluate experimental vaccines against leishmaniasis, the results provide a more stringent set of screening criteria that might be used to predict vaccine success against fly challenge, relating to the rapid appearance of multifunctional effector cells within the challenge site.
Pertinent to our findings are those of Rogers et. al. who demonstrated that vaccination of BALB/c with glycoconjugates derived from PSG diminishes disease severity following sand fly challenge [29]. Collectively, these findings should be especially informative for ongoing and future clinical development of “second-generation” Leishmania vaccines [48], and reinforce the rationale for inclusion of molecules specific to natural transmission, such as selected components of sand fly saliva or promastigote-secretory gel, in an anti-Leishmania vaccine [49].
Female C57BL/6 mice were obtained from Jackson Laboratories. C57BL/6 LYS-eGFP knock-in mice [37] were a gift from T. Graf (Albert Einstein University, NY) and were bred at Taconic Laboratories through a contract with the NIAID. Mice were maintained at a NIAID animal care facility under specific pathogen-free conditions. All animal experiments were performed under a study protocol approved by the NIAID Animal Care and Use Committee.
All experiments were carried out using the L. major Friedlin strain obtained from the Jordan Valley NIH/FV1 (MHOM/IL/80/Friedlin). In some experiments, a stable transfected line of FV1 L. major promastigotes expressing a red fluorescent protein was employed, as described previously [36]. Briefly, the DsRed gene was amplified by PCR employing the pCMV-DsRed-Express plasmid (BD Biosciences/Clontech) as a template and cloned into the SpeI site of the pKSNEO Leishmania expression plasmid. FV1 promastigotes were transfected with the resulting expression plasmid construct [pKSNEO-DsRed] and selected for growth in the presence of 50 µg/ml Geneticin (G418) (Sigma).
L. major or L. major-RFP were grown at 26°C in medium 199 supplemented with 20% heat-inactivated FCS (Gemini Bio-Products), 100 U/ml penicillin, 100 µg/ml streptomycin, 2 mM L-glutamine, 40 mM Hepes, 0.1 mM adenine (in 50 mM Hepes), 5 mg/ml hemin (in 50% triethanolamine), and 1 mg/ml 6-biotin. Infective-stage metacyclic promastigotes were isolated from stationary cultures (4–6 day-old) by negative selection of non-infective forms using peanut agglutinin [50] (PNA, Vector Laboratories Inc). In some experiments metacyclic promastigotes of L. major were isolated from sand flies on day 14 following infection with L. major, as previously described [36]. Briefly, Infected flies were killed, dissected aseptically, and the stomodeal valve and anterior gut of each fly was transferred into Dulbecco's modified Eagle's medium (DMEM). The guts were macerated briefly using a plastic pestle, spun twice to remove the debris, and washed once in DMEM followed by metacyclic promastigote isolation as described above. Mice were subsequently infected with the specified number of parasites in the ear dermis by intra-dermal (i.d.) injection using a 29 ½ GA needle in a volume of 10 µl unless specified otherwise.
Analysis of protective immunity in mice with a healed primary lesion was carried out using animals that had been infected 16–20 weeks previously with 104 L. major metacyclic promastigotes in the left hind footpad by sub-cutaneous injection using a 29 ½ gauge needle in a volume of 40 µl. Autoclaved Leishmania antigen (ALM) plus CpG oligodeoxynucleotides (ODN) vaccination was performed in a manner similar to that published previously [11]. Briefly, B6 mice were injected subcutaneously in their left hind footpad with 50 mg of clinical grade ALM, prepared from whole cell heat-killed L. major promastigotes (WHO) plus 50 µg of CpG ODN sequence 1826 (Coley Pharmaceutical Group), graciously provided by Dr. P. Darrah (VRC/NIH), using a 29 ½ gauge needle in a volume of 40 µl, three times, at 2 week intervals.
Transmission of L. major parasites was performed as described [34],[36]. Briefly, 2–4 day old P. duboscqi (Mali colony) female sand flies were infected via feeding through a chick skin membrane on heparinized mouse blood containing L. major or L. major-RFP amastigotes or promastigotes. After 14–15 days, individual flies were transferred to plastic vials covered at one end with nylon mesh. Mice were anesthetized by intraperitoneal injection of 30 µl of ketamine/rompin (100 mg/ml). Specially designed clamps were used to bring the mesh end of each vial into contact with the ear of an anesthetized mouse, allowing flies inside the vial to feed on the ear skin for a period of 2 to 3 hours in the dark. In some experiments mice were exposed to empty vials. The number of flies with blood meals was employed as a means of checking for equivalent exposure to potential transmission by sand fly bite among animals in different treatment groups. The median number of flies with blood meals in vials with 4 flies was 2.
Ear tissue was prepared as previously described [34]. Briefly, the ventral and dorsal sheets of needle or sand fly inoculated ears were separated, deposited in DMEM containing 100 U/ml penicillin, 100 µg/ml streptomycin and 0.2 mg/ml Liberase CI purified enzyme blend (Roche Diagnostic Corp.), and incubated for 2 hours at 37°C and 5% CO2. Digested ear sheets were subsequently homogenized for 3 minutes using the Medicon/Medimachine tissue homogenizer system (Beckton Dickinson). Individual retromaxillary (ear) lymph nodes were removed, and mechanically dissociated using tweezers and a syringe plunger. Single cell suspensions of tissue homogenates were then filtered using a 70 µm-pore size Falcon cell strainer (BD Biosciences).
Mice were sacrificed and single cell suspensions from the ear dermis were obtained as described above. Cells were incubated without fixation with an anti-Fc-γ III/II (CD16/32) receptor Ab (2.4G2, BD Biosciences) in RPMI without phenol red (Gibco) containing 1.0% FCS for 10” followed by incubation for 20” with a combination of 4 or 6 of the following anti-mouse antibodies: PE-Cy7 or APC anti-CD11b (M1/70 BD Biosciences); Per-CP Cy5.5 anti-Gr-1(Ly6G/C) (RB6-8C5, BD Biosciences); FITC or PE anti-Ly6G (1A8, BD Biosciences); PE anti-CD11c (HL3, BD Biosciences); Per-CP Cy5.5 anti-CD11c (N418, BioLegend); APC anti-F4/80 (BM8, eBioscience), FITC anti-I-Ab (AF6-120.1, BD Biosciences); or Alexafluor-700 anti-mouse MHC II (M5/114.15.2, eBioscience). The isotype controls employed were rat IgG1 (R3-34) and rat IgG2b (A95-1). The data were collected and analyzed using CELLQuest software and a FACScalibur or FacsDIVA software and a FacsCANTO flow cytometer (BD Biosciences). Gating of ear-derived cells was carried out as described previously [34]. Ears were analyzed individually, or pooled with ears from the same group, as indicated in the text.
Whole ear single-cell suspensions in RPMI 1640 containing 10% FCS, 10 mM Hepes, L-glutamine, and penicillin/streptomycin, obtained as described above, were incubated at 37°C in 5% CO2 for 16–18 hours in flat-bottom 48-well plates with 2.5×105 BMDCs, with or without 50 mg/ml freeze-thaw Leishmania antigen prepared from L. major V1 stationary phase promastigotes, in a final volume of 1 ml. During the last 5–6 hours of culture Brefeldin A (Golgiplug; BD Biosciences) was added to block golgi transport according the manufacturers' instructions. Following in vitro culture, cells were washed and stained with anti-Fc III/II (CD16/32) receptor Ab (2.4G2) for 10 minutes in RPMI without phenol red containing 1.0% FCS, followed by PE-Cy7 or PE-Cy5 anti-mouse CD4 (RM4-5) for 15 minutes. In some experiments cells were also stained with FITC anti-TcR β (145-2 C11). Cells were then fixed with BD Cytofix/Cytoperm (BD Biosciences) and stained with anti-Fc III/II (CD16/32) receptor Ab (2.4G2) followed by a combination of the following anti-mouse antibodies: PerCP-Cy5.5 anti-CD3 (145-2C11), FITC-, APC-, or AlexaFluor 700 anti-IFN-g (XMG1.2), and FITC or PE anti-TNF-α (MP6-XT22). Intracellular staining was carried out for 30 minutes on ice. All antibodies were acquired from BD Biosciences. For each sample, greater then or equal to 4000 CD4+CD3+ cells were collected using a FACS Caliber or FACS Canto flow cytometer and analyzed using either Cell Quest Pro or FACS Diva Software, respectively (BD Biosciences). For measurement of IFN-γ in culture supernatants, pooled, single-cell suspensions of ear tissue as described above were incubated in triplicate at 37°C in 5% CO2 for 72 hours in 96-well round bottom plates with 2.5×105/ml BMDC with or without freeze-thaw Leishmania antigen in a total volume of 200 ml. Following incubation, the concentration of IFN-γ in the culture supernatant was determined by ELISA according the manufactures instructions (eBioscience).
Parasite titrations were performed as previously described [31]. Briefly, tissue homogenates were serially diluted in 96-well flat-bottom microtiter plates containing biphasic medium, prepared using 50 µl NNN medium containing 20% of defibrinated rabbit blood and overlaid with 100 µl M199/S. The number of viable parasites in each ear was determined from the highest dilution at which promastigotes could be grown out after 7–10 days of incubation at 26°C.
Because individual sand flies, or more then one sand fly may deposit parasites in more than one location, sand fly bitten ears often have more then one lesion. Total lesion diameter was determined by measuring the diameter of individual lesions using a caliper and in cases where there was more then one lesion per ear the diameters were added together.
Two photon intravital imaging and image analysis was performed as described previously [34]. Briefly, anesthetized mice were imaged in the lateral recumbent position that allowed the ventral side of the ear pinna to rest on a coverslip. A strip of Durapore tape (3 M) was stuck to a bench top several times (to ensure that subsequent removal would not cause undue damage) and placed lightly over the ear pinna and affixed to the imaging platform in order to immobilize the tissue. Care was taken to minimize pressure on the ear.
Images were acquired using an inverted LSM 510 NLO multiphoton microscope (Carl Zeiss Microimaging) enclosed in an environmental chamber that was maintained at 30°C. This system had been custom fitted with 3 external non-descanned PMT detectors in the reflected light path. Images were acquired using either a 20×/0.8 air objective or a 25×/0.8 NA water immersion objective. Fluorescence excitation was provided by a Chamelon XR Ti:Sapphire laser (Coherent) tuned to 920 nm for eGFP excitation. Voxel dimensions were 0.64×0.64×2 µm using the 20× objective and 0.36–0.51×0.36–0.51×2 µm using the 25× objective.
Raw imaging data were processed with Imaris (Biplane) using a Gaussian filter for noise reduction. All images are displayed as 2D maximum intensity projections. Movie files of 3-dimentional images were generated using Imaris.
Animals were treated with three 0.5 mg injections of a neutrophil depleting (NIMP-R14) or control (GL113) IgG antibody, i.p., on days 3.5, 9, and 14 following sand fly transmission. The first dose of antibody was delayed until 3.5 days after exposure to infected sand fly bite as earlier observations demonstrated that L.m. infection is established in macrophages at this time [34]. Antibody treatments were spaced 5 days apart as preliminary experiments suggested excessive administration of the NIMP-R14 antibody, such as on successive days, led to depletion of cell types other then neutrophils. Success and specificity of depletions were determined as described in the text. The NIMP-R14 hybridoma was a gift from Dr. Y. Belkaid (NIAID).
Parasite loads in the ears of mice transmitted with L. major by infected sand fly bite do not follow a Gaussian distribution. This is likely the result of variability in the infectious burden and feeding behavior of individual, infected, sand flies [36]. Therefore, data sets were compared using a nonparametric Mann Whitney test. Mann Whitney calculations were done using Prism 4 (Graphpad Software, Inc. San Diego, CA). In Figure 4, D and E, parasite loads and lesion size were compared using an exact stratified Wilcoxon rank sum test, stratified by experiment in order to allow pooling of experiments as described in the text. The stratified Wilcoxon calculations were done in StatXact 8 Procs (Cytel, Inc., Cambridge, MA). Comparisons in which the data represented replicate samples were carried out using t-tests. All p-values are two-sided.
|
10.1371/journal.ppat.1003946 | Identification of Host-Targeted Small Molecules That Restrict Intracellular Mycobacterium tuberculosis Growth | Mycobacterium tuberculosis remains a significant threat to global health. Macrophages are the host cell for M. tuberculosis infection, and although bacteria are able to replicate intracellularly under certain conditions, it is also clear that macrophages are capable of killing M. tuberculosis if appropriately activated. The outcome of infection is determined at least in part by the host-pathogen interaction within the macrophage; however, we lack a complete understanding of which host pathways are critical for bacterial survival and replication. To add to our understanding of the molecular processes involved in intracellular infection, we performed a chemical screen using a high-content microscopic assay to identify small molecules that restrict mycobacterial growth in macrophages by targeting host functions and pathways. The identified host-targeted inhibitors restrict bacterial growth exclusively in the context of macrophage infection and predominantly fall into five categories: G-protein coupled receptor modulators, ion channel inhibitors, membrane transport proteins, anti-inflammatories, and kinase modulators. We found that fluoxetine, a selective serotonin reuptake inhibitor, enhances secretion of pro-inflammatory cytokine TNF-α and induces autophagy in infected macrophages, and gefitinib, an inhibitor of the Epidermal Growth Factor Receptor (EGFR), also activates autophagy and restricts growth. We demonstrate that during infection signaling through EGFR activates a p38 MAPK signaling pathway that prevents macrophages from effectively responding to infection. Inhibition of this pathway using gefitinib during in vivo infection reduces growth of M. tuberculosis in the lungs of infected mice. Our results support the concept that screening for inhibitors using intracellular models results in the identification of tool compounds for probing pathways during in vivo infection and may also result in the identification of new anti-tuberculosis agents that work by modulating host pathways. Given the existing experience with some of our identified compounds for other therapeutic indications, further clinically-directed study of these compounds is merited.
| Infection with the bacterial pathogen Mycobacterium tuberculosis causes the disease tuberculosis (TB) that imposes significant worldwide morbidity and mortality. Approximately 2 billion people are infected with M. tuberculosis, and almost 1.5 million people die annually from TB. With increasing drug resistance and few novel drug candidates, our inability to effectively treat all infected individuals necessitates a deeper understanding of the host-pathogen interface to facilitate new approaches to treatment. In addition, the current anti-tuberculosis regimen requires months of strict compliance to clear infection; targeting host immune function could play a strategic role in reducing the duration and complexity of treatment while effectively treating drug-resistant strains. Here we use a microscopy-based screen to identify molecules that target host pathways and inhibit the growth of M. tuberculosis in macrophages. We identified several host pathways not previously implicated in tuberculosis. The identified inhibitors prevent growth either by blocking host pathways exploited by M. tuberculosis for virulence, or by activating immune responses that target intracellular bacteria. Fluoxetine, used clinically for treating depression, induces autophagy and enhances production of TNF-α. Similarly, gefitinib, used clinically for treating cancer, inhibits M. tuberculosis growth in macrophages. Importantly, gefitinib treatment reduces bacterial replication in the lungs of M. tuberculosis-infected mice.
| Tuberculosis continues to be a cause of significant morbidity and mortality world-wide due to numerous factors, including the rise of drug resistance and the absence of an effective vaccine. The challenge of adherence to long treatment regimens and the limited number of effective therapeutics drive the need for innovative therapeutic strategies that are predicated on a better understanding of the biology of infection. For example, the fact that the majority of people infected with the bacillus never develop active disease suggests that human immunity is actually quite effective at controlling M. tuberculosis; a novel, potentially more effective therapeutic strategy could emerge if we were able to understand and leverage the basis for this control.
The cellular interaction between M. tuberculosis and macrophages is crucial for determining the outcome of infection. Early in infection, macrophage microbicidal mechanisms actively work to try to clear the bacteria; however, macrophage responses that are adequate to kill other bacterial pathogens often fail to clear M. tuberculosis. In a majority of individuals, the activation of macrophages by IFN-γ can result in control but not sterilization of infection, instead driving it into a latent state. During latency, the macrophage can be a niche where the bacteria are protected from assaults by the immune system and antibiotic therapy, thus facilitating their persistence and ultimate dissemination.
Recent studies have uncovered a number of processes that are important to tubercular infection. The ability of M. tuberculosis to arrest the normal progress of phagosome maturation is critical for its survival in macrophages [1]; however, the molecular mechanisms on both the pathogen and host sides that account for this arrest are unclear. For example, while calcium signaling in macrophages appears to be important in this process, the nature of the calcium signal and the mechanisms by which M. tuberculosis actively affects calcium signaling are debated [2], [3]. In addition to phagosome maturation arrest, M. tuberculosis may actively suppress many other macrophage innate immune responses. For example, virulent strains of M. tuberculosis actively prevent apoptosis of infected macrophages, thus preventing bacterial killing by macrophage efferocytosis and avoiding activation of T-cells through cross-presentation of antigens by dendritic cells [4], [5], [6]. M. tuberculosis may also actively prevent activation of the inflammasome and induction of autophagy [7], [8]. In addition to subversion of immune responses, M. tuberculosis manipulates the host microenvironment in order to acquire nutrients to promote its own survival. For example, virulent mycobacteria are able to induce the development of intracellular lipid bodies which fuse with M. tuberculosis containing phagosomes and provide a critical source of carbon [9].
Although we have some insight into the pathways that are important for M. tuberculosis infection of macrophages, our current understanding of the mechanisms that determine whether the macrophage controls bacterial infection or succumbs to its virulence is incomplete. In order to obtain greater insight into host factors involved in M. tuberculosis infection, unbiased screening using RNAi or small molecules targeting host proteins have recently been performed. Two published RNAi screens, one genome-wide and one focused on kinases and phosphatases, identified mammalian proteins that are candidate regulators of M. tuberculosis infection [10], [11]. To provide a functional context for the identified regulators, the authors constructed a signaling network by integrating the RNAi screening data with data from transcriptional profiling. Over half of identified genes were found to be negative regulators of autophagy, affirming the importance of this pathway for host defense against M. tuberculosis [10]. In addition to regulators of autophagy, the networks implicated were enriched for modules that govern metabolism and signal transduction, with many of these modules centered around the serine/threonine kinase AKT.
Kinases are central to mammalian signaling pathways. AKT/PKB is a key modulator of cellular processes such as growth and proliferation, glucose metabolism, apoptosis, and autophagy. AKT is specifically activated during Salmonella infection of host cells by the bacterial effector SopB and promotes bacterial survival by prevention of phago-lysosome fusion [12]. Treatment of M. tuberculosis macrophages with the AKT and PKA inhibitor H-89 also results in inhibition of bacterial growth. However, in contrast to Salmonella infection, the role of AKT is unknown in M. tuberculosis infection [12]. Importantly, although AKT was identified in the network that emerged from the genome-wide RNAi screen of M. tuberculosis infected THP-1 macrophages, the kinase itself was neither identified in the primary genome-wide screen nor in a more directed kinase/phosphatase screen conducted by the same group [11].
Of note, in the RNAi screens that have been reported, the siRNAs used to decrease host factor expression were added only after M. tuberculosis had already entered and adapted to the macrophage microenvironment; thus, these screens were not designed to identify factors that are crucial for the earliest events in the host-pathogen interaction. Effective silencing of gene expression using transfection of siRNA is in part dependent on the half-life of the targeted protein and occurs on the timescale of hours to days after transfection. In contrast, a chemical biological approach has some advantages over RNAi with regard to studying early events. The rapid binding of small molecules to proteins facilitates probing the early period immediately after phagocytosis. Because this period is accompanied by the most significant transcriptional responses on the part of both the bacterium and the macrophage [13], events during this time frame are likely important determinants of the outcome of infection. Further, in general, unlike genetic approaches which can target only one isoform or homologue at a time, small molecules can inhibit multiple, closely related isoforms of the same target, thus facilitating the identification of host activities performed by two or three closely related targets with redundant function. In addition, unlike RNAi, small molecules can inhibit enzymatic function without disrupting larger complexes with subsequent pleiotropic effects, or can inhibit a specific function of a protein while leaving other functions intact [14]. Thus small molecule based screens can provide a valuable complement to existing datasets obtained using siRNA based knockdown.
Three recently conducted screens were designed to identify small molecules that disrupt M. tuberculosis replication in macrophages. The first screen used a high-content imaging approach to identify compounds that directly target bacterial processes during macrophage growth [15]. A subsequent study by the same group and using the same screening approach identified an inhibitor of M. tuberculosis cytochrome bc1 [16]. More recently, a group reported a screen in which microscopy was used to identify host-targeted inhibitors that prevent replication of Mycobacterium bovis BCG in macrophages. This study characterized existing neurotropic drugs that diminish replication of M. tuberculosis in infected macrophages by inducing autophagy or altering endosomal trafficking, however the targets of these drugs and their modes of action have yet to be elucidated [17]. We performed a complementary screen of a library of known bioactive small molecules to identify inhibitors of M. tuberculosis replication in macrophages that are biased towards disrupting host functions. Our goal was to identify small molecules that specifically target host proteins to gain insight into host functions required for controlling M. tuberculosis infection. To focus on early events crucial for the adaptation of the macrophage for bacterial survival but not on uptake itself, we added molecules to infected macrophages immediately after phagocytosis. From this screen, we identified several classes of host-targeted compounds that limit the ability of M. tuberculosis to proliferate in macrophages, including kinase inhibitors, G-protein coupled receptor modulators, and ion channel inhibitors. Our results are complementary to previously reported screens, expanding our knowledge of host pathways that are crucial determinants of infection. Based on the bioactive molecules identified in our screen, we determined that the SSRI fluoxetine inhibits M. tuberculosis growth in macrophages, induces enhanced expression of TNF-α, and enhances autophagy. We additionally demonstrated a new role for the tyrosine kinase EGFR during M. tuberculosis infection. Importantly, we showed that the clinically used EGFR inhibitor gefitinib has efficacy for preventing bacterial replication in both infected macrophages and mice, suggesting that EGFR is relevant in vivo. The inhibitors we have identified will be important tools for studying the host-pathogen interaction and for testing the importance of host pathways during in vivo infection in animal models.
For identification of small molecules that inhibit M. tuberculosis replication in macrophages, after significant assay testing (see Methods S1) we designed a microscopy-based assay with simultaneous imaging of macrophages and mycobacteria that provides the sensitivity to detect modest growth inhibition, easily identifies compounds with significant macrophage toxicity, and can be largely automated to allow for the rapid testing of thousands of compounds (for assay design and development details, see Methods S1). We used this high-content imaging assay to monitor the growth of GFP-expressing M. tuberculosis in infected macrophages.
In designing the assay, we sought to establish growth conditions that would allow for robust and reproducible replication of M. tuberculosis in macrophages and a readout that would accurately reflect bacterial number. Macrophages were infected with M. tuberculosis constitutively expressing GFP (H37Rv-GFP [18]), and at various time-points after infection, the cells were fixed and stained with DAPI to allow for enumeration of macrophages by microscopy. Because intracellular mycobacteria grow in clumps rather than as discrete, easily quantifiable spots, we utilized the open-source image analysis software CellProfiler, which can simultaneously determine and integrate multiple parameters from our images [19] to determine an optimal visual parameter that correlates with bacterial number. We found that GFP pixel intensity integrated across the field and normalized to macrophage number best represented growth while also accounting for differences in surviving macrophage number, and most accurately reflected bacterial census enumerated by plating for colony forming units (CFU).
Using this parameter, we tested growth of H37Rv-GFP in a number of cell lines including THP-1 cells, RAW 264.7, and J774A.1 cells, and found that J774 murine macrophages allowed for the most consistent and robust growth, with a bacterial doubling time equivalent to that measured in axenic culture. The optimal time-point for measuring growth was 3 days post infection, when intracellular mycobacterial growth was most homogeneous and little apparent macrophage death was observed.
We next assessed the heterogeneity of mycobacterial growth and ability of the assay to reproducibly detect growth inhibition across large numbers of wells in 96-well format. Using a gradient of concentrations of rifampicin to model varying degrees of growth inhibition, we demonstrated that the assay was able to distinguish growth inhibition of 50% inhibition or less (Figure 1A, Figure S1). With the highest dose of rifampicin we consistently obtained Z′-factors of 0.4–0.5, which is borderline in robustness for a high-throughput screen. However, because bacterial growth based on imaging across the wells did not follow a Gaussian distribution, standard statistical methods to assess high-throughput assay quality, such as Z′-factor, that assume a normal distribution for high-throughput assays may not accurately assess the robustness of the assay. We therefore used a bootstrap Monte Carlo analysis as another means to assess the quality of the assay [20], [21] and found that we were able to consistently distinguish 50% inhibition of growth as established with our positive controls with a predicted false negative rate of 0.2% and a false positive rate of 0.13%. We found these rates to be acceptable, allowing us to progress to high-throughput screening. Ultimately we compared the list of hits obtained using either composite z-scores or p-values obtained using the Monte Carlo bootstrap analysis (both metrics included in Table S1); we found that the list of hits was essentially the same using both metrics.
We used the assay to screen a library of 1920 small molecules from the Broad Institute bioactives collection and an additional 159 kinase inhibitors obtained from Nathanael Gray at the Dana Farber Cancer Institute (Table S1, Figure S2). One advantage of screening a library of known bioactives is that compound annotation can provide initial suggestions about mechanisms of action, thus facilitating the considerable challenge of target identification from potential host and bacterial targets. Additionally, the bioactives library is enriched for molecules that target mammalian proteins, thus allowing for elucidation of novel host/pathogen biology by focusing on macrophage functions.
In a 96 well format, ∼3000 macrophages were plated in each well and infected with M. tuberculosis strain H37Rv-GFP at a multiplicity of infection (MOI) of 1∶1 for 4 hours, after which time extracellular bacteria were removed by washing. To facilitate identification of compounds that restrict intracellular M. tuberculosis growth rather than inhibit uptake, the compound library was added after phagocytosis. The average final concentration of small molecules was 5 µM. After three days, cells were washed, fixed, stained with DAPI, and imaged with a 4× objective lens on an ImageXpress Micro high-throughput microscope (Molecular Devices).
We analyzed our data using both a bootstrap Monte Carlo analysis and composite z-score and were reassured that both methods identified the same small molecule hits. (The z-score cutoff was <−1.5; the p-value cutoff was <0.025.) From the primary screen we identified a total of 164 unique small molecules that resulted in statistically significant growth inhibition when compared to control wells. We discarded compounds with significant macrophage toxicity (<50% macrophage survival relative to controls), known antibiotics, and compounds identified to have significant activity against M. tuberculosis grown in axenic culture in a parallel screen [18]. After applying these filters, we were left with 133 unique non-antibiotic compounds (Table S1) that restrict growth of M. tuberculosis in macrophages. Based on known annotation, most identified molecules fall into 5 broad categories: G-protein coupled receptor (GPCR) modulators, ion channel modulators, membrane transport protein-acting compounds, anti-inflammatories, and kinase modulators (Figure 1B). Categorizing compounds based on known annotations raises the caveat that activity in this assay may not necessarily be related to their annotated activity but rather to another “off-target” effect. However, we compared the relative representation of activity classes from active molecules with the library that was screened and found a slight over-representation of GPCR modulators, ion channel modulators, and membrane transport protein-acting compounds, and similar representation among anti-inflammatory compounds and kinase modulators. The specific categories identified and relative proportions of hits in each category reflect both underlying biology and representation of these categories in the screened library. To assess the success of the screen, we selected and repurchased a group of 22 compounds with a range of composite z-scores (−4 to −1.5) for retesting. Importantly, our retest rate from these compounds was 90%.
As the imaging assay is only a proxy for mycobacterial growth, selected compounds representing each of the major five classes were retested using the gold-standard for mycobacterial growth, colony-forming units (CFU). Compounds repurchased from commercial sources were tested at multiple concentrations to demonstrate the dose-dependence of growth-inhibition. J774 murine macrophages were infected with M. tuberculosis strain H37Rv, treated with three concentrations of each compound, and allowed to grow for three days. Prior to harvesting cells, we verified microscopically that there was no significant macrophage toxicity. Cells were then washed with PBS, lysed, and plated for CFU. Relative to dimethylsulfoxide (DMSO) controls, fluoxetine (a selective serotonin reuptake inhibitor; membrane transport protein), farnesyl thiotriazole (FTT) (a protein kinase C activator; kinase modulator), AKTi1/2 (an inhibitor of PKA/PKB; kinase modulator), quinidine (a sodium channel inhibitor; ion channel inhibitor), and ritanserin (an antagonist at the serotonin 2A receptor; GPCR modulator) all demonstrated significant dose-dependent activity against intracellular mycobacteria. At the maximum concentrations used, inhibition ranged from 50% for FTT to 75% for fluoxetine (Figure 2A). While this is less than would be expected from a traditional antibiotic with direct bactericidal activity, it is consistent with the magnitude potentially expected when host regulators of infection are targeted [10], [11]. A subset of compounds were additionally retested for activity in 5 point dose-response using our imaging assay (Figures S3, S4, S5, S6, S7, S8); observed inhibition was similar to that seen using CFU.
Because immortalized cell lines may contain mutations that alter signaling pathways and gene expression, we next verified that these same compounds have similar activity in a primary macrophage model. Primary mouse bone marrow-derived macrophages (BMDM) were infected with H37Rv and treated with the same panel of bioactive compounds. Because mycobacteria grow more slowly in BMDM than in cell lines, infection was allowed to progress for 5 days. After light microscopy was used to confirm the absence of macrophage death, cells were washed and lysed, and then released bacteria were plated for CFU. All tested compounds had significant activity against M. tuberculosis in BMDM (Figure 2B).
To confirm that the bioactive compounds were acting primarily on the host cells and were not directly toxic to the bacilli at the concentrations used in the macrophage infection model, compounds were tested for growth inhibitory activity against M. tuberculosis growing in axenic culture. At concentrations higher than the highest concentration used against infected macrophages, none of the inhibitors had significant activity against M. tuberculosis growing in axenic culture at any point during a 14-day time course (Figure S9).
Host targeted inhibitors that restrict the ability of M. tuberculosis to proliferate in a macrophage could function in various ways, such as by activating or enhancing a functional immune response like autophagy, or by inhibiting a signaling pathway activated by M. tuberculosis as a virulence mechanism. As a recent genome-wide RNAi screen that sought to identify host factors that regulate the ability of M. tuberculosis to replicate in macrophages predominantly identified negative regulators of autophagy [10], we tested whether our small molecule inhibitors might function by activating autophagy. LC3 is a ubiquitin-like protein that localizes to autophagolysosomes and is used as a specific molecular marker of autophagy. Upon induction of autophagy, LC3-I is converted to LC3-II by lipidation by a ubiquitin-like system; this conversion can be used as a readout of autophagy. Using the LC3 conversion assay, we determined that relative to uninfected macrophages or infected macrophages treated with DMSO control, both the EGFR inhibitor gefitinib and the serotonin transport inhibitor fluoxetine significantly enhanced autophagy (Figure 3A). Although the degree of LC3 conversion varied from experiment to experiment, these inhibitors consistently increased the LC3-I to LC3-II ratio. Other small molecules, including AKTi1/2, imatinib, quinidine, and ritanserin, did not consistently or significantly enhance LC3 conversion. These results suggest that gefitinib and fluoxetine restrict intracellular mycobacterial growth at least in part by enhancing host autophagy pathways. As the two molecules are not structurally related and act within distinct signaling pathways, there may be multiple routes through which autophagy can be activated to restrict the growth of M. tuberculosis.
The capacity to mount an inflammatory cytokine response to infection is critical to the host capacity to control tuberculosis. Patients with latent tuberculosis treated with TNF-α inhibitors or soluble receptors are at significantly increased risk of developing reactivation disease [22], [23]. Similarly, patients with mutations in the IFN-γ receptor are at increased risk of severe mycobacterial disease [24]. On a cellular level, infection of macrophages with bacterial pathogens results in the induction of inflammatory responses that can ultimately lead to control of infection. M. tuberculosis infection of unactivated macrophages results in an induction of cytokines including TNF-α and the production of reactive oxygen and nitrogen species; however, in the absence of IFN-γ activation, the response is insufficient to control replication. Host targeted inhibitors could limit intracellular proliferation by upregulating TNF-α, a cytokine which has been shown to inhibit mycobacterial growth in macrophages through both nitric oxide dependent and independent mechanisms [25]. To determine whether treatment of infected macrophages with our inhibitors results in upregulation of TNF-α, macrophages were infected with M. tuberculosis and TNF-α levels were measured in the supernatants 24 h after infection. We found that treatment of infected macrophages with fluoxetine led to a significant increase in the amount of TNF-α produced (Figure 3B). This result is consistent with previous reports that serotonergic drugs including fluoxetine increase the in vivo levels of inflammatory cytokines, including TNF-α, both in human subjects and in animal models [26]. Thus, the effects that we observe in cell culture may be generalizable to a whole organism. Induction of autophagy has been described among the effects of TNF-α on macrophages [27]. As fluoxetine both induces TNF-α and autophagy, fluoxetine may induce autophagy by increasing levels of TNF-α. These results suggest that host-targeted therapies have the potential to inhibit M. tuberculosis proliferation by enhancing host protective inflammatory responses, at least at the level of the individual macrophage. Given the widespread clinical use and safety profile of serotonergic drugs such as the SSRIs, further investigation on whether their administration could impact tuberculosis treatment would be interesting.
Kinases are important signaling molecules that regulate many aspects of mammalian cell biology including functions important for cellular response to infection, such as innate immune pathways, apoptosis, and autophagy. Kinase signaling has previously been demonstrated to be important for supporting M. tuberculosis replication in macrophages [11]. We identified 15 compounds that are annotated as kinase inhibitors in our screen (Table S1). Several of the targets of these kinase inhibitors have been previously implicated in the pathogenesis of M. tuberculosis or related pathogenic mycobacterial species, validating the capacity of the screen to identify biologically meaningful results [10], [12], [28]. H-89, an ATP-competitive inhibitor of kinases including both AKT and PKA, has been shown to inhibit proliferation of M. tuberculosis in mouse bone marrow derived macrophages [10], [12]. Imatinib, an inhibitor of Ableson family kinases (ABL) was also recently demonstrated to decrease the ability of M. tuberculosis to replicate in both macrophages and mice [28], [29]. However, the specific role that these kinases play in supporting replication of M. tuberculosis in host cells is unclear.
Understanding the role that host kinase signaling has upon M. tuberculosis macrophage growth is important for understanding the basic biology of infection, but might also have significant implications for the development of host-targeted therapeutics. Several of the kinase inhibitors identified in our screen are compounds currently used clinically for non-infectious disease related indications, raising the possibility of repurposing of such drugs for the treatment of tuberculosis. To confirm the efficacy of kinase inhibitors, we focused on three kinases for which inhibitors are currently in clinical use or development for cancer treatment: AKT, ABL and Epidermal growth factor receptor (EGFR). For each of these kinases we identified and tested at least two structurally unrelated compounds annotated as specific inhibitors of the individual enzymes. Although in practice a target responsible for any given phenotype is not necessarily the annotated target, the correlation between structurally unrelated compounds annotated to have the same target increased our confidence in compound specificity. Treatment of M. tuberculosis-infected macrophages with inhibitors targeting AKT (AKTi1/2, H-89), Abl (imatinib, GNF2), or EGFR (gefitinib, lapatinib) lead to a >2 fold decrease in bacterial load at the highest tested concentration of each inhibitor in both J774 macrophages (Figure 4A) and mouse bone marrow macrophages (Figure 4B).
The serine/threonine kinase AKT/PKB has been previously implicated in the ability of M. tuberculosis and other bacterial pathogens to replicate in host cells [10], [12]. Treatment of infected cells with the AKT1 inhibitor H-89 was shown to limit proliferation of M. tuberculosis in infected mouse macrophages [12] and knocking down AKT1 and AKT2 in human THP-1 cells also led to decreased replication of intracellular M. tuberculosis [10]. In our screen, in addition to previously identified ATP competitive inhibitor H-89, which preferentially targets PKA in addition to AKT, we found that a more specific allosteric inhibitor of AKT, AKTi1/2 [30], [31] also restricts intracellular M. tuberculosis replication. To confirm that AKT plays a role in M. tuberculosis infection of macrophages, we first tested to see whether AKT was activated by infection with virulent M. tuberculosis. AKT activation requires phosphorylation at two residues, Thr-308 and Ser-473 [32]; Western blot analysis for this phosphorylation is a standard assay for activation. Using Western blot analysis for phosphorylation of Ser-473, we observed rapid activation of AKT upon infection of J774 macrophages with M. tuberculosis (Figure 5A). Importantly, the activation of AKT was completely abrogated by treatment with the allosteric inhibitor AKTi1/2 that is known to block the phosphorylation and activation of AKT1 and AKT2 (Figure 5B). To genetically confirm the role of various AKT isoforms in macrophage infection, we knocked down the expression of AKT1, AKT2, and AKT3 in J774 macrophages (Figure S10). Similar to previously reported results [10], we found that the maximum effect was observed with simultaneous knockdown of AKT1 and AKT2. We also found that knocking down AKT3 provided little additional benefit (Figure 5C). That silencing both AKT1 and AKT2 is required for maximal effect may account for the failure to identify AKT in the original RNAi screens that target only one isoform at a time, while a small molecule inhibitor inhibits both isoforms simultaneously.
Imatinib mesylate (Gleevec, STI-571) is a tyrosine kinase inhibitor that is currently used therapeutically for treating chronic myelogenous leukemia (CML). Imatinib inhibits the ABL family tyrosine kinases ABL1 and ABL2 and other related tyrosine kinases [33]. In fact, imatinib's activity against CML may require activity at more than one target [34]. Recently, imatinib was shown to diminish entry of Mycobacterium marinum in mouse fibroblasts and M. tuberculosis in mouse macrophages [28]. In addition, imatinib treatment decreased bacterial replication in M. marinum and M. tuberculosis infections of mice [28]. Identification of the ABL inhibitors GNF-2 and imatinib in our screen, which was designed to avoid inhibition of initial phagocytosis, supports the idea that ABL is important not simply for cell entry but also for later replication and survival of mycobacteria in macrophages. A subsequent study has shown that imatinib increases acidification of the lysosomal compartment of macrophages. This effect was required for the anti-mycobacterial activity of the compound [29].
As imatinib inhibits both ABL1 and ABL2, in addition to other tyrosine kinases, the relevant target during M. tuberculosis infection of macrophages is not clear. Further, in apparent contradiction, silencing of abl1 using siRNA has been reported to increase M. tuberculosis replication in macrophages [11] suggesting that this family of kinases may play complicated roles during infection. We therefore sought to characterize the role of ABL1 in entry and survival of M. tuberculosis in macrophages. To specifically abrogate ABL1 function, we transfected macrophages with siRNA targeting abl1. Although we observed significant silencing of ABL1 expression (Figure S10), we observed no difference in M. tuberculosis entry into cells treated with abl1-specific siRNA, demonstrating that ABL1 may not in fact play a role in M. tuberculosis entry (Figure 5D, Figure S10B). However, after three days of infection we observed a ∼1.9 fold decrease in bacterial replication in cells treated with siRNA targeting abl1 compared to cells treated with a non-specific siRNA control, indicating that ABL1 likely plays a role in M. tuberculosis intracellular survival post entry into macrophages (Figure 5D). Importantly, knocking down expression of only ABL1 gave approximately the same magnitude effect observed in imatinib treated cells (∼2.1 fold), suggesting that the effects of imatinib are likely mediated primarily through ABL1 during M. tuberculosis infection of macrophages.
Although primarily studied in the context of cancer, EGFR has been linked to influenza uptake [35], to regulation of inflammation following rhinovirus infection [36], and to prevention of apoptosis in host cells in bacterially infected gastric epithelial cells [37]. EGFR has not previously been linked to mycobacterial infection. We identified gefitinib (Iressa, ZD-1839), an EGFR inhibitor used for treating non-small cell lung cancer [38], and two other EGFR inhibitors in our primary screen (Table S1). Subsequent testing in J774 macrophages and BMDM confirmed that gefitinib treatment controls M. tuberculosis intracellular growth (Figures 4A–4B); we confirmed that EGFR is in fact expressed in macrophages using reverse transcription and PCR-based detection of transcript (Figure S11).
To confirm that the small molecules function by specifically inhibiting EGFR signaling, we sought to perturb EGFR signaling in an alternative manner. As EGFR is a well-validated target for anticancer therapy, several monoclonal antibodies for blocking human EGFR have been developed. In particular, treatment of cells with combinations of non-competitive antibodies against EGFR results in synergistic receptor downregulation via recycling inhibition in a manner that does not result in activation of EGFR signaling [39]. We first confirmed that gefitinib effectively blocks M. tuberculosis replication in primary human monocyte derived macrophages using a total of four donors (data not shown). Next, we treated infected human macrophages with two EGFR-neutralizing antibodies [39]. Treatment with neutralizing antibodies inhibited intracellular M. tuberculosis growth to the same degree as treatment with gefitinib. (Figure 6A), confirming the role of EGFR in M. tuberculosis intracellular survival and replication in primary human macrophages.
There are several signaling pathways that can be triggered downstream of EGFR activation, including activation of AKT, ERK, JNK, and MAPK p38. To determine whether the effect of EGFR inhibition was mediated by any of these signaling pathways, we first sought to determine whether they were activated by M. tuberculosis infection of macrophages. Both AKT and p38 were phosphorylated upon infection (Figures 5A and 6B), consistent with activation. ERK and JNK did not appear to have increased phosphorylation upon infection (data not shown). Surprisingly, AKT phosphorylation was not inhibited by gefitinib treatment (Figure 5B), suggesting that the effect of inhibiting EGFR during M. tuberculosis infection is independent of AKT signaling. To determine whether p38 signaling is impacted by gefitinib, we measured phosphorylation of p38 and found that it was consistently inhibited in the presence of gefitinib treatment (Figure 6B). While p38 phosphorylation decreased over time after infection even in the absence of compound, gefitinib-treated cells had significantly less p38 phosphorylation than DMSO-treated cells at multiple timepoints. To determine whether inhibition of p38 phosphorylation could be the mechanism by which EGFR inhibition of gefitinib treatment restricts M. tuberculosis growth in macrophages, we tested whether selective p38α inhibitor AMG548 [40] would also restrict intracellular mycobacterial growth. We found that inhibition of p38α by AMG548 does in fact restrict mycobacterial growth in a dose-dependent fashion (Figure 6C). MAPK p38 has been demonstrated to regulate autophagy via p38IP and mATG9, with signaling through p38 associated with inhibition of autophagy and depletion of p38 with activation of autophagy [41]. Consistent with our finding that gefitinib induces autophagy (Figure 3A), this report that p38 inhibition can induce autophagy [41], and our implication of p38 downstream of EGFR activation after M. tuberculosis macrophage infection, we suggest that inhibition of p38 activity downstream of gefitinib-mediated EGFR inhibition may trigger an increase in autophagy within the cell, resulting in enhanced clearance of M. tuberculosis.
As there are few studies implicating EGFR in bacterial pathogenesis, we sought to determine whether EGFR signaling is relevant to M. tuberculosis replication in vivo. Given the problem of drug-resistant tuberculosis and the long courses of therapy currently required for treatment, there is growing interest in host-directed therapies as adjunctive therapy, with the potential to shorten therapy and increase the efficacy of current antitubercular antibiotics [11], [17], [28], [42]. To determine whether gefitinib has efficacy against M. tuberculosis in vivo, we examined the efficacy of this inhibitor during early acute infection with M. tuberculosis in the mouse model. We chose to test during early infection prior to the recruitment of IFN-γ producing T cells to the lungs, as this is the time period that is best modeled by in vitro infection of unactivated macrophages. BALB/c mice were infected with M. tuberculosis via aerosol inoculation. Infection was allowed to progress for 7 days prior to initiation of treatment. Mice were subsequently treated with gefitinib at a dose of 100 mg/kg every day for six days. On day 14 after infection, 7 days after initiation of gefitinib treatment, mice were sacrificed, and lungs were homogenized and plated for CFU. The experiment was performed three times and the combined results are shown. Relative to DMSO-treated control mice, mice treated with gefitinib had statistically significantly less bacterial burden (Figure 6D). These results support the in vivo relevance of EGFR in infection and the idea that host-targeted therapies, and kinase inhibitors specifically, may be useful as adjunctive treatment for tuberculosis. Further work will clearly be required to demonstrate efficacy in additional, more chronic models of M. tuberculosis as well as to determine any potential benefit when used in combination with traditional anti-tuberculosis chemotherapeutic agents.
We have developed a high-content, high-throughput imaging assay to identify small molecules that inhibit the growth of M. tuberculosis within macrophages. We applied our assay to screen a library of known bioactive molecules, and from that screen identified several compound classes with diverse annotated targets that result in reproducible, dose-dependent intracellular M. tuberculosis growth restriction. Out of the compounds of interest in our five primary categories, several hits validated previously described targets [12], [28], while others pointed to the involvement of potentially novel factors important for infection, including EGFR, the serotonin transporter targeted by SSRIs, serotonin and dopamine receptors, and sodium channels. An important caveat is that an annotated mechanism of action may not be relevant to the actual activity of a compound in a given assay; observed activity may be due to some other “off-target” effect. Thus, we have used complementary genetic approaches to validate a few exemplary small molecules and implicate their annotated targets in our identified phenotype of M. tuberculosis growth restriction within macrophages.
Three previous chemical screens had identified compounds that restrict M. tuberculosis growth within the context of macrophages. The first, by Christophe et al., focused on screening compounds that are predominantly without known activities [15]. Moreover, the compounds selected for follow-up target essential bacterial processes rather than host proteins relevant to the host-pathogen interaction. A similar screen by the same group also focused on compounds without known activities that targeted essential bacterial processes [16]. A third chemical screen was, like our screen, skewed toward bioactive compounds that target host proteins [17]. Of our 133 identified active compounds, 21 were screened by Sundamarthy et al., and of those compounds 13 were identified as active by their criteria. While distinct libraries screened explain some of the differences between the lists of hits, more subtle differences contribute as well. Of the 8 compounds identified in our screen that were screened but not identified as hits in the screen by Sundamarthy et al., several score close to their threshold for consideration as a positive result. The discrepancies for those compounds likely reflect differences in the precise cutoff for determining hits and the sensitivity of the respective assays. However, some clear hits from our screen scored poorly in their assay, possibly reflecting biological differences between the assays. While our assay used M. tuberculosis and was designed to identify a visual output that approximated differences in mycobacterial growth as optimized using typical antimycobacterial agents, the assay used by Sundamarthy et al. used Bacille Calmette Guerin and had a complex visual output based on phenotypic changes including host cell morphologic changes, phenotypes felt to represent toxicity to the host cell, bacterial size, bacterial number, and bacterial intensity. In fact, they report that traditional antimycobacterial agents did not perform well in their screen, suggesting that it was not designed primarily to identify molecules that affect bacterial load when compared to the gold standard of CFU. Their phenotype likely represented growth restriction mediated by the particular host processes captured by their visual output, such as enhanced autophagy, rather than overall growth restriction. Overlapping hits between the screens may all have a similar mechanism of growth restriction, while compounds that were hits in our screen but not in the screen by Sundamarthy et al. may have distinct mechanisms of action.
We identified several ion channel inhibitors in our screen, including niguldipine and verapamil, which block calcium channels, and flecainide and quinidine, which block sodium channels. A role for voltage-gated calcium channels in mycobacterial infection has previously been suggested [43]. In addition to altering calcium signaling, at least one calcium channel blocker, verapamil, has been shown to inhibit a bacterial efflux pump that is important in the context of intracellular infection [44]. Multiple calcium channel blockers, including verapamil, are used extensively in the clinic for cardiac applications. Our work adds to the growing body of work that suggests a role for calcium channel blockers as adjunctive therapy. In fact, in a recent paper, Gupta et al. tested the addition of verapamil to the standard background regimen of isoniazid, rifampin, and pyrazinamide to treat tuberculosis in a murine model, and found that addition of verapamil both reduced CFU during the active phase of infection and reduced rates of relapse [45]. Whether these reductions are the result of the effect of verapamil on the host or the bacterium is unclear. Nevertheless, this work validates the possibility that compounds identified in this screen may ultimately be useful as adjunctive agents for the treatment of tuberculosis. A potential role for sodium channels has not previously been described and merits further investigation.
Our second class of targets, GPCRs, has only recently been suggested to have a role in mycobacterial infection. While recent work implicated a single GPCR, the D-3-hydroxybutyrate receptor GPR109A [46], our work identified primarily inhibitors of central nervous system-associated GPCRs, such as serotonin receptors and dopamine receptors. Although their function in macrophages is not well-understood, several central-nervous system associated GPCRs have been described to be expressed in macrophages [47], [48], [49]. While some of these receptors have previously been shown to have a role in modulating inflammatory cytokines in response to specific stimulation [50], [51], none have explicitly been shown to be involved in infection or the host-pathogen interface. Our results raise the intriguing possibility that these receptors play roles in the host immune response that are quite distinct from their roles in the central nervous system.
The membrane transport proteins identified in this screen, including the serotonin transporter targeted by SSRIs, a dopamine transporter, and an acetylcholine transporter, have similarly not been shown to be involved in infection previously. Fluoxetine has been noted to increase systemic TNF-α [26], consistent with our results. The role of TNF-α during infection with M. tuberculosis is complex. Low TNF-α levels are clearly detrimental for the host and lead to impaired control of bacterial replication. However, overproduction of TNF-α can have host detrimental effects resulting from excessive tissue damage, induction of macrophage necrosis, and potentially from signaling bacteria to enter a nonreplicating antibiotic tolerant state [52], [53], [54]. Whether the levels of TNF-α produced during infection of humans with M. tuberculosis are optimal for bacterial replication or for host protection is not clear, and this likely varies with the genotype of both the infecting strain and the infected individual. Tobin et al. demonstrated in a zebrafish model of M. marinum infection that particular host genotypes of LTA4H resulted in either high or low levels of TNF-α production in response to infection, either of which was detrimental for outcomes. They went on to demonstrate that promoter polymorphisms for LTA4H in humans similarly resulted in high or low levels of TNF-α production. In an elegant translational component of their study, they demonstrated that among patients with tuberculosis meningitis, only those with the high- TNF-α producing genotype benefited from the standard addition of corticosteroids to their therapeutic regimen [52]. Conversely, it is likely that individuals that produce low levels of TNF-α upon infection with M. tuberculosis may benefit from adjunctive therapy that enhances production. Given the safety profiles of SSRIs, their widespread use, and their ability to modulate TNF-α levels and induce autophagy during infection, they are attractive candidates for further exploring the possibility of tailoring TNF-α levels to optimize host response in the infected individual.
The bulk of the compounds in our anti-inflammatory category were non-steroidal anti-inflammatories (NSAIDs). Their identification is somewhat surprising as their inhibition of cyclooxygenase should block the production of a metabolite, prostaglandin E2, previously shown to be protective for infected macrophages and to reduce the bacterial burden in infected cells [5]. In fact, one would expect an increase in metabolites generated by competing pathways, including lipoxygenase-mediated production of lipoxins such as lipoxin A4, which has been shown to impair the host response to infection [5]. Thus, an NSAID-induced shift in metabolic balance away from prostaglandin E2 toward lipoxins might be expected to worsen the outcome of infection. In contrast, the identification in our screen of multiple distinct NSAIDs that restrict M. tuberculosis growth in macrophages suggests that the role of eicosanoid pathways in mycobacterial infection is potentially more complex than we understand thus far. Given how inexpensive, well-studied, and generally well-tolerated NSAIDs are, they certainly merit further study in animal models of infection. Corticosteroids were also represented in our anti-inflammatory category. Anti-inflammatory corticosteroids have been used as adjunctive therapy for tuberculosis for decades [55]. The use of corticosteroids for tuberculosis meningitis has been accepted into clinical practice; however, the potential benefit of adjunctive steroids for any other form of tuberculosis has been a matter of debate within the literature. A recent meta-analysis found that steroids significantly reduced mortality associated with tuberculosis infection of all organ groups [56]. Our finding that corticosteroids reduce bacterial burden in a cellular model of infection raises the question of whether steroids might impact infection at the level of infected macrophages that counteract the immunosuppressive effects on the whole organism level.
We identified multiple kinase inhibitors in our screen. The identification of AKT and ABL family inhibitors validated our screen, as inhibitors of both classes have been previously shown to disrupt mycobacterial proliferation in macrophages. While a previous siRNA screen had identified AKT indirectly through association with genes identified in their screen [10], AKT was not directly identified using siRNA. Our study demonstrates the inherent advantage of small molecules over RNAi in simultaneously targeting multiple redundant isoforms of a given protein. Imatinib, an inhibitor of ABL family kinases, had similarly been demonstrated to be important for mycobacterial infection of macrophages, predominantly by both inhibiting bacterial uptake and subsequent replication [28]. Using primarily a model with M. marinum, which has significant differences from M. tuberculosis in its intracellular lifestyle, Napier et al. found that the effects of imatinib were to restrict M. marinum uptake and limit its intracellular growth, thus implicating imatinib-sensitive tyrosine kinases as important for virulence. In contrast, a previous study targeting ABL1 with siRNA suggested that depletion of ABL1 results in increased proliferation of M. tuberculosis in macrophages [11]. Here we demonstrate that ABL1 is specifically required for proliferation of M. tuberculosis in macrophages. Our study does not show reduced uptake of M. tuberculosis in cells treated with imatinib or with reduced ABL1 expression. Instead, reduction of ABL1 expression or inhibition of ABL function seems to specifically restrict intracellular growth at a later stage.
Finally, we identified EGFR/p38 MAPK signaling pathway as a novel regulatory pathway during mycobacterial infection that functions to suppress effective antimicrobial responses. Inhibition of EGFR appears to restrict growth of intracellular mycobacteria through induction of autophagy in an AKT-independent mechanism, potentially through downstream inhibition of p38 MAPK. Previous studies have demonstrated a role for p38 MAPK as a negative regulator of both basal and starvation-induced autophagy via p38IP and mATG9 [41]. Our data suggests that EGFR signaling during M. tuberculosis infection activates a similar p38 dependent pathway that prevents clearance of the bacteria by autophagy. Using an EGFR inhibitor in a murine model of infection, we show the relevance of EGFR signaling in vivo and demonstrate that targeting the host with compounds already in clinical use for other applications holds potential for novel therapeutics for tuberculosis. This work thus demonstrates a possible therapeutic strategy of targeting host factors that modulate intracellular M. tuberculosis infection and replication. Repurposing of agents already in clinical use for other indications could expedite the testing of this strategy for tuberculosis.
The general idea of modulating the host response to improve the outcome of tuberculosis treatment has circulated for some time. This concept however, is enormously complex as any benefit to the patient must integrate the impact of host-targeted intervention across all cell and tissue types and all systemic responses, and throughout the whole course of infection. Thus, benefits on a cellular, macrophage level may be counterbalanced by detrimental systemic effects involving numerous cell types and cytokine responses, or vice versa. Further, interventions with benefits at one point in infection, for example early in an inflammatory process, may not necessary be beneficial late in infection. Nevertheless, treatments that limit inflammation have been used both clinically and in experimental settings as adjunctive therapy for conventional antibiotics. Treatment of human patients with corticosteroids results in a modest decrease in mortality and is helpful in some forms of extrapulmonary tuberculosis including meningitis and pleural disease [56], possibly by limiting host inflammatory related tissue damage and/or by allowing M. tuberculosis to transition into a state of active replication in which it is sensitive to antibiotics [57]. Similarly, while anti-TNF-α agents lead to reactivation of mycobacterial disease, at the same time, blocking TNF-α has been suggested to favor a host microenvironment that favors bacterial clearance, particularly in the face of tubercular chemotherapy. Specifically, agents that either directly block TNF-α or inhibit signaling mechanisms that indirectly result in TNF-α production have been used to enhance the responsiveness of bacteria to conventional antibiotics. The TNF-α blocking agent etanercept increased the efficacy of conventional antibiotics during the chronic phase of infection, when the bulk of the bacterial are thought to be replicating slowly [58]. FDA-approved phosphodiesterase (PDE) inhibitors, which alter intracellular levels of cAMP resulting in reduced TNF-α secretion, likewise have been shown to reduce bacterial burden in rabbit and mouse models of infection when combined with current antimycobacterial antibiotics [42], [59]. Of note, an inhibitor of PDE-4 enhanced the effect of isoniazid on clearance of bacteria, but did not have an effect alone [42]. Inhibitors of PDE-3 and PDE-5 were not tested in the absence of standard tuberculosis therapy, so whether they would have an effect alone is unknown [59].
In contrast to work studying therapies that potentially manipulate immunity at a systemic level, a growing body of literature including this current work supports the idea that targeting specific host-pathways to enhance molecular mechanisms for bacterial clearance on the cellular level may be an effective adjunctive strategy for treatment [60]. Based on known host-pathogen biology or the RNAi screens described above, a variety of host targets have been identified and tested in animal models. Targeting lipid metabolism, a liver X receptor inhibitor reduced bacterial burden in the lungs of mice infected with M. tuberculosis [61]. As noted above, imatinib, an ABL inhibitor, has been shown to reduce bacterial burden in a murine model of infection [28]. Additionally, a TGF-β receptor inhibitor [11] and a GPR109A inhibitor [46], selected for study based on findings from the RNAi screen of M. tuberculosis infected THP-1 cells, were both shown to be similarly effective in reducing the burden of disease in mice. As the models for infection and treatment differ from study to study, it is difficult to directly compare efficacy between studies. Unlike the PDE inhibitors however, the liver X receptor inhibitor, ABL inhibitor, TGF-beta inhibitor, GPR109A inhibitor, and now an EGFR inhibitor all have efficacy even in absence of traditional tuberculosis therapy. These results suggest that enhancing host mechanisms of intracellular killing is a viable option for novel TB therapies, while reiterating the need to move quickly from in vitro cellular models to infected animals to determine the effect of an intervention on the complex, intact immune response of a whole organism. Additionally, efforts to determine the timing during an infectious process or the host genetic background where host intervention might be most beneficial will be important to such approaches.
As the problems of MDR- and XDR-TB grow globally, identifying new therapeutic approaches will be critical for decreasing morbidity and mortality, and potentially disrupting the transmission of these highly resistant strains. Traditional drug development pipelines for anti-tuberculosis antibiotics have proven slow to move from lead compounds to clinically deployed medications. As several host-acting compounds that are already in clinical use with well-understood pharmacology and side effects have been shown to be effective in animal models of disease, such compounds could be rapidly tested in expanded animal models of infection and perhaps even moved to humans directly, given the limitations of existing animal models, to clarify the role of host-targeted therapies on treatment efficacy and duration. Alternatively, for some host-modulating compounds that are commonly used within patient populations (i.e, calcium channel blockers, NSAIDs, SSRIs), retrospective data may exist to provide initial support for any potential benefit of these host-targeted inhibitors. While some host-targeted therapies, including some kinase inhibitors, would currently be relatively expensive to administer, they are within the cost range of other medications being studied for use against for highly drug-resistant tuberculosis, including linezolid, and with time, their costs will drop as they become available off patent [62]. Others, such as some SSRIs or calcium-channel blockers, are currently inexpensive and would be amenable to inclusion in treatment regimens. Given the current state of tuberculosis, with rising incidence of MDR, XDR, and even TDR-TB, novel strategies are required that move beyond the conventional paradigm of an antibiotic that kills the bacterium in axenic culture. Targeting the host is one such strategy that can be tested as a feasible path forward exists, facilitated by the repurposing of current drugs.
Animal work was approved by Massachusetts General Hospital IACUC (protocol number 2009N000203) or the Harvard Medical School HMA Standing Committee on Animals (protocol number 03000). All protocols conform to the USDA Animal Welfare Act, institutional policies on the care and humane treatment of animals, the “ILAR Guide for the Care and Use of Laboratory Animals,” and other applicable laws and regulations.
J774 macrophages were seeded into 96-well black clear-bottom plates. M. tuberculosis strain H37Rv constitutively expressing GFP [18] was grown to mid-log phase in axenic culture, washed in PBS, briefly opsonized in heat-inactivated horse serum, and used to infect cells at an MOI of 1∶1. Infection was allowed to progress for four hours, then media was aspirated, cells were washed once with PBS, and media containing the screening compounds at an average concentration of 5 uM was added back to cells. Three days after infection, media was aspirated, cells were washed once with PBS, and fixed with 4% paraformaldehyde with Triton X-100 and DAPI.
Plates were imaged using an Image Xpress Micro high-throughput microscope (Molecular Devices). Images were taken with a 4× objective at four sites per well. Images were then analyzed using CellProfiler open-source software [19]. The imaging-analysis pipeline is openly available (http://cellprofiler.org/published_pipelines.shtml) and included correction to homogenize illumination over each field, a filter to remove any large debris from the analysis, identification and quantitation of DAPI-stained nuclei, identification, quantitation, and pixel intensity calculation for GFP-expressing bacteria. The final output was calculated as (average GFP pixel intensity per bacterium across the field)×(number of bacteria identified in the field)/number of nuclei per field. The four images sites per well were averaged. To identify hits we used p-values obtained by bootstrap Monte Carlo and composite z-scores (see Methods S1).
J774 cells were seeded into 24 well plates. M. tuberculosis strain H37Rv was prepared as described above, then used to infect cells at an MOI of 1∶1. Phagocytosis was allowed to progress for 4 hours; cells were then washed once with PBS and fresh media containing compound was added back. Day 3 after infection, cells were washed once with PBS, lysed in 0.5% Triton X-100, and plated on 7H10 plates in serial dilutions.
Bone marrow was obtained from C57BL/6 mice. In brief, adult male mice were euthanized and femurs and tibias were harvested. Bone marrow was flushed from the cells, resuspended in DMEM, and plated non-tissue culture treated dishes in DMEM media containing 2 mM L-glutamine, 20% fetal bovine serum, and 25 ng/ml recombinant mouse M-CSF. Cells were harvested day 6 after bone marrow isolation and either plated for subsequent infection or frozen. For infections, cells were plated in 24 well plates and infected as above with H37Rv. Media was changed every two days. Day 5 after infection, cells were washed once with PBS, lysed with 0.5% Triton X-100, and plated for CFU.
For siRNA silencing experiments, J774 cells were plated in a 6-well dish. 20 pmol siRNA duplex was added in Optimem (Gibco) with 9 µl Lipofectamine RNAiMax (Invitrogen) at 24 h and again at 48 h after plating. The following day, cells were harvested, counted and re-plated for M. tuberculosis infections. 24 h after replating, the cells were infected with H37Rv as described above. To assess the efficiency of silencing, lysates were prepared at the same timepoint that the cells are infected with M. tuberculosis. For experiments blocking EGFR signaling, anti-EGFR antibody 225 (Millipore) and EGFR antibody Ab-5 (Thermo Scientific) were used at a final concentration of 20 nM. Media containing antibodies was replenished every 24 h.
For monitoring LC3 conversion by Western blot analysis, 2×105 J774 macrophages were plated per well in 12 well dishes and were infected with H37Rv at an MOI = 1. After 4 h the infected monolayer was washed once with PBS and media containing inhibitors was added. Three hours later the cells were washed, and protein lysates were prepared and run on a 15% SDS-PAGE gel. LC3-I and LC3-II were detected using an antibody from Cell Signaling Technologies. For TNF-α ELISAs, J774 or mouse BM macrophages were plated at 5×104 cells per well in a 96 well plate and infected at an MOI = 1. Monolayers were washed with PBS after 4 h phagocytosis and media with inhibitors was added. Cell supernatants were collected 24 h later, and assayed for TNF-α using an ELISA kit (Invitrogen). For phospho-p38 and p38 westerns, J774 cells were infected with H37Rv at an MOI of 1. Phagocytosis was allowed to proceed for 4 hours. Cells were then washed with warmed media, and then treated with gefitinib or DMSO carrier. Cells were harvested at 0 minutes, 15 minutes, 30 minutes, and 60 minutes after drug treatment and probed for phospho-p38 or p38 using antibodies from Cell Signaling Technologies.
BALB/c mice were infected in a Madison aerosol chamber as previously described [63]. 5 mice were sacrificed day 1 after infection, and their lungs were homogenized and plated for CFU to determine the number of implanted bacteria. Infection in the remaining mice was allowed to progress for 7 days. Beginning day 8 after infection, mice were then given intraperitoneal injections of gefitinib at 100 mg/kg or DMSO carrier for 6 days. Day 14 after infection, 5 mice in each experimental group were sacrificed, and lungs were homogenized and plated for CFU.
|
10.1371/journal.pgen.1005539 | Dynamic Contacts of U2, RES, Cwc25, Prp8 and Prp45 Proteins with the Pre-mRNA Branch-Site and 3' Splice Site during Catalytic Activation and Step 1 Catalysis in Yeast Spliceosomes | Little is known about contacts in the spliceosome between proteins and intron nucleotides surrounding the pre-mRNA branch-site and their dynamics during splicing. We investigated protein-pre-mRNA interactions by UV-induced crosslinking of purified yeast Bact spliceosomes formed on site-specifically labeled pre-mRNA, and analyzed their changes after conversion to catalytically-activated B* and step 1 C complexes, using a purified splicing system. Contacts between nucleotides upstream and downstream of the branch-site and the U2 SF3a/b proteins Prp9, Prp11, Hsh49, Cus1 and Hsh155 were detected, demonstrating that these interactions are evolutionarily conserved. The RES proteins Pml1 and Bud13 were shown to contact the intron downstream of the branch-site. A comparison of the Bact crosslinking pattern versus that of B* and C complexes revealed that U2 and RES protein interactions with the intron are dynamic. Upon step 1 catalysis, Cwc25 contacts with the branch-site region, and enhanced crosslinks of Prp8 and Prp45 with nucleotides surrounding the branch-site were observed. Cwc25’s step 1 promoting activity was not dependent on its interaction with pre-mRNA, indicating it acts via protein-protein interactions. These studies provide important insights into the spliceosome's protein-pre-mRNA network and reveal novel RNP remodeling events during the catalytic activation of the spliceosome and step 1 of splicing.
| The spliceosome is a highly dynamic RNP machine that during the catalytic cycle undergoes many changes in composition and conformation. The pre-catalytic Bact spliceosome contains the U2, U6 and U5 snRNAs and ~40 proteins, which are evolutionarily conserved between budding yeast and metazoans. The Bact spliceosome is converted to a catalytically-activated B* spliceosome and following recruitment of the Cwc25 protein, step 1 of splicing is catalyzed and the C spliceosome is generated. The U2 snRNP plays an essential role in branch-site selection and pre-mRNA splicing catalysis. During the Bact to B* transition the affinity of several U2 SF3a/b proteins for the spliceosome is significantly reduced. Whether this is due to remodeling events affecting U2 snRNP contacts with the pre-mRNA is not known. Information about conserved spliceosomal protein-pre-mRNA contacts and their dynamics during splicing remains limited. Here we investigated pre-mRNA–protein contact sites in yeast Bact spliceosomes by UV-induced crosslinking. We detected contacts of nucleotides surrounding the branch-site with several of the U2 SF3a/b proteins, and we show that these interactions are evolutionarily conserved. We carried out a similar investigation with B* and C spliceosomes and provide important insights into the dynamics of pre-mRNA–protein interactions involving the essential U2, RES, Cwc25, Prp8 and Prp45 proteins.
| The removal of introns from nuclear pre-mRNAs proceeds by way of two phosphoester transfer reactions and is catalyzed by the spliceosome, a large ribonucleoprotein (RNP) complex composed of the snRNPs U1, U2, U4/U6 and U5 and several proteins [1]. The spliceosome is a highly dynamic RNP machine that undergoes many changes in composition and conformation during its work cycle [2].
Initially, the U1 snRNP recognizes the 5’ splice site (5’ SS) and U2 snRNP recognizes the branch-site (BS) of the pre-mRNA, resulting in the formation of the pre-spliceosome or A complex. The pre-formed U4/U6.U5 tri-snRNP is then recruited, generating the B complex, which does not yet have an active site. Subsequent activation of the spliceosome (leading to the Bact complex) involves major rearrangement of the spliceosomal RNA–RNA interaction network. This rearrangement is catalyzed by the combined action of the RNA helicases Prp28 and Brr2 and leads to the displacement of the U1 and U4 snRNAs and the formation of new base-pair interactions between the U2 and U6 snRNAs and the 5’ SS [3]. Thus, a web of RNA–RNA interactions holds the 5’ SS and the BS together for step 1 catalysis [4].
The Bact complex, which contains U2, U6 and U5 and ~40 proteins in the yeast S. cerevisiae [5], is converted by the DEAH-box NTPase Prp2, in co-operation with the G-patch protein Spp2, into a catalytically activated complex (B*) [6–8]. Following the recruitment of the splicing factor Cwc25 [7,9], step 1 is catalyzed, whereby the 2’-OH of the BS adenosine attacks the 5' SS of the pre-mRNA generating the cleaved 5’ exon and intron 3’ exon; concomitantly the C complex is formed. This then catalyses step 2, in which the 3’ SS is cleaved, resulting in the excision of the intron and ligation of the 5’ and 3’ exons, after which the mRNA product is released. The excised intron lariat remains associated with U2, U5 and U6 snRNPs, which then dissociate and take part in subsequent rounds of splicing.
During the transformation of complex B into Bact, not only is the spliceosome's RNA network radically rearranged, but also its protein composition changes significantly; as a result, several proteins are released, while twelve Bact proteins are recruited. At the same time the NTC (nineteen complex) and the NTC-related proteins are stably integrated into the Bact complex [5]. Likewise, the three proteins comprising the RES (pre-mRNA retention and splicing) complex [10] are also stably integrated into the Bact complex.
Despite the substantial restructuring that the spliceosome has undergone at this point, it does not yet have a functional active site. Previous studies showed that the binding affinity of several proteins is significantly changed during the Prp2-mediated transition of Bact spliceosomes to catalytically activated B* spliceosomes [7,11]. During this step, the essential splicing factor Cwc24 is quantitatively displaced from the B* complex. The U2-associated SF3a and SF3b proteins Prp11 and Cus1 and the RES protein Bud13 all remain bound to the B* spliceosome under near-physiological conditions, but their binding is reduced at high salt concentrations [11]. The destabilization of these proteins' binding by Prp2 and Spp2 indicates that the structure of the catalytic core of the spliceosome near the BS is remodeled. This could lead to a proper 5’SS and BS configuration for nucleophilic attack on the 5' SS phosphodiester bond during step 1 catalysis [7,12]. However, while it is clear that the affinity of the U2 and RES proteins for the spliceosome is significantly reduced during catalytic activation, it is not known whether this implies remodeling events involving contact of U2 and RES proteins with the pre-mRNA. Likewise, information about the set of spliceosomal protein-pre-mRNA contacts and their dynamics during splicing remains limited, but is crucial for unraveling potential functions of spliceosomal proteins for the formation and maintenance of the spliceosome's RNA–RNA network during catalysis.
The U2 snRNA/BS interaction is established in the A complex and is thought to make the BS adenosine bulge out for nucleophilic attack on the 5’ SS during step 1 catalysis [13]. In human pre-spliceosomes and spliceosomes the U2 SF3a/b proteins help to recruit the U2 snRNP to the BS, and all of them except SF3b130 can be crosslinked, in a sequence-independent manner, to a region upstream of the BS (the so-called “anchoring site”), to the BS itself and to a region downstream of it [14,15]. The BS sequence is highly conserved in yeast but only weakly conserved in metazoans. Given the short length of the BS sequence, and its degeneracy in metazoans, it has been suggested that spliceosomal proteins function together with the U2 snRNA/BS duplex to tether the U2 snRNP to pre-mRNA during spliceosome assembly. In yeast there is perfect complementarity between the BS sequence and the U2 RNA. Thus the anchoring/stabilization of the U2 snRNP to the BS sequence in yeast could be different from that in human, and it may not depend critically on protein–RNA interactions. Although most of the U2 proteins in the yeast S. cerevisiae are evolutionarily conserved [16–18], it is not known whether they interact in a similar way with the BS region in yeast spliceosomes, or whether a site equivalent to the human anchoring site exists in yeast pre-mRNAs. So far, only Hsh155, the yeast homologue of human SF3b155, has been shown to crosslink to pre-mRNA between the BS and the 3'SS [19]. Furthermore, it was recently shown that the RES subunit Snu17 is in contact with the pre-mRNA downstream of the BS in proximity of U2-Hsh155 [20]. However, is it currently not known whether additional components of the yeast U2 SF3a/b and RES subunits make direct contact with the pre-mRNA in spliceosomal complexes.
To address these questions, we have investigated protein–pre-mRNA interactions by UV-induced crosslinking of purified yeast spliceosomes stalled at the Bact assembly stage or after conversion of Bact to B* and C complexes, using a purified yeast splicing system [7]. Results of these studies revealed contacts in Bact complexes between pre-mRNA nucleotides directly upstream of the BS and the yeast U2 proteins Prp9, Prp11, Hsh49, Cus1 and Hsh155; the latter were also in contact with the intron further downstream of the BS. Thus, these interactions are evolutionarily conserved between yeast and man. Consistent with previous results demonstrating a Snu17-pre-mRNA crosslink [20], we observed that also the RES components Pml1 and Bud13 crosslinked to the intron downstream of the BS in Bact complexes. Subsequent UV crosslinking with purified spliceosomes that had been stalled after catalytic activation by Prp2/Spp2 and consecutive step 1 catalysis by Cwc25 revealed remodeling events involving contacts between U2 SF3a/b proteins upstream of the BS and the RES proteins downstream of it. Finally, concomitantly with these remodeling events, enhanced contacts of Cwc25, Prp8 and the NTC-related protein Prp45 with the BS and/or 3'SS regions were observed. These studies thus provide novel insights into the extensive protein–pre-mRNA interactions and their dynamics within and surrounding the pre-mRNA BS and 3'SS regions during step 1 of splicing in yeast spliceosomes.
To obtain insights into the nature and number of proteins that are in direct contact with the region at the 3’ end of the intron in purified yeast spliceosomes, we constructed a pre-mRNA which was body-labeled with 32P-UTP during transcription in the 3’ third of the intron, including exon 2 and 47 nucleotides (nts) upstream of the BS (termed hereinafter “3’-region-labeled pre-mRNA”; Fig 1A). The experimental strategy used to produce the 3’-region-labeled pre-mRNA is outlined in S1A Fig. Briefly, the 3’ fragment was obtained by T7 transcription. For this purpose, a T7 promoter was added by PCR and the PCR product was transcribed in vitro with an excess of GMP to ensure the presence of a monophosphate at the 5’ end and with α-32P UTP to randomly trace-label the entire RNA transcript (see S1 Text for details). To produce the 5’ fragment, unlabeled actin pre-mRNA, prepared by transcription in vitro, was specifically cleaved between nucleotides 425 and 426 by a DNA enzyme based on the “8–17” catalytic motif [21] (S1A Fig, upper panel). The 5’ cleavage fragment was dephosphorylated, gel purified and ligated to the T7-transcribed 3’ fragment by DNA splint directed RNA ligation [22]. The 3’-region-labeled pre-mRNA allows the analysis by UV crosslinking of protein–pre-mRNA interactions at the BS site, the region directly upstream of the BS as well as around the 3’SS.
Protein–pre-mRNA interactions were analyzed initially in purified Bact spliceosomes, which were assembled in vitro by incubating heat-inactivated splicing extracts from a temperature-sensitive prp2-1 yeast strain with the 3’-region-labeled pre-mRNA that also contained an MS2 binding site at its 5’ end [5,7]. Bact spliceosomes were purified by glycerol-gradient centrifugation and MS2-MBP affinity chromatography and then were irradiated with UV light at 254 nm, and digested under denaturing conditions with a mixture of RNases T1, A and I. The entire protein mixture was then separated by two-dimensional (2D) gel electrophoresis as described for human spliceosomal complexes [23]. Our 2D gel electrophoresis method is based on charge-driven separation of proteins under denaturing conditions at acidic pH in the first dimension and further separated by molecular weight though SDS gradient PAGE in the second. In contrast to the commonly used isoelectric focusing (IEF), this system prevents proteins from reaching zero charge and allows separation without in-gel precipitation over a wide range of isoelectric points (IEPs) and with masses greater than 300 kDa [23].
Fig 1B shows a RuBPS-stained 2D gel (S1 Text) of the total proteins isolated from non-irradiated Bact complexes. Individual protein spots were cut out of the gel and peptides were identified by mass spectrometry (MS). Only a few contaminant proteins were found, such as Xnrn1/Kem1 and Hrb1/Tom34, which are usually present in small amounts in preparations of yeast spliceosomes [5,7]. All the previously identified Bact complex proteins were observed [5]; these included nearly all of the U2 SF3a/b proteins (i.e. SF3b: Rse1, Hsh155, Cus1 and Hsh49; SF3a: Prp9, Prp11 and Prp21), which could be well separated from each other. The low-MW U2 proteins Msl1, Rds3 and Ysf3 and the Sm proteins D2, E, F and G ran out of the gel but could be identified in 2D gels which were modified to improve the resolution of smaller proteins [23]. A subset of U5 proteins (Prp8, Brr2 and Snu114), and most proteins of the NTC complex and NTC-related proteins, were also located as single, distinct spots. Proteins of the RES complex (Ist3/Snu17, Pml1 and Bud13) [10] were also identified. A comparison of previous MS analyses of purified yeast Bact spliceosomal complexes [5,7,24] with those of our 2D analysis indicates a general reliability of this method for separating and identifying proteins that co-purify with yeast spliceosomal complexes [23].
Fig 1C shows an autoradiography of the 2D gel performed as described above but with UV-irradiated Bact complexes. Exposure to 254-nm UV light is known to induce direct (zero-length) crosslinks between nitrogenous bases of nucleic acids and amino-acid side chains when they are in a favorable configuration. We observed prominent 32P-labeled spots of U2-Hsh155 and the NTC-related protein Prp46, both of which could be superimposed on the RuBPS-stained spots (indicated by circles in Fig 1C). This indicates that the covalent attachment of a few RNA nts to proteins larger than 50 kDa, after UV-irradiation, did not alter significantly their migration behavior. A predominant crosslink in the middle of the autoradiogram was due to the contaminating poly(A)-binding protein Hrb1/Tom34, while other contaminant proteins were crosslinked to pre-mRNA at very low levels or not at all (marked by asterisks in Fig 1C). Prominent radioactive spots were also observed for smaller U2 proteins (MW < 50 kDa), such as U2-Hsh49 and also two proteins of the RES complex (Pml1 and Snu17). The covalent attachment of RNA nts to smaller proteins led to a shift of their crosslinked species to the acidic region (i.e. left side) of the gel in the first dimension; in addition, they were separated into several spots and did not co-localize with the RuBPS stained spot, which were located slightly below (Fig 1C). Nevertheless, in all three cases crosslinked species migrated to the left side of the gel, where no other co-migrating proteins were visible in the RuBPS stained gel, indicating that the crosslinked proteins of interest were not contaminated with other proteins. We recently showed that the RES complex subunit Snu17 crosslinks in the Bact complex to a 14-nt-long region of the pre-mRNA intron downstream of the BS, as shown after digestion with RNase T1 [20]. Thus, the presence of several spots in the 2D gel may indicate that crosslinked pre-mRNA–protein species included shorter digestion products of the 14-nt-long RNA fragment (note that treatment with three different RNases was performed for 2D gel analysis). The same is likely to be true for RES-Pml1 and U2-Hsh49, whose crosslinked species showed a similar separation behavior (Fig 1C). The unequivocal identification of these proteins will be demonstrated below.
We also observed that low levels of additional proteins crosslinked to the 3’ part of the intron, such as the U5 proteins Prp8 and Brr2, the U2 proteins Prp9, Prp11, Cus1, the RES protein Bud13 and a few Bact-specific proteins (i.e. Cwc22, Cwc27/Cwc2 and Isy1, indicated by dots; Fig 1C). These proteins crosslinked much less strongly to the 3’ part of the intron in the Bact spliceosome than those described above, suggesting that they are in contact with the 3’ region of the pre-mRNA but are in a conformation that does not favor the formation of UV-induced crosslinks. Alternatively, digestion with a mixture of three different RNases before 2D gel electrophoresis may lead to partial digestion of the crosslinking site.
Next, we focused on the characterization of the crosslinked U2 SF3a/b and RES complex proteins with two major objectives: (i) identification of the candidate proteins crosslinked to pre-mRNA by pull-down of their tagged version, and (ii) identification of the region/position of crosslinks within the intron. Therefore, we generated yeast strains carrying a C-terminally TAP-tagged version of most of the U2 SF3a/b proteins and the two RES subunits Pml1 and Bud13. In addition, to localize protein–pre-mRNA interactions to well-defined short regions in the RNA stretch directly upstream of the BS or around and downstream of it, we synthesized site-specifically labeled pre-mRNA (S1B Fig).
We prepared eight different actin–pre-mRNA constructs, each of which harbored a single 32P label directly 5’ of a distinct guanosine residue in the neighborhood of the BS (G452–G516, summarized in Fig 2A). For this purpose, full-length non-32P-labeled pre-mRNAs were cleaved into two pieces at a specific position by using a distinct DNA enzyme; after 5’ 32P-labeling of the 3' piece, the two fragments were ligated by using the DNA splint-directed RNA-ligation method of Moore and Sharp. In this way, full-length pre-mRNAs were recreated, each containing a 32P-label at the desired position [22,25,26] (S1B Fig and Methods for details).
Extracts from the prp2-1 strain harboring the TAP-tagged versions of proteins were then used for assembly of yeast Bact complexes on each of the site-specifically 32P-labeled pre-mRNAs. The purity of Bact complexes was determined by analyzing the composition of their associated snRNAs and pre-mRNA (i.e. for the presence of U2, U5L, U5S and U6 snRNA and the absence of splicing intermediates of the pre-mRNA; S2 Fig). Each purified Bact complex was irradiated with UV light at 254 nm and disrupted by incubating at 70°C in 3% SDS. After complete digestion with RNase T1 (which cleaves 3’ of guanosine residues), the RNA fragments shown in Fig 2A were obtained, each of which contained a single radioactive phosphate 5’ of the terminal guanosine residue. We then immunoprecipitated TAP-tagged crosslinked proteins with IgG Sepharose beads and analyzed the immunoprecipitates by western blotting using the PAP complex (peroxidase-anti-peroxidase) (Fig 2B and 2C, upper panels, western blot). Autoradiography of the membrane revealed the 32P-labeled RNA fragment crosslinked to each precipitated protein (lower panels). Thus, we were able to assign a well-defined pre-mRNA region crosslinked to a known protein and could map the entire intron area spanning from nts 447–516.
We first analyzed the U2 SF3a/b proteins and initially focused on the region upstream and across the BS (Fig 2B). For the RES complex proteins we focused on the region downstream of the BS (Fig 2C) because our earlier results showed that Snu17 is in direct contact with this region [20]. Western blotting confirmed that U2 proteins were immunoprecipitated either before irradiation (–UV) or after it (+UV); however, when UV irradiation was omitted, 32P-labeled fragments were not precipitated with the proteins (Fig 2B, lower panels, autoradiography, lanes 1–4). After UV irradiation, the U2 proteins Prp9, Cus1, Prp11 and Hsh49 were found crosslinked to the pre-mRNA fragments 32P-labeled at the G452 and G460 positions (i.e. fragments 447–452 and 453–460; Fig 2B, lower panels, lanes 5 and 6). None of the U2 proteins analyzed crosslinked to the downstream fragments 461–467 and 468–478, with the exception of Hsh155, which crosslinked to the RNA region 461–467 and with lower intensity to the BS region 468–478 (Fig 2B, lanes 7 and 8). The U2 proteins Rse1 and Prp21 and the two small proteins Rds3 and Ysf3 did not crosslink to pre-mRNA. This is consistent with earlier reports that the putative human homologue of Rse1, SAP130, could not be crosslinked to pre-mRNA [15] and Rds3 did not bind RNA in vitro [27]. Taken together, these results indicate that the U2 proteins Prp9, Cus1, Prp11, Hsh49 and Hsh155 are in direct contact with the pre-mRNA in the Bact complex, and their interaction is confined to a 14-nt-long region of the intron upstream of the BS (447–460), with the exception of Hsh155, which interacts in addition with the intron downstream of the BS [19] (see also below).
Fig 2C shows a similar pull-down experiment performed with purified, irradiated and RNAse-T1-digested Bact complexes carrying TAP-tagged RES Pml1 and Bud13 proteins. Western blotting showed that Pml1-TAP and Bud13-TAP were immunoprecipitated both before and after UV-irradiation. Pml1-TAP crosslinked the most strongly to the pre-mRNA fragments 483–496 and 500–511, while Bud13-TAP crosslinked to the fragment 500–511 (Fig 2C, lower panels). Thus, Pml1 interacts directly with the 14-nt-long region of the intron downstream of the BS (483–496), and both Pml1 and Bud13 interact further downstream than Snu17 [20].
Next, we expanded our analysis to the dynamics of protein–RNA interactions during catalytic activation and step 1 catalysis by using a purified yeast splicing system to investigate changes of UV crosslinking intensities in purified yeast spliceosomes stalled at specific assembly stages after Bact, such as the B* and C complex stages [7]. Bact spliceosomes were prepared as above by incubating distinct site-specifically labeled actin pre-mRNAs (Fig 3A) with prp2-1 heat-inactivated splicing extracts in which proteins were untagged. Each Bact spliceosome was affinity-purified as above. One portion of Bact spliceosomes was complemented with ATP and recombinant Prp2 and Spp2 (whereby transformation of complex Bact to B* occurs) and one portion was complemented with ATP plus Prp2, Spp2 and Cwc25 (whereby transformation of complex B* to C occurs and step 1 is catalyzed). Spliceosomes were then further purified by glycerol-gradient sedimentation. The actual conversion from Bact to B* to C was analyzed for each purified complex by gel electrophoresis (S3 and S4 Figs). The presence of U2, U5 and U6 snRNA, and the total (or, for B*, nearly total) absence of splicing intermediates of the pre-mRNA confirmed the identity of the Bact and B* complexes, while the presence of step 1 products confirmed the identity of the C complex (S3C and S4C Figs). In addition, the efficiency of conversion of Bact to B* was determined by western-blot analysis, which revealed the nearly complete dissociation of the splicing factor Cwc24 from the B* complex during catalytic activation, as previously shown by MS and dual-color fluorescence cross-correlation spectroscopy (dcFCCS) [7,11]. The efficiency of conversion of B* to C complexes was monitored from the formation of step 1 splicing products, analyzed by 8% denaturing RNA PAGE and quantified by phosphorimager. The % of step 1 products (compared to the total RNA in a lane) was calculated to be ~40% (S3D and S4D Figs).
Peak fractions of purified Bact, B* and C complexes were irradiated with 254-nm UV light and–after denaturation and digestion with RNase T1 –the crosslinked 32P-labeled proteins were analyzed by SDS-PAGE. The gel was subjected to autoradiography (Fig 3A). Each site-specifically labeled pre-mRNA showed a distinct crosslinking pattern, revealing bands of different intensities and masses. Fig 3A (lanes 1 and 4) shows different degrees of crosslinking of four proteins with sizes consistent with the apparent molecular masses of untagged Prp9, Cus1, Prp11 and Hsh49, which crosslinked to the pre-mRNA fragments 447–452 and 453–460 in the Bact complex. To ascertain that the four untagged crosslinked proteins corresponded to Prp9, Cus1, Prp11 and Hsh49 as shown in Fig 2B, we compared untagged and tagged proteins in parallel experiments. The molecular masses of tagged proteins are increased by a predicted 21kDa, along with the complete disappearance of the untagged version. S5 Fig lane 2 shows the patterns of untagged Hsh49, Prp11, Cus1 and Prp9 crosslinked to the pre-mRNA fragment 32P-labeled at G460 in the Bact complex. When the crosslinked proteins were compared with their tagged versions, we observed that Hsh49 shifted from 25kDa to ~50kDa (compare lanes 2 and 5, red arrow), Prp9 shifted from 60kDa to ~90kDa (compare lanes 2 and 3, yellow arrow), and Prp11 and Cus1 showed the expected size-shifts consistent with the addition of the TAP-tag (compare lane 2 with lanes 4 and 6; green and blue arrows, respectively). Similar comparisons were performed for the identification of Hsh155 and the RES proteins (S5B and S5C Fig). Taken together, these data allow assignment of the radioactive bands shown in Fig 3A to the proteins indicated (on the left and right of the gel).
The pattern in lane 1 of Fig 3A shows that Hsh49 crosslinked in highest yield to the ~6-nt-long region 447–452, whereas Prp9 and Cus1 crosslinked at low levels to the same fragment. Prp11 was not clearly distinguishable from Hsh49; however, as shown by a light exposure of the gel in Fig 3B, its crosslinking yield was very weak. Although the chemistry of the different sites in the RNA and proteins may affect the intensity of the crosslinks, these results do suggest that Prp9, Cus1 and Prp11 make no close contacts with this particular RNA region. Remarkably, however, the intensities of Prp9 and Cus1 crosslinks were much stronger in the ~8-nt-long region 32P-labeled further downstream (i.e. at G460) in the Bact complex, whereas crosslinks of Hsh49 (and Prp11) remained unchanged in this downstream ~8-nt-long region (Fig 3A, lane 4; see also Figs 3B and S5A for the crosslinking intensity of Prp11).
Intriguingly, upon conversion of the Bact to the B* and C complexes, crosslinks of Hsh49 to the ~6-nt-long region (447–452) were greatly reduced, as shown by light exposure of the gel in Fig 3B (lanes 1–3). Quantification of the intensity of Hsh49 crosslinks indicated that it was reduced by 40% and 80% in the B* and C complexes, respectively, relative to the Bact complex (S6A Fig). This indicates that remodeling of the spliceosome leads to a reduced interaction of Hsh49 with the ~6-nt-long region of the intron. Quantification of Prp11 crosslinks was difficult, as its weak signals did not resolve well from the strong signals of Hsh49. Interestingly, reduced crosslinking yield of Hsh49 to the downstream ~8-nt-long fragment (453–460) was also observed during spliceosome remodeling. However, the yield of crosslinks of Hsh49 to this region were reduced to a less significant extent in the B* and C complexes compared with the Bact complex (20% and 40%, respectively; S6A Fig). This indicates that Hsh49 maintains a relatively strong interaction with the ~8-nt-long region of the intron during catalytic activation and step 1 catalysis, compared to the upstream ~6-nt-long region. Similarly, during spliceosome remodeling the levels of crosslinking of Prp9 to the ~8-nt-long fragment (453–460) decreased by ~40% in both B* and C complexes compared with the Bact complex, indicating that also the binding site of Prp9 is destabilized (Figs 3A and 3B, lanes 4–6, and S6A). Although Cus1 was difficult to quantify, the pattern of its crosslinking seemed reduced to a similar extent (see Figs 3B, lanes 4–6, and S6A). Taken together, these results revealed that contacts of Hsh49, Prp11, Cus1 and Prp9 with adjacent regions of the intron were reduced to various degrees during the spliceosome’s conformational changes, indicating remodeling of the binding sites of these proteins.
To obtain independent evidence of the decrease over time of Hsh49 crosslinks, we performed 2D gel electrophoresis with UV-irradiated and RNase-digested B* and C complexes assembled on the 3’-region-labeled pre-mRNA, and compared the intensities of the signal from their crosslinked species with those observed in corresponding experiments with the Bact complex (Fig 3C). Consistently with the data shown in Fig 3A, the crosslinking level of Hsh49 to the 3’-region-labeled pre-mRNA was high in the Bact complex. However, the crosslinking level of Hsh49 decreased by ~10% in the B* complex and by more than 40% in the C complex relative to the Bact complex (Fig 3C), as determined by quantification of the radioactive spots (S6B Fig for quantification of Hsh49 in 2D gels). Thus, these results confirmed that contacts of Hsh49 with the 3’ regions of the intron were reduced during the spliceosome’s conformational changes, indicating remodeling of its binding sites.
A protein with the molecular mass of ~110 kDa, which was identified as Hsh155 (S5B and S5C Fig), crosslinked upstream and downstream of the BS. Hsh155 crosslinked strongest to the pre-mRNA fragment immediately upstream of the BS (461–467), and weakest to the BS fragment itself (468–478; Fig 3A, lanes 7–12), consistently with the result of the pull-down experiment shown in Fig 2B. Furthermore, we observed enhanced crosslinking of Hsh155 to the entire region downstream of the BS (Fig 3A, lanes 13–21). These results indicate that Hsh155 is in contact with a large 50-nt-long region of the intron and it spans the BS; however, its pattern of interaction with the intron does not seem to change significantly during remodeling of the Bact to B* and C complexes, with a decrease in the yield of crosslinking of only ~20% (S6A Fig).
There were also additional crosslinks that were not compatible with any obvious U2 snRNP protein equivalent (Fig 3A, marked with question marks). This suggests that there are additional proteins that also contribute to the protein–pre-mRNA interaction network in this region. Of interest is a 25 kDa protein that crosslinked with low intensity to the 11-nt-long BS fragment in the C complex (see below for the characterization of this protein). In addition, a ~15 kDa protein was observed that crosslinked to the BS fragment in Bact, B* and C complexes; the identity of this protein could not be determined either by 2D gel electrophoresis or by tagging the small U2 proteins Rds3 or Ysf3. Furthermore, crosslinking of Prp8 and Prp45 was identified (see below for a detailed description).
To shed some light on the dynamic interactions between the RES complex proteins and the intron, we investigated possible changes of their crosslinking pattern as described above for the U2 proteins. A protein of ~20 kDa was efficiently crosslinked to the pre-mRNA fragment 483–496 (Fig 3A, lanes 16–18). This protein, which was identified as Snu17 (see S5B and S5C Fig, lane 3), crosslinked strongest in the Bact complex; however, the intensity of crosslinking decreased by ~ 70% in the B* and C complexes relative to the Bact complex (Figs 3B and S6C), indicating that the interaction of Snu17 with the 14-nt-long region 483–496 is weakened after activation of the spliceosome by Prp2 (Fig 3A and 3B, lanes 16–18, see S7 Fig for an independent experiment). Analysis of Snu17 crosslinks by 2D gel electrophoresis, confirmed that the binding of Snu17 to the 3' region of the pre-mRNA is drastically reduced during spliceosome remodeling (Figs 3C and S6B for quantification of Snu17 in 2D gels)
The intensity of Pml1 crosslinking was much lower than that of Snu17 in the same region of the intron, suggesting that Pml1 either makes no close contacts with this region or simply binds in a manner unsuitable for forming crosslinks. Nonetheless, during transition of the spliceosome from Bact to B* the crosslinking intensity of Pml1 was reduced by ~50% and by another 10% from B* to C (Figs 3A, lanes 16–18 and S6C). A similar decrease in crosslinking intensity was confirmed by 2D gel electrophoresis (Figs 3C and S6B for quantification of Pml1 in 2D gels). Likewise, the levels of Pml1 and Bud13 crosslinks to the region further downstream (500–511) decreased by ~50–60% in the B* and C complexes relative to Bact (Figs 3A, lanes 19–21, and S6C). The intensity of Snu17 crosslinking was dramatically reduced in the 500–511 region of the intron, as compared with that upstream (483–496), indicating that the closest interaction of Snu17 with the intron is with the 14-nt-long region 483–496. This is consistent with results of our earlier pull-down experiments, which showed that Snu17 interacts directly with this region [20]. Taken together, these data indicate remodeling events involving RES protein contacts with the intron region downstream of the BS upon spliceosome conformational changes.
A large protein with a molecular mass of ~250 kDa, which is the expected size for Prp8, crosslinked in very low yield to regions 461–467 and 468–478 of the intron (Fig 3A, lanes 7–12, indicated by arrowheads). Quantification of these crosslinked species revealed that Prp8 crosslinked to the region 461–467 of the intron: first in the B* complex and with increased level (by ~30%) in the post-step 1 spliceosome (Figs 3A, lanes 8 and 9, and S6C). Interestingly, Prp8 crosslinked weakly also to the BS sequence 468–478 in the C complex, indicating that during/after step 1 catalysis, Prp8 is favorably positioned for interaction with the BS (Fig 3A, lane 12). A stronger Prp8 crosslink was observed further downstream, to the 14-nt-long region 483–496 in the B* complex, the intensity of which was enhanced in the C complex (Fig 3A, lanes 16–18; see also S7 Fig for an independent experiment). Thus, our results indicate that Prp8 is favorably positioned for its interaction with the BS upon catalytic activation of the spliceosome by Prp2/Spp2, and with the 3’SS region upon subsequent step 1 catalysis by Cwc25.
Again independent evidence of the temporal increase of Prp8 crosslinks was obtained by 2D gel electrophoresis performed with UV-irradiated and RNase-digested B* and C complexes assembled on the 3’-region-labeled pre-mRNA, and the intensities of their crosslinked species was compared with those observed in the Bact complex (Fig 3D). Consistent with the data shown in Fig 3A, the crosslinking level of Prp8 to the 3’-region-labeled pre-mRNA was very low in the Bact complex, indicating that Prp8 makes no close contacts with the 3' region of the intron before catalytic activation by Prp2/Spp2. However, the crosslinking level of Prp8 increased more than 60% in the B* complex and even more than 90% in the C complex relative to the Bact complex (Figs 3D and S6B). Taken together, these results indicate that contacts of Prp8 with the BS and 3’SS regions begin during/after the catalytic activation by Prp2/Spp2 and the interaction with the 3'SS is enhanced after step 1 catalysis promoted by Cwc25, and are consistent with previous results that showed contacts of Prp8 with the 3'SS subsequent to step 1 catalysis in yeast extracts [28–31].
To determine the identity of the 25kDa protein, which crosslinked to the BS fragment 468–478 in the post-step 1 spliceosome, we used recombinant full-length Cwc25 and truncated variants thereof in reconstitution of the C complex. Reconstituted C complexes were UV-irradiated and RNAse-T1-digested as above. Fig 4A shows that recombinant Cwc25 crosslinked to the BS fragment in the C complex (lane 6). A truncated variant of Cwc25 (residues 1–168), lacking 11 amino acids at the C-terminus, showed a similar crosslinking yield (lane 5), In contrast, the two variants Cwc25 1–102 and 1–125, lacking 77 and 54 amino acids at their C-termini, did not crosslink to this region (lanes 3 and 4). Consistently with previous observations [32], this experiment demonstrates that Cwc25 is in contact with the BS sequence and that at least the N-terminal 168 amino acids of Cwc25 are needed for this.
Intriguingly, the addition of the truncated version of Cwc25 1–168 (Fig 4A, lane 5) to B* spliceosomes promoted step 1 catalysis, which was even more efficient than that observed with the full-length version (Fig 4B, compare lanes 5 and 2). Surprisingly, Cwc25 1–125 promoted step 1 catalysis even in the absence of RNA crosslinking (Fig 4B and 4A lanes 4), indicating that Cwc25’s activity in promoting step 1 can be uncoupled from its activity in RNA-binding/crosslinking. This result suggests that Cwc25 1–125 may still interact with one or more proteins in the neighborhood of the BS and thus render the microenvironment of the catalytic center favorable for step 1 catalysis. Candidate proteins for interaction with Cwc25 are Prp8 and Hsh155, which are shown here to crosslink to the BS region concomitantly with Cwc25 (Fig 4A). Furthermore, Yju2 may also interact with Cwc25, as it was previously shown to be involved in recruiting Cwc25 to the spliceosome [9].
In addition to the NTC, two splicing factors, namely Prp45 and Prp46 [33] that interact with components of the NTC, and whose function is related to NTC in human and yeast, were also observed in the 2D gel carried out with Bact spliceosomes (Fig 1B). We observed that Prp45 did not crosslink in Bact complexes assembled on the 3’-region-labeled pre-mRNA after irradiation with UV light (Fig 1C), in contrast, Prp46, which was previously shown to interact with Prp45 in vitro and in vivo [33], crosslinked in high yield to the 3' end of the intron in Bact spliceosomes (Fig 1C). To determine whether this crosslink was retained during spliceosome remodeling, we prepared 2D gels from crosslinked, RNase-digested B* and C complexes assembled on the 3’-region-labeled pre-mRNA. Fig 5A shows that the crosslink of Prp46 was preserved with a similar yield in the B* complex but it increased by ~20% in the C complex (S6B Fig). This indicates that Prp46 remains in contact with the 3’ end of the intron during remodeling of the Bact to B* and to C complexes. To map more precisely the RNA interaction site of Prp46, we performed UV crosslinking of Bact spliceosomes assembled on site-specifically labeled pre-mRNAs in Prp46-TAP extract (Fig 5B). After pull-down, we observed that Prp46-TAP crosslinked weakly to both RNA fragments labeled at G511 and G516 (Fig 5B, lanes 7 and 8). Despite the strong crosslink of Prp46 observed in 2D gels obtained from the Bact complex (Fig 5A), we detected low levels of Prp46 crosslinks in the Bact complex assembled on each of the two site-specifically labeled pre-mRNAs (lanes 7 and 8). Taken together, these results suggest that the prominent crosslink of Prp46 in the 2D gel may be due either (i) to an additional crosslinked protein co-migrating with Prp46 or (ii) to interaction with a region located further upstream than the region 479–482. Alternatively, or additionally, the TAP-tag fused to Prp46 may prevent efficient crosslinking of Prp46 to the intron (Fig 5B).
Although Prp45 did not crosslink in Bact complexes assembled on the 3’-region-labeled pre-mRNA after UV irradiation (Figs 1C and 5A), we nonetheless observed a protein with the expected size of Prp45 (i.e. ~42kDa), which crosslinked in low yield to the fragment 483–496, in C complexes (Figs 3A, lane 18 marked by a dot, and S7). The identity of Prp45 was determined by pull-down of crosslinked and T1-digested complexes containing Prp45-TAP, assembled on pre-mRNA site-specifically labeled at G496 (Fig 5C). Prp45-TAP crosslinked (with low intensity) only in the C complex (lane 6), indicating that Prp45 makes contact with the region of the intron 483–496 after step 1 catalysis. Independent evidence that Prp45 contacts the pre-mRNA upon step 1 catalysis was obtained again from the analysis of 2D gels obtained from crosslinked B* and C complexes that were assembled on the 3’-region-labeled pre-mRNA (Fig 5A). Prp45 crosslinked in the B* complex at low levels, yet the intensity of this crosslinked species increased more than 80% in the C complex (Figs 5A, S6B and S7). This result suggests a temporal interaction of Prp45 with the intron's region near the 3' SS upstream of Prp46, and indicates that Prp45 contacts the 3’ end of the intron after/during step 1 catalysis.
Here we have investigated pre-mRNA–protein contact sites in affinity-purified yeast Bact spliceosomes by UV crosslinking. A number of crosslinked proteins of the U2 snRNP, including the SF3a subunits Prp9 and Prp11 and the SF3b proteins Cus1, Hsh49 and Hsh155, as well as RES complex proteins and their contact sites on the pre-mRNA intron, could be precisely assigned by performing crosslinking followed by 2D gel electrophoresis and immunoprecipitation. Taken together, the results indicate that the branch-site region is contacted at several positions, apparently over its entire length, by proteins. A similar investigation was carried out with affinity-purified spliceosomal B* and C complexes. The results presented here provide much-needed information regarding the spliceosomal pre-mRNA–protein network, and they show for the first time that also yeast U2 SF3a/b proteins, as their human counterpart, are tightly anchored around the BS region. They also provide insight into the dynamics of pre-mRNA–protein interactions involving Cwc25, Prp8 and Prp45 within the spliceosome upon its conversion into the B* (i.e., catalytically activated) complex and during the subsequent conversion of the latter into the C complex (i.e., the step 1 spliceosome).
In the human system, the U2 protein-pre-mRNA interactions are already established in the early A complex but remain in the rearranged, activated spliceosome [15,34]. Here, we analyzed U2 protein–pre-mRNA interactions initially in purified Bact complexes, likely our data obtained with Bact complexes apply also to earlier complexes (i.e. A and B complexes), which for practical reasons were not analyzed here.
Using a combination of UV crosslinking and immunoprecipitation of TAP-tagged proteins, we were able to assign a number of U2 snRNP proteins crosslinked to specific sites using pre-mRNAs that were labeled at specific positions by a combination of DNA enzymes cleavage and splint-directed ligation [22]. Site-specific labeling of the RNA with 32P is a much more promising approach for UV crosslinking studies, because the RNA–protein interaction site can be precisely mapped on the RNA. For the first time, we were able to use purified yeast spliceosomes to perform a comprehensive protein–pre-mRNA interaction analysis and thus to assign a well-defined RNA region crosslinked to a known protein. In this way we were able to map an extensive area, spanning a 70-nt-long region of the intron.
Consistent with previous studies with human spliceosomal complexes, crosslinking sites involving the yeast U2 SF3a proteins Prp9 and Prp11, as well as SF3b proteins Cus1 and Hsh49, were observed within a 14-nt-long region upstream of the BS of affinity-purified Bact complexes (Fig 2B). Furthermore, consistent with previous results obtained in yeast [19] and human [14] spliceosomes, immunoprecipitation revealed contacts between a region located further downstream (surrounding the BS) and SF3b Hsh155. These results provide evidence that the region directly upstream of the BS, and surrounding it, is the main interaction platform of the yeast U2 snRNP proteins. Likewise all human U2 snRNP-associated SAPs, except for SAP130, were found in direct contact with a 20-nt-long region upstream of the BS in the isolated spliceosomal complexes A, B, and C [34,35] and SF3b155 was also found to bind to a site downstream of the BS [14]. Thus, our data furthermore suggest that U2 protein-pre-mRNA interactions with the regions upstream and downstream of the BS are conserved between yeast and human (S1 Table). Furthermore, consistent with earlier findings [15], an oligoribonucleotide complementary to the 14-nt-long region upstream of the BS inhibits formation of the yeast A, B and Bact complexes (S8 Fig). Thus, interactions of SF3a and SF3b with the pre-mRNA appear to be a prerequisite for pre-spliceosome formation also in yeast, indicating that the perfect complementarity between the BS sequence and the U2 snRNA is not sufficient to anchor the U2 snRNP to the BS sequence, and that stable binding of U2 is largely dependent on U2 protein–pre-mRNA interactions.
The RES complex is a conserved, spliceosome-associated module that has been shown to enhance splicing of a subset of transcripts and to promote the nuclear retention of unspliced pre-mRNAs in yeast [10]. Furthermore, it was shown to be required for efficient splicing of TAN1 pre-mRNA, and the intron sequence between the 5'SS and the BS was necessary and sufficient to mediate dependence upon RES [36]. Here, we identified low-yield crosslinks between the BS and the 3’SS of both Pml1 and Bud13. Consistent with results from our earlier studies [20], we show that the other RES complex protein, Snu17, is in direct contact with the pre-mRNA in the region between the BS and the 3’SS within a 14-nt-long RNA stretch upstream of the G nucleotide at position 496. Our data do not conflict with the result observed with the TAN1 pre-mRNA because the requirement of TAN1 intron nts upstream of the BS for RES dependence could be transient, and an interaction may occur earlier during spliceosome assembly.
Our observation of direct contact between Snu17 and the intron is in agreement with earlier reports showing that Snu17 consists primarily of a RRM, which is probably involved in contacting the RNA. It was also reported that the RRM of Snu17 is atypical and acts as a central binding platform that provides two separate interaction surfaces, which interact with disordered parts of Bud13 and Pml1 at the same time [37–39]. While Bud13 and Pml1 do not harbor typical RNA-binding domains, Bud13 contains a conserved lysine-rich region that might bind RNA. In addition, Pml1p or U2 proteins interacting with RES in the spliceosome (such as the SF3b Hsh155) might facilitate the recognition of RNA by RES. A recent NMR solution structure of the core of the RES complex revealed that complex formation leads to intricate folding of the three components that stabilize the RRM fold of Snu17 upon binding of Bud13 and Pml1, while RNA binding efficiency is increased [20]. Taken together, our results indicate that Snu17 crosslinks directly to the intron between the BS and the 3’SS in the Bact complex, while Pml1 and Bud13 may make contact with the intron through their elaborated interconnection with Snu17, but they may be in a conformation that does not favor the formation of UV-induced crosslinks (see Fig 6 for a summary and S2 Table).
As Hsh155 is in contact with nucleotides of the pre-mRNA between the BS and 3’SS (Fig 3), which are also in contact with all three components of RES, this indicates that Hsh155 and the RES proteins are in close proximity to one another in the Bact spliceosome. This is consistent with earlier studies showing by a yeast two-hybrid screen and co-immunoprecipitation experiments that Snu17 interacts with U2 SF3b proteins [18]. Furthermore, the RES complex subunit Snu17 was shown to bind to the U2 snRNP [41]. Taken together, all these studies indicate that there is a direct interconnection between RES, the U2 SF3b proteins and the pre-mRNA downstream of the BS.
Examination of U2 protein–pre-mRNA interactions in purified spliceosomal complexes stalled after catalytic activation by Prp2/Spp2 (B* complex) and subsequent step 1 catalysis by Cwc25 (to form the C complex), revealed that the spliceosome structure involving the region of the intron upstream of the BS and the SF3 proteins Prp9, Hsh49 and Cus1 undergoes a conformational change during spliceosome activation and subsequent step 1 catalysis. That is, crosslinks of Prp9, Hsh49 and Cus1 were significantly reduced in both spliceosomal complexes compared with those observed with the Bact complex, indicating that binding to pre-mRNA of these proteins is destabilized after ATP hydrolysis by Prp2. This is consistent with the remodeling of the structure of the catalytic core of the spliceosome near the BS upon nucleophile attack on the 5' SS phosphodiester bond during step 1 catalysis. That is, alterations in U2 protein binding are probably due to conformational changes that destabilize the interactions of these proteins with the pre-mRNA upstream of the BS concomitant with step 1. Intriguingly, our data reveal that contacts of Hsh49, Cus1 and Prp9 (and to a lesser extent Prp11) with two adjacent short regions of the intron upstream of the BS were reduced, to different degrees, during the remodeling of the spliceosome. This indicates that SF3 proteins remain in contact with the intron upstream of the BS even after step 1 catalysis, yet their binding affinity to the pre-mRNA is significantly reduced at a certain site and partially abolished at another. Our results further indicate that the complete set of U2 proteins remains in contact with the U2 snRNA via protein–RNA or protein–protein interaction. Indeed, it was shown that the U2 snRNP is released from the intron-lariat spliceosome in vitro as an integral snRNP, indicating that it remains intact during the entire splicing cycle and that none of its proteins are lost under physiological conditions in vitro [24].
Furthermore, as suggested by their decreased efficiency of crosslinking during spliceosome remodeling (Fig 3), the binding of the RES complex as a whole is reduced. Remodeling events involving the RES complex proteins are intriguing because the same stretch of the intron is also bound by Prp2 and is essential for Prp2- and Spp2-mediated catalytic activation [8,42]. Indeed, Prp2 was crosslinked to the same nucleotides of the intron as the RES proteins [8]; thus, it may be possible that Prp2 recognizes this stretch of RNA "productively" only when it is in contact with the RES proteins. The RES complex could be recognized as an entry point or primary target by Prp2/Spp2 to initiate translocations along the intron [42], thereby destabilizing RNA-bound proteins and acting as a classical RNPase. Alternatively, it was recently suggested that Prp2, in addition to binding the intron, is probably involved in several protein–protein interactions in the spliceosome [8]. This would lead to the formation of a relay system that could transmit a power stroke within the motor module of Prp2/ATP, through the various anchor points that Prp2 shares with other components of the spliceosome [8]. Thus, the RES–the binding of which is destabilized upon Prp2/Spp2-mediated B* formation [11] (and this work)–could be an important primary element of this communication system. Importantly, it was recently reported that a prp2 mutant was suppressed by deletion of PML1, indicating that Pml1 stabilizes an interaction that Prp2 destabilizes [43].
To date, Prp8 is the only spliceosomal protein that has been shown to crosslink to all the three regions in pre-mRNA that are required for splicing (5’SS, 3’SS, and BS), as well as to U5 and U6 snRNAs [44,45]. Here, by 2D gel electrophoresis of affinity-purified Bact complexes, we provide evidence that Prp8 –although already stably associated with Bact–makes no close contacts with the 3’ region of the intron before catalytic activation by Prp2/Spp2. Using affinity-purified spliceosomes stalled at the B* and C stages, we show that conformational changes leading to step 1 catalysis bring Prp8 to a position near the BS and the 3’SS (Figs 3 and S7) [28–31]. That is, remodeling at the catalytic core of the spliceosome accompanies stabilization of Prp8–pre-mRNA contacts. Intriguingly, previous work showed that high-affinity binding sites are created in the B* complex–also for additional factors required for step 1 catalysis such as Yju2 and Cwc25 –during catalytic activation [11]. Thus, the ATP-dependent Prp2-driven activation of the spliceosome leads not only to reduced contacts with the pre-mRNA of U2 and RES proteins, but also to stabilization of other protein–pre-mRNA interactions by promoting direct contact with the pre-mRNA. Taken together, these results provide new insight into the dynamics of protein–pre-mRNA interactions (simultaneous reduction of some and enhancement of others) within the spliceosome during its catalytic activation and catalysis.
Step 1 catalysis cannot occur efficiently without Cwc25 [7,9]. After Prp2-mediated catalytic activation of the spliceosome, a strong binding site is created on the B* spliceosome for the step 1 factor Cwc25. While Cwc25 only shows background binding to complex Bact, its binding to complex B* has a Kd value in the subnanomolar range [11]. Consistent with the enhanced binding of Cwc25 upon the action of Prp2, we show here that Cwc25 crosslinks to the 11-nt-long BS fragment. These data are in agreement with earlier reports showing that Cwc25 crosslinks to the intron sequence three bases downstream of the BS [32]. Interestingly, using truncated versions of recombinant Cwc25 for reconstitution of C complexes, we observed that Cwc25 1–125 (lacking 54 amino acids at its C-terminus) promoted step 1 catalysis even in the absence of RNA crosslinking, indicating that Cwc25’s step-1-promoting activity is not coupled to its pre-mRNA interacting activity. This indicates that contacts of Cwc25 1–125 would theoretically occur with one or more proteins in the proximity of the BS, thus making the microenvironment of the catalytic center suitable for step 1 catalysis. This would be consistent with Cwc25 being one of the intrinsically disordered proteins, which are highly connected or “promiscuous” proteins that undergo several simultaneous or sequential interactions and use regions of disorder as a scaffold for assembling an interacting group of proteins [46]. Thus, Cwc25 might act as an important hub in the catalytic center of the spliceosome. Indeed, we observed Cwc25’s contacts in the BS region of C complexes concomitant with enhanced crosslinking of Prp8 and Prp45 to the same or a slightly downstream region (Figs 3–5), suggesting that Cwc25 co-ordinates the catalytic center through protein–protein interaction.
We showed that Prp45 contacts the pre-mRNA only after step 1 catalysis (Figs 5C and S7), although it is already associated with the spliceosome at the Bact stage (Fig 1B). Earlier results showed that, in addition to their interaction in two-hybrid screens, Prp45 and Prp46 interact in vitro, most probably through direct protein–protein contact [33]. Here, Prp45 crosslinked to the pre-mRNA region of the intron 483–496 and Prp46 to the region immediately downstream (i.e. 500–516), indicating that the two proteins are also in close contact during spliceosome remodeling. Interestingly, Prp45 crosslinked during or after step 1 catalysis to the same pre-mRNA region of the intron (i.e. 483–496) where Prp8 was also found to crosslink with highest yield (S7 Fig), indicating that profound remodeling events involving this region occur. Indeed, a simultaneous reduction of Snu17 and Pml1 contacts was also observed (Figs 6 and S7 for a summary).
The contact of Prp45 to this region is consistent with earlier work that showed that Prp45 interacts with Prp22, a DEAH-box RNA helicase involved in spliceosome disassembly [33,47], which was also crosslinked to the 3’SS in post-step 1 spliceosomes [48]. In addition, a temperature-sensitive allele of Prp45 was shown to be synthetically lethal with alleles of several second-step splicing factors (Slu7, Prp17, Prp18 and Prp22) and with several NTC components. Thus, Prp45 may be required for Prp22 as well as the recruitment or stabilization of additional step 2 factors, and the positioning of Prp45 close to the 3’SS after step 1 may determine the timing of this event. The timing of the direct interaction of Prp45 with the pre-mRNA may explain the contribution of this protein to step 2 catalysis: this could be effected either (i) by participating in maintaining the step 2 conformation, or (ii) by binding and regulating the Prp22 ATPase/translocase activity [47].
In a first step, actin pre-mRNA was prepared by transcription in vitro with T7 RNA polymerase (S1 Text). The transcription reaction was not gel-purified, but instead was precipitated with ethanol. After washing twice with 70% ethanol, the precipitated RNA was dried, dissolved in 50μl CE buffer (10 mM cacodylic acid-KOH, pH 7.0, 0.2 mM EDTA-KOH, pH 8) or water and applied to a G 50 spin column (GE Healthcare). The eluted RNA was then subjected to DNA enzyme cleavage, essentially as described previously [25]. First, a threefold molar excess of the DNA enzyme over the pre-mRNA was added to the reaction mixture. The solution was then adjusted to 15 mM NaCl and 5 mM TRIS-HCl, pH 7.7. After denaturation at 70°C for 2 min the mixture was kept at room temperature for 5 min. Finally, 150 mM NaCl, 50 mM Tris-HCl, pH 7.7 and 2 mM of both MgCl2 and MnCl2 were added and the mixture was incubated at 30°C for 3 hrs. To remove the cyclic phosphate produced at the 3’ end of the 5’ fragment by the DNA enzyme, the intrinsic 3'-phosphatase activity at low ATP concentration of T4 polynucleotide kinase (T4 PNK) was used [26]. The reaction was supplemented with 2 unites/μl T4 PNK, PNK buffer and 0.4 mM ATP and incubated for 1 h at 37°C [26]. The RNA digestion fragments were gel-purified as described for in vitro transcriptions (S1 Text).
For the production of site-specifically labeled pre-mRNAs, the purified 3’ pre-mRNA fragment created by DNA enzyme cleavage was 5’-phosphorylated with 2 μM γ-32P ATP, T4-PNK buffer, 2 units/μl RNAsin and 1 unit/μl T4-PNK in a total volume of 20 μl or more, depending on the experiment. The reaction mixture was incubated for 1 h at 37°C and then purified by using a G 50 spin column, followed by phenol-chloroform-isoamyl alcohol (PCI) extraction and ethanol precipitation.
RNA fragments were ligated by DNA splint-directed RNA ligation [22]. The 5’-ligation fragments were prepared by DNA enzyme cleavage followed by 3’-dephosphorylation as described above. For site-specific labeling, the 3’ ligation fragment was labeled at the 5’ end with γ-32P ATP as described above. For region-specific labeling, the 3’ fragment was produced by radioactive in vitro transcription using GMP as a starting nucleotide (S1 Text). The 5’ ligation fragment, the DNA splint and the 3’ ligation fragment were mixed in a 1.4:1.2:1 ratio. After addition of T4 DNA ligase buffer and water, the reaction was denatured for 2 min at 70°C and the sample was then cooled to 30°C at 6°C per min. Thereafter 1 mM ATP, 2 units/μl RNAsin and 3 units/μl T4 DNA ligase were added and the reaction was incubated for 3 hrs at 30°C. Finally, the ligation product was gel-purified. The efficiency of ligation was ~30–60%.
|
10.1371/journal.pcbi.1005444 | redGEM: Systematic reduction and analysis of genome-scale metabolic reconstructions for development of consistent core metabolic models | Genome-scale metabolic reconstructions have proven to be valuable resources in enhancing our understanding of metabolic networks as they encapsulate all known metabolic capabilities of the organisms from genes to proteins to their functions. However the complexity of these large metabolic networks often hinders their utility in various practical applications. Although reduced models are commonly used for modeling and in integrating experimental data, they are often inconsistent across different studies and laboratories due to different criteria and detail, which can compromise transferability of the findings and also integration of experimental data from different groups. In this study, we have developed a systematic semi-automatic approach to reduce genome-scale models into core models in a consistent and logical manner focusing on the central metabolism or subsystems of interest. The method minimizes the loss of information using an approach that combines graph-based search and optimization methods. The resulting core models are shown to be able to capture key properties of the genome-scale models and preserve consistency in terms of biomass and by-product yields, flux and concentration variability and gene essentiality. The development of these “consistently-reduced” models will help to clarify and facilitate integration of different experimental data to draw new understanding that can be directly extendable to genome-scale models.
| Reduced models are used commonly to understand the metabolism of organisms and to integrate experimental data for many different studies such as physiology, fluxomics and metabolomics. Without consistent or clear criteria on how these reduced models are actually developed, it is difficult to ensure that they reflect the detailed knowledge that is kept in genome scale metabolic network models (GEMs). The redGEM algorithm presented here allows us to systematically develop consistently reduced metabolic models from their genome-scale counterparts. We applied redGEM for the construction of a core model for E. coli central carbon metabolism. We constructed the core model irJO1366 based on the latest genome-scale E. coli metabolic reconstruction (iJO1366). irJO1366 contains the central carbon pathways and other immediate pathways that must be connected to them for consistency with the iJO1366. irJO1366 can be used to understand metabolism of the organism and also to provide guidance for metabolic engineering purposes. The algorithm is also designed to be modular so that heterologous reactions or pathways can be appended to the core model akin to a “plug-and-play”, synthetic biology approach. The algorithm is applicable to any compartmentalized or non-compartmentalized GEM.
| Stoichiometric models have been used to study the physiology of organisms since 1980s [1–3], and with the accumulation of knowledge, and progressing techniques for genome annotation, these models have evolved into Genome Scale Metabolic Reconstructions (GEMs), which encapsulate all known biochemistry that takes place in the organisms by gene to protein to reaction (GPRs) associations [4]. Since the first Genome Scale models developed [5,6], the number of annotated genomes and the corresponding genome scale metabolic reconstruction has increased tremendously [7–9].
With increasing popularity of GEMs, different techniques to analyse these networks have been proposed [10,11]. Flux Balance Analysis (FBA), a constraint-based method (CBM) enables many forms of analysis based solely on knowledge of network stoichiometry and incorporation of various constraints, such as environmental, physicochemical constraints [12]. FBA has been further expanded by other methods such as Thermodynamics-based Flux Analysis (TFA) [13–16] and others [17,18] for the integration of available thermodynamics data with GEMs. TFA utilizes information about the properties of reaction thermodynamics and integrates them into FBA. Such properties now can be estimated by Group Contribution Method [19–21] and high-level Quantum Chemical Calculations[22]. Metabolic networks are valuable scaffolds that can also be used to integrate other types of data such as metabolic [23,24], regulatory and signalling [25–27], that can elucidate the actual state of the metabolic network in vivo. However, both FBA, TFA and other steady-state approaches cannot capture the dynamic response of metabolic networks, which requires integration of detailed enzyme kinetics and regulations [28]. Hatzimanikatis and colleagues have developed a framework that utilizes FBA, TFA and generates kinetic models without sacrificing stoichiometric, thermodynamic and physiological constraints [29–31]. Recently, another approach has been proposed to integrate kinetics into large-scale metabolic networks[32].
As the quality and the size of the models increase with better annotation, the complexity of the mathematical representations of the models also increases. Hatzimanikatis and colleagues [33] observed that majority of the studies and applications using metabolic models have still revolved around the central metabolism and around “reduced” models instead of genome-scale models, indicating that the full potential of GEMs remains largely untapped [34–38]. These reduced models have the advantage of having small sizes as they are built with a top-down manner, but they lack the quality of bottom-up built models since they have been reduced ad hoc, with different criteria and aims, which have not been consistently and explicitly justified [39–41]. An algorithmic approach called NetworkReducer [42] has been recently proposed following a top-down reduction procedure. The main purpose of this approach is to preserve selected so-called “protected” metabolites and reactions, while iteratively deleting the reactions that do not prevent the activity of the selected reactions. This algorithm has been further extended [43] to compute the minimum size of subnetworks that still preserve the selected reactions.
In this study, we have developed redGEM, a systematic model reduction framework for constructing core metabolic models from GEMs. Herewith, we propose an approach that focuses on selected metabolic subsystems and yet retains the linkages and knowledge captured in genome-scale reconstructions. redGEM follows a bottom-up approach that allows us to handle the complexity and to yield comprehensive insights in connecting the metabolic model to actual cellular physiology. redGEM can be tailored to generate minimal models with conserved functions. However, our approach is not strictly focused only on the reduction of the stoichiometry for the generation of highly condensed network, but aims also to preserve the constitutive characteristics of metabolic networks.
In redGEM, we use as inputs: (i) a GEM, (ii) metabolic subsystems that are of interest for a physiology under study; (iii) information about utilized substrates and medium components; and (iv) available physiological data (Fig 1). After a series of computational procedures, we generate a reduced model that is consistent with the original GEM in terms of flux profiles, essential genes and reactions, thermodynamically feasible ranges of metabolites and ranges of Gibbs free energy of reactions. We applied redGEM on the latest GEM of E. coli iJO1366 [44] under both aerobic and anaerobic conditions with glucose and other possible carbon sources and generated a family of reduced E. coli iJO1366 models.
We performed the redGEM algorithm on the latest GEM of E. coli, iJO1366 to generate a reduced model consistent with its parent GEM model. Firstly, we selected 6 central carbon metabolism subsystems (glycolysis, pentose phosphate pathway, citric acid cycle, glyoxylate cycle, pyruvate metabolism, and oxidative phosphorylation), as they are defined in original E. coli GEM. In addition, we have included all the reactions that use quinone/quinol pool metabolites (Ubiquinone/ubiquinol, menaquinone/menaquinol, 2- dimethyl menaquinone/2- dimethyl menaquinol for E. coli) in oxidative phosphorylation subsystem to capture the coupling between the core carbon metabolism and energy/redox metabolism. Some of those reactions had different subsystem definition in original GEM. These subsystems include a total of 185 reactions and 126 metabolites. We next redefined the content of each starting subsystem by performing an intra-expansion analysis to identify the RT (See Material and Methods for definitions) reactions. We include a reaction in RT when it only interconverts metabolites that are already included in one subsystem, and these reactions belong to a different subsystem in original GEM. This analysis established that there are many reactions in GEM whose reactants and products belong to a specific subsystem but are assigned to a different subsystem in the original GEM (Table 1). Some of the reactions defined in RT are common between subsystems, since the subsystems share many metabolites, especially cofactor pairs such as ATP/ADP, NAD+/NADH etc.
After the intra-expansion, the network expansion by directed graph search finds metabolites and reactions between subsystems in a pairwise manner for non-common metabolites (postulate 3 in Material and Methods) with respect to the degree of connection D. D is the distance between a subsystem pair and can be either equal to the inherent minimum distance between each pair, or imposed by the user for all subsystem pairs. Depending on the network topology, the inherent minimum distance can be equal to the input D imposed by the user (postulate 5 in Material and Methods). redGEM also performs pairwise connections between the metabolites of the same subsystem. The algorithm calculates MS, MijD and MiiD (all pairs i, j), RSi, RijD, RiiD (all pairs i, j), which overall define the core network CND with respect to selected degree of connection parameter D (Table 2). The additional reactions for every degree of connection D are specific for the corresponding D (postulate 2 in Material and Methods). As a final step, redGEM performs an additional intra-expansion, and scans through every reaction in GEM to identify the reactions RT, which are not captured by RSi, RijD, RiiD (all pairs i, j) but include only MS, MijD and MiiD (postulate 4 in Material and Methods). This procedure finalizes the steps that define the final core network for further analysis for redGEM. We performed redGEM on E. coli iJO1366 and we generated all core networks with degree of connection up to D = 6.
At D = 1, redGEM captured many connecting reactions that are part of many ad hoc built models, such as malic enzymes 1–2 between glycolysis and TCA cycle that connect L-malate to pyruvate, phosphoenolpyruvate carboxylase and phosphoenolpyruvate carboxykinase that connect oxaloacetate and phosphoenolpyruvate. Moreover, it captures many other reactions, such as 2 types of L-aspartate oxidases, which are using quinone/quinol cofactor pairs and labeled as electron transport chains reactions. There are two more L-aspartate oxidase reactions that are added to the D = 1 core network by redGEM (S1 Table). One uses O2/H2O2 and the other one is using fumarate/succinate as cofactor pairs. These reactions are captured by RiiD and RT simultaneously. Finally redGEM added 10 reactions whose reactants and products are only cofactors and small metabolites belonging to D = 1 core network in their stoichiometry, such as NAD+ kinase, NADP phosphatase, adenylate kinase, nucleoside-triphosphatase etc. as a part of RT. Along these reactions, the non-growth associated ATP maintenance (ATPM) reaction is explicitly included in the reduced model, and its corresponding minimum requirement of the GEM is preserved for further analysis in this study.
When we analyze the pairwise connections between subsystems with respect to different connection parameter D, we observe that there is no D = 1 connection between certain pairs, such as pentose phosphate pathway (PPP) and glyoxylate metabolism (GLX) (Fig 2). However, zero connection between two subsystems by D = 1 does not necessarily mean that these subsystems are far from each other, as we observe that there are 5 and 15 reactions that are connecting PPP and GLX in 2 and 3 steps, respectively. As another extreme, tricarboxylic acid (TCA) cycle and electron transport chains (ETC) have 15 different reactions that connect each other with 1 reaction, demonstrating the strong connection between TCA cycle and redox metabolism.
Following the analysis for reactions, we identified the metabolites that connect the subsystems in a pairwise manner. There are no such intermediate metabolites between subsystems connected by D = 1, since this degree of connection only captures reactions between the unshared metabolites of a subsystem pair (Table 3). When the subsystems are connected pairwise with D = 2, there are 21 metabolites that become intermediates between all subsystem pairs. This number increases to 51 when degree of connection is increased to 3. By definition, a metabolite that connects a subsystem pair in 2 steps can also connect them in 3 steps through different reactions.
There are metabolites, such as pyruvate and succinate, that already participate in D = 0 reactions (in the initial starting subsystems), and they appear later to connect at least one subsystem pair with D = 3 connection. This indicates that there is no path in GEM with length less than 3 that can connect any starting subsystem pair using these intermediates, excluding the reactions that already belong to this subsystem pair.
Methylglyoxal is known to be a hub metabolite, since it can connect dihydroxyacetone phosphate to lactate in 2 reactions. Lactate is a metabolite that belongs to different starting D = 0 subsystems such as oxidative phosphorylation and pyruvate metabolism. Moreover, it can be converted to pyruvate by lactate dehydrogenase, and pyruvate is already known as a hub metabolite that can connect different subsystems. As another example, L and D tartrate connect pentose phosphate pathway and citrate cycle in 3 steps through the following path: With an antiporter, cytosolic succinate transports L and D forms of tartrate to cytosol. Then, L and D-tartrate dehydratase enzymes convert these two forms of tartrate to oxaloacetate and water. Following this biotransformation, oxaloacetate can be converted to pyruvate by many enzymes. As we observed in methylglyoxal case, pyruvate is part of many different starting D = 0 subsystems (glycolysis/gluconeogenesis, oxidative phosphorylation, citrate cycle, pyruvate metabolism and extracellular subsystem), and L and D tartrate appear as intermediates that connect 7 pairs of subsystems in D = 3.
Another layer of information that we can extract through this analysis is the subsystems that connect the selected starting subsystems, thus demonstrating the proximity of these subsystems to the defined starting core carbon ones. By starting from 7 subsystems (including extracellular metabolites as extracellular subsystem), the network expansion procedure results in capturing reactions as core from 32 different subsystems for D = 6 (Table 4). In GEM, there are 37 subsystems, which signifies that only 6 steps expansion captures reactions from ~90% of all subsystems defined in GEM, thus showing the tight connections between metabolites/subsystems in the network. For 2 subsystems defined in GEM, anaplerotic reactions and methylglyoxal metabolism, more than half of the all reactions within these subsystems are captured by network expansion procedure with connection parameter D being up to 3 (Table 5). An important observation is that components of the same subsystems can be parts of the connection of more than 1 subsystem pairs, since different subsystems can share the same metabolites.
The wild-type biomass reaction of the iJO1366 model contains 102 biomass building blocks (BBBs). The size and the complexity of the composition makes it necessary to develop techniques to keep the information stored in GEM for the biosynthesis, but yet reduce the size of the network significantly. Methods, such as graph-search algorithms can be used for identification of biosynthetic routes between two metabolites in metabolic networks [45,46]. However, these graph theory based approaches cannot be used for our purposes due to two main issues and limitations: i) they do not make use nor obey mass conservation; hence the pathways they generate are not guaranteed to be able to carry flux in metabolic network or to be elementally balanced, ii) and they cannot manage pathways that are not linear, such as branched pathways. To overcome these limitations, we used lumpGEM [47], an in-built tool, which identifies subnetworks that can produce biomass building blocks starting from precursor metabolites. These precursors are provided by redGEM through the systematically generated core network based on degree of connection parameter, D. Each subnetwork is then transformed into a lumped reaction and inserted in the reduced model. lumpGEM forces mass conservation constraints during optimization to identify subnetworks, thus preventing the generation of lumped reactions, which cannot carry flux in the metabolic networks. As an example, for D = 1, by minimizing the number of non-core reactions In GEM, lumpGEM generated a 17 reactions subnetwork to synthesize histidine from core carbon metabolites (Fig 3). Histidine is synthesized from ribose-5-phosphate, a precursor from pentose phosphate pathway. The linear pathway from this core metabolite to histidine is composed of 10 steps. However, due to the mass balance constraint, two metabolites, 1-(5-Phosphoribosyl)-5-amino-4-imidazolecarboxamide and L-Glutamine cannot be balanced in a network that is composed of core reactions and the linear pathway from ribose-5-phophate to histidine. These metabolites are balanced in the network by other non-core reactions. Hence, the generated sets of reactions are not linear routes from precursor metabolites to biomass building blocks, but branched, balanced subnetworks (for formulation of lumpGEM, see Material and Methods).
Using lumpGEM, we replicated all the biosynthetic pathways in databases such as EcoCyc [48], either as a part of subnetworks or the exact pathway. In addition, we identified subnetworks that can qualify as alternative biosynthetic pathways. E. coli is well-known to be robust against deletions by having many duplicate genes and alternate pathways[49]. Some of these routes may not be active due to energetics or regulatory constraints but using lumpGEM we can map these possible alternate pathways completely and also derive different biosynthetic lumped reactions. The introduction of such lumped biosynthetic reactions simplifies the core models considerably and allows the use of these models in important computational analysis methods such as dynamic FBA [50] extreme pathway analysis [51,52] and elementary flux modes [53,54], as well as for TFA formulations and kinetic modeling.
For D = 1 core network, lumpGEM generated 1216 subnetworks and 254 unique lumped reactions for 79 biomass building blocks in total for aerobic and anaerobic case. The remaining BBBs of the total 102 can be produced within the D = 1 core network. For some biomass building blocks, it is possible that all the alternatives for Smin (the minimal subnetwork size) subnetworks generated under aerobic conditions are using molecular oxygen, thus cannot carry flux under anaerobic conditions. This necessitates the generation of lumped reactions without any oxygen in the media. For Smin, lumpGEM generated only 4 new lumped reactions for anaerobic case, for 3 metabolites, namely, heme O, lipoate (protein bound) and protoheme. All the other lumped reactions generated for anaerobic case are a subset of the 250 lumped reactions (S2 Table) for aerobic conditions. In the subsequent studies, we used all lumped reactions in order to allow for studies under different oxygen limitations without changing the model structure. The core model can be found in the supplementary material (S1 File).
Reduced models have been used to understand and investigate cellular physiology for many years. Before the emergence of genome scale models (GEMs), different groups with different aims built reduced models for their studies with a top-down approach. Conversely, GEMs provide the platform to understand all the metabolic capabilities of organisms, since GEMs encapsulate all the known biochemisty that occurs in cells. However the complexity of GEMs make their use impractical for different applications, such as kinetic modeling or elementary flux modes (EFMs). The need to focus on certain parts of these networks without sacrificing detailed stoichiometric information stored in GEMs makes it crucial to develop representative reduced models that can mimic the GEM characteristics. Within this scope, we developed redGEM, an algorithm that uses as inputs genome-scale metabolic model and defined metabolic subsystems, and it derives a set of reduced core metabolic models. These family of core models include all the fluxes across the subsystems of interest that are identified through network expansion, thus capturing the detailed stocihiometric information stored in their bottom-up built parent GEM model. Following the identification of the core, redGEM uses lumpGEM, an algortihm that captures the minimal sized subnetworks that are capable of producing target compounds from a set of defined core metabolites. lumpGEM expands these core networks to the biomass building blocks through elementally balanced lumped reactions. Moreover, redGEM employs lumpGEM to include alternative lumped reactions for the synthesis of biomass building blocks, thus accounting for alternative sytnhesis routes that can be active under different physiological conditions.
redGEM builds reduced models rGEMs that are consistent with their parent GEM model in terms of flux and concentration variability and essential genes/reactions. These reduced models can be used in many different areas, such as kinetic modeling, MFA studies, Elementary Flux Modes (EFM) and FBA/TFA. redGEM algorithm is applicable on any compartmentalized or non-compartmentalized genome scale model, since its procedure does not depend on any specific organism. As a demonstration, we have applied the redGEM algorithm on different organisms, namely P. putida, S. cerevisiae, Chinese Hamster Ovary cell (CHO) and human metabolism. For instance, redGEM algorithm has generated core networks of sizes between 168 metabolites/164 reactions to 360 metabolites/414 reactions for iMM904 [58] GEM reconstucted for S. cerevsiae with degree of connection parameter D varied from 1 to 6. The generated reduced model irMM904 with D = 1 has the same biomass yield with the parent model GEM as 0.29/hr under 10 mmol/gDWhr glucose uptake. Similar to E. coli case, flux and concentration variability, and gene essentiality characteristics of the rGEM are in agreement with the GEM counterparts (Ataman et al., manuscript in preparation). Moreover, reduced models are promising platforms for the comparison of central carbon (or any other) metabolism of different species. This approach can help us to better investigate the metabolic capabilities and limitations of organisms and to identify the sources of physiological differences across different species.
We applied redGEM algorithm on the latest genome scale model of E. coli iJO1366 [44], which is composed of 2251 enzymatic reactions (including transporters), 1136 unique metabolites across cytoplasm, periplasm and extracellular media. We used glucose as the sole carbon source and constrained the model for aerobic conditions.
In redGEM, we introduce and use the following definitions:
We can also generate the core network from the chosen subsystems using the minimum distance between the chosen subsystems and report the connecting reactions and metabolites. In this case, the degree of connection D is the minimum distance between Si and Sj.
redGEM uses the following inputs and parameters:
The central workflow of redGEM involves 4 steps:
The core carbon network is defined as all the reactions and metabolites in MS, MijD and MiiD (all i, j pairs), RSi,RijD,RiiD (all i, j pairs), RT (reactions that only cofactor pairs, small metabolites and inorganics participate).
We used the lumpGEM algorithm to generate pathways for all biomass building blocks (BBB) as they are defined in GEM. lumpGEM identifies the smallest subnetwork (Smin) that are stoichiometrically balanced and capable of synthesizing a biomass building block from defined core metabolites. Moreover, it identifies alternative subnetworks for the synthesis of the same biomass building block. Finally, lumpGEM generates overall lumped reactions, in where the cost of core metabolites, cofactors, small metabolites and inorganics are determined for the biosynthesis. redGEM defined the core network by the algorithm above, and then we generated all minimum sized subnetwork (Smin) for each BBB. Then lumpGEM calculated the unique lumped reactions for all the BBBs, and we used these lumped reactions for further validation and other analysis. lumpGEM takes the following steps to build elementally balanced lumped reactions for the biomass building blocks. In the workflow, lumpGEM
Maximize
∑i#ofRnCzrxn,i
such that:
S.v=0
vBBB,j≥nj,GEM.μmax
where,
To identify alternative Smin subnetworks for a BBB, lumpGEM further constrains the GEM with the following integer cuts constraint after generating each subnetwork with an iterative manner[59]. The reactions that belong to each subnetwork are denoted as RSmin
∑k#ofRSminzRSmin,k>0
We validate the consistency between rGEM and GEM performing the following consistency checks by comparing:
While these are the basic consistency tests, one could define additional checks, which can be specific to the organism and problem under study. We recommend that in all cases one should perform the checks using FBA and TFA, i.e. with and without thermodynamics constraints.
The first release of the redGEM toolbox is available upon request to the corresponding author.
|
10.1371/journal.ppat.0030073 | Crystal Structure of the P Pilus Rod Subunit PapA | P pili are important adhesive fibres involved in kidney infection by uropathogenic Escherichia coli strains. P pili are assembled by the conserved chaperone–usher pathway, which involves the PapD chaperone and the PapC usher. During pilus assembly, subunits are incorporated into the growing fiber via the donor–strand exchange (DSE) mechanism, whereby the chaperone's G1 β-strand that complements the incomplete immunoglobulin-fold of each subunit is displaced by the N-terminal extension (Nte) of an incoming subunit. P pili comprise a helical rod, a tip fibrillum, and an adhesin at the distal end. PapA is the rod subunit and is assembled into a superhelical right-handed structure. Here, we have solved the structure of a ternary complex of PapD bound to PapA through donor–strand complementation, itself bound to another PapA subunit through DSE. This structure provides insight into the structural basis of the DSE reaction involving this important pilus subunit. Using gel filtration chromatography and electron microscopy on a number of PapA Nte mutants, we establish that PapA differs in its mode of assembly compared with other Pap subunits, involving a much larger Nte that encompasses not only the DSE region of the Nte but also the region N-terminal to it.
| Bacterial adhesion to a host is a crucial step that determines the onset of bacterial infection. It is mediated through recognition of a receptor on the host cell surface by a protein called an adhesin displayed on the surface of the bacterium. Many adhesins are displayed at the tip of specialized organelles called pili, some of which are assembled by the ubiquitous chaperone–usher pathway. In this pathway, each pilus subunit is assisted in folding by a chaperone. The resulting chaperone–subunit complex is targeted to a pore located in the outer membrane, called the usher, that serves as assembly platform. There, pilus subunits dissociate from the chaperone and polymerize, resulting in a surface organelle, the pilus, that protrudes out of the usher. Here, we have determined the structure of the major subunit of the P pilus, PapA. The P pilus, produced in uropathogenic Escherichia coli, displays the adhesin PapG responsible for targeting the bacterium to the kidney epithelium. We have determined the structure of PapA either bound to its cognate chaperone, PapD, or bound to another PapA subunit. These structures provide a view of PapA before and after its assembly in the pilus and shed light on the mechanism of PapA assembly.
| Urinary tract infections, which include infections of the bladder (cystitis) and kidney (pyelonephritis), are some of the most common bacterial infections. These infections are caused mainly by uropathogenic Escherichia coli [1]. Once uropathogenic E. coli is introduced, survival and persistence of these bacteria in the urinary tract require a specific set of virulence factors, including the expression of type P pili. P pili are specifically required for the ability of uropathogenic E. coli to bind Gal-α (1–4)-Gal moieties in human kidney cells and cause pyelonephritis [2,3]. P pili are encoded by the pap gene cluster and are assembled via the highly conserved chaperone–usher pathway, involving the periplasmic immunoglobulin (Ig)–like chaperone PapD and an outer membrane usher PapC [4,5]. P pili consist of six subunits making up a composite fiber with a short tip fibrillum composed of the PapE subunit joined to a more rigid helical rod composed of the PapA subunit [6,7]. PapG is the adhesin at the end of a tip fibrillum; PapK and PapF are adaptor subunits between the PapA rod and the PapE fibrillum and between the PapE fibrillum and the PapG adhesin, respectively; finally, PapH terminates P pilus formation [8,9]. The PapA rod is formed by more than 1,000 PapA molecules assembled in a right-handed helical manner, with 3.3 molecules per turn [6,10].
All pilin subunits adopt an Ig-like fold but lack the seventh, C-terminal G β-strand, thus producing a large hydrophobic groove on the side of the protein (Figure 1 and [11,12]). In a process called donor–strand complementation (DSC), the G1 β-strand of PapD inserts a conserved motif of three alternating hydrophobic residues (called the P1 to P3 residues) plus N101 (P4 residue) into four binding pockets in the hydrophobic groove of the pilus subunits (P1 to P4 binding pockets). The G1 strand provides the structural information lacking in the pilus subunit by completing its Ig fold [11–13]. Pilus subunit assembly proceeds through a noncovalent polymerization process called donor–strand exchange (DSE; Figure 1). All subunits, except for the adhesin, possess an N-terminal extension (Nte) peptide of 11 (PapK), 12 (PapE and PapF), 19 (PapA), or 33 (PapH) residues (Figures 1 and S1) that is disordered and not part of their Ig-like structure. The Nte comprises a highly conserved array of alternating hydrophobic residues, called the P2 to P5 residues [14,15]. This array constitutes the DSE region of the Nte (see Figure S1 for location of this DSE region). As chaperone–subunit complexes are differentially targeted to the usher [16,17], each subunit donates its Nte to complete the Ig-fold of the subunit previously assembled by inserting its P2–P5 residues into the corresponding P2–P5 binding pockets, thus first displacing and then replacing the chaperone G1 strand in the groove of the previously assembled pilus subunit [18–21]. This process occurs through a zip-in–zip-out process whereby the DSE reaction is initiated by the insertion of the P5 residue of the Nte of one subunit into the P5 pocket of the groove of the other [21]. This binding event leads to the formation of a transient ternary complex, the formation of which is essential for the DSE reaction to proceed. Insertion of the Nte of PapH into the groove of a PapA subunit terminates pilus biogenesis because PapH lacks a P5 pocket and thus cannot provide the initiator-binding event required for the exchange reaction with another subunit [9].
Recently, in a departure from the more conventional model described above, Mu et al. [22] suggested, based on electron microscopy and image reconstruction of the PapA rod, that the DSE region of the Nte of PapA is not involved in DSE, but instead the region of the Nte N-terminal to the DSE region is involved in the process. Interestingly, the Nte of PapA is longer than the Nte of most Pap subunits except PapH (Figure S1), and thus the model proposed by Mu et al. [22] could potentially explain why residues N-terminal to the DSE region would be required in the process of P pilus biogenesis. This is explored further in this report, in which we describe the structures of a binary complex of PapD bound to PapA and of a ternary complex containing the chaperone PapD and two PapA subunits. In this ternary complex, PapD is bound to PapA through DSC, and this subunit is itself bound to another PapA subunit through DSE. These structures are used as a basis for a detailed mutational study dissecting the requirements for PapA polymer formation.
In order to investigate the PapD/PapA and PapA/PapA interactions within the P pilus rod, the first structure which we solved was that of PapD/PapA, where two mutations were introduced in the papA gene. This was necessary because PapA in purified wild-type PapD/PapA complexes tends to spontaneously polymerize, and self-polymerization of PapA has prevented structural studies. As PapA, like any other pilus subunits, polymerizes through DSE, mutating residues in the Nte was an obvious starting point to obtain a PapA mutant unable to polymerize. We first deleted the entire Nte (residues 2 to 19; PapANtd2; Figure S1), but coexpression of such mutants with PapD did not result in a material amenable to purification (unpublished data). Thus, a second mutant was designed that conserved the entire Nte but introduced an Asn at position 15. Position 15, a Gly residue in the wild-type Nte of PapA, locates in the DSE region of the Nte and is strictly conserved among all Pap Ntes (Figure S1). G15 is indeed required because, after DSE, it lies in the P4 pocket of the subunit's grooves, which, in this region and in all Pap subunits, contains a bulky phenylalanine or a tyrosine (F152 in PapA, Y146 in PapK, F138 in PapE, F137 in PapF, and Y162 in PapH) [11,18]. We thus mutated G15 to N; however, PapAG15N in complex with PapD also undergoes spontaneous polymerization (Figure S2A, top panel). Thus, in addition to the G15N mutation, a mutant where residues 2 to 8 were deleted was next constructed. The deleted region is just N-terminal to the DSE region of the Nte of PapA. This mutant (PapANtd1_G15N) did not undergo spontaneous polymerization, and formed a stable complex with PapD (Figure S2A, middle panel). This experiment indicates that, as suggested by the electron microscopy study of Mu et al. [22], the region of the Nte N-terminal to the DSE region is involved in PapA/PapA interaction. Crystals of PapD/PapANtd1_G15N diffracted to a resolution of 2.6 Å, and the structure was solved by molecular replacement using the PapD/PapK structure as a search model. This structure is very similar to the already known PapD/PapK, PapD/PapENtd, or PapD/PapHNtd1 structures [9,11,18]. Like PapK, PapE, or PapH, PapA lacks strand G of its Ig-fold; PapD complements this by donating its G1 strand.
By itself, the PapD/PapANtd1_G15N structure is not very informative. However, by investigating a different PapA mutant with the additional mutation T101L, we obtained crystals of a complex containing one PapD molecule and two PapANtd1_G15N_T101L molecules. This T101L mutant was initially designed to fill in the PapA P5 pocket. Indeed, as explained in the Introduction, all Pap subunits, except PapH, have a clear P5 pocket, which serves as an initiator point, the occupation of which triggers the DSE reaction [9,21]. Thus, the PapANtd1_G15N_T101L was made to test the possibility that by filling its P5 pocket, PapA would become more like PapH in being unable to undergo DSE. As shown in Figure S2B, indeed, PapANtd1_G15N_T101L has reduced DSE activity compared with that of PapANtd1_G15N. However, this T101L mutation had the additional unexpected consequence of stabilizing a complex containing a 1:2 molar ratio of PapD and PapA (Figure S2A, lower panel). PapD/(PapANtd1_G15N_T101L)2 was thus purified. Crystals were produced diffracting to 2.5 Å resolution. The structure of this ternary complex is shown in Figure 2. It clearly shows one PapA molecule bound to PapD through DSC in a complex very similar to PapD/PapANtd1_G15N; however, this time, the Nte of the donor–strand-complemented PapA molecule is bound to the groove of another PapA molecule, and thus this ternary complex crystal structure provides a snapshot of PapA before and after DSE. In that respect, the PapD/(PapANtd1_G15N_T101L)2 complex is similar to the one obtained by Zavialov et al. for the Caf system [19].
Figure 3A shows a superimposition of the two PapA molecules in the PapD/(PapANtd1_G15N_T101L)2 structure, with the donor–strand-complemented PapA subunit (dscPapA) in purple and the donor–strand-exchanged PapA subunit (dsePapA) in orange. The core sheet structure of dsePapA is in a closer conformation than that of dscPapA, as the β-strands on each side of the groove of dsePapA are nearer to each other. Also, the “63–74” loop is ordered in dsePapA and not in dscPapA, as this molecule is missing residues 70 to 73 in this region. The truncated Nte of dscPapA (as indicated above, residues 2 to 8 were removed to create PapANtd1) is clearly visible in the groove of dsePapA from residue 10 (the two first residues of PapANtd1_G15N_T101L, which are residues 1 and 9 of full-length PapA, were not defined in the electron density). DsePapA is only visible in the electron density from residue 20, as its Nte is not interacting in the groove of another PapA molecule and is thus disordered. Figure 3B shows the surface of dsePapA bound to the truncated Nte of dscPapA (left panel) and that of dscPapA bound to the G1 strand of the chaperone (right panel): the PapD G1 strand interacts as expected at the P1 to P4 positions in the groove of dscPapA (right panel), the Nte of which interacts in the P2 to P5 pockets in the groove of dsePapA (left panel). It is noticeable in the left panel of Figure 3B that the dsePapA groove is extending beyond the region occupied by the P2 residue of dscPapA Nte. However, the groove of dscPapA is not extending beyond the P1 pocket, due to the disordered “63–74” loop. Modeling of the first nine residues of PapA Nte in the extended groove of dsePapA shows that this extended groove has the right length and shape to accommodate the nine missing residues in the Nte of PapANtd1 (unpublished data). Thus, the groove of PapA is long enough to accommodate the extended Nte of another PapA molecule.
Figure 4A shows details of the grooves of dscPapANtd1_G15N (left panel) and dsePapANtd1_G15N_T101L (right panel). Figure 4A on the left panel shows that in the PapD/PapANtd1_G15N structure, the P4 position of the groove is formed by F152. This configuration of F152 is similar to that seen in the equivalent position of the PapD/PapK, PapD/PapE, PapD/PapH, and PapE/KNte complexes [9,11,18]. As mentioned above, the bulk created in the P4 pocket by F152 and equivalents, by being able to accommodate only a conserved Gly in the Ntes, acts as a registering device that calibrates the positioning of Ntes in the subunits' grooves.
In contrast to what is observed in the PapE/KNte or any other chaperone–subunit complex structures, the side chain of F152 has moved out of the P4 pocket of the DsePapA structure (Figure 4A, right panel) and the P4 pocket can now accommodate the N15 mutation (substituted for wild-type G15 in the PapANtd1_G15N_T101L mutant to prevent higher-order polymerization). F152 in the dsePapA structure is allowed to move away from the groove position because of the T101L mutation. Indeed, as shown in Figure 4B, the T101L mutation and the insertion of N15 induce a rearrangement of the side-chains in the P4–P5 region, leading to T99 moving out of the P5 region, thereby leaving room for F152 to substitute in its place, the new F152 position being stabilized by L101.
Like dsePapA in the PapD/(PapANtd1_G15N_T101L)2 complex structure, PapE in the PapE/KNte structure is also donor–strand exchanged (dsePapE), but with the Nte of PapK (KNte) [18]. Comparing dsePapE (Figure 5A, right) with dsePapA (Figure 5A, left) shows that, while the groove of dsePapE stops at the P2 pocket, the groove of dsePapA is extending beyond this pocket. This is due to the presence of the “63–74” loop in dsePapA, a loop which is much shorter in dsePapE, and to the closure of dsePapE groove by PapE N- and C-termini compared with the open groove in dsePapA (Figure 5B).
A number of PapA Nte mutants were next produced in order to evaluate the effect of these mutations on polymerization and pilus formation. In addition to PapA wild-type, six constructs were studied. In the DSE region, a G15N mutation and a Δ11–17 deletion (where the entire DSE region is removed) were made. In the region N-terminal to the DSE region of the Nte, an I4G single-site mutant and the Ntd1 deletion described above were studied. I4 is a bulky residue in that region of the Nte and thus would contribute significantly to the interface, were it to be involved in groove/Nte interaction (see [21] for consideration regarding surface area contribution of residues in the Nte). We also combined the Ntd1 (Δ2–8) deletion with the G15N mutation and the I4G and G15N mutations. Polymerization was assessed by gel filtration immediately after purification of the corresponding PapD/PapA complexes, and pilus formation was assessed after freeze–thaw of PapD/PapA complex preparations using electron microscopy (EM). Thus, gel filtration provides information on the limited polymerization events taking place early on during polymerization, while freeze–thaw of PapD/PapA complexes followed by analysis by EM provides information on the ability of the various PapA molecules to form pili. Results obtained for each of the PapA constructs are presented in Figure 6, where each panel provides the elution profile and EM micrograph for each wild-type and mutated PapA. For those complexes that exhibited a gel filtration peak corresponding to PapD/(PapA)2, further polymerization was checked by pooling the fractions corresponding to PapD/(PapA)2 and running another gel filtration the next day. For those complexes that did not exhibit polymerization, the PapD/(PapA)1 peak was rerun on gel filtration to make sure that indeed no polymers were formed.
Gel filtration of PapD/PapAwt complexes reveals a major 1:2 PapD/PapA (PapD/(PapA)2) complex. Yet, this complex polymerizes, as (1) rerunning a gel filtration on the 1:2 complex 24 hours later results in higher order polymers being formed (unpublished data), and (2) pili are formed readily (Figure 6A). These results confirm the existence of a rate-limiting step in the DSE reaction whereby PapD/(PapA)2 complex formation appears to be required before DSE can proceed to higher-order polymer forms [23]. The reason for such a rate-limiting step is unclear; it may be that this provides a checkpoint mechanism before committing to full biogenesis of the rod. The PapD/PapAI4G behaves very much like the PapD/PapAwt in the gel filtration, but fewer pili appear to be made (Figure 6B). The single G15N mutation in the DSE region appears to affect polymerization with the detection of 1:3 and 1:4 PapD/PapA complexes (PapD/(PapA)3 and PapD/(PapA)4), and while pilus formation does occur, the diameters of the pili and their central channels are increased (Figure 6C). Thus, these mutations appear to slow down the reaction in such a way that polymer intermediates are observed after 24 h, but overall, these mutations do not seem to affect the process so stringently that it is unable to proceed to completion. More important is the effect of deleting the region preceding the DSE region (PapD/PapANtd1; Figure 6D) or deleting the DSE region (PapD/PapAΔ11–17; Figure 6E). Neither of these mutants produces pili. Also, while PapD/PapANtd1 appears to be able to form 1:2 PapD/PapA complexes and aggregates of protein are visible by EM, the PapD/PapAΔ11–17 mutant is totally impaired. Thus, both regions (the DSE region and the region N-terminal to it) are important for PapA pilus formation, with the deletion of each of the regions blocking the process at two different stages of PapA polymerization. Combining the DSE (G15N) and non-DSE (I4G) mutations (Figure 6F), or combining the non-DSE deletion (Ntd1) and the single-site DSE (G15N) mutation (Figure 6G) lead to results that confirm our conclusion that both regions of the Nte are important for Pap polymerization and pilus formation. For the combined PapAI4G_G15N, large, linear aggregates are found, which are not pili. Higher-order limited polymers (PapD/(PapA)2, PapD/(PapA)3, and PapD/(PapA)4) are observed, showing that this double mutant is not totally impaired, while PapANtd1_G15N appears to be severely impaired.
In this report, we solved the structures of the PapA subunit, the major subunit of the P pilus, before and after DSE, and, based on these structures, we have examined the roles that the various regions in the Nte play in polymerization and pilus formation. We show that polymerization of wild-type PapA transitions through a 1:2 PapD/PapA complex (PapD/(PapA)2), and that a triple alteration combining a deletion of the residues preceding the DSE region of the Nte (Ntd1), a mutation of a conserved Gly residue in the DSE region (G15N), and a mutation in the P5 pocket (T101L) stabilizes the PapD/(PapA)2 intermediate. The Ntd1 and G15N mutations, individually, do not appear to block the formation of the PapD/(PapA)2 intermediate complex, nor do they block formation of higher-order complexes (Figure 6C and 6D). However, the combined Ntd1 and G15N mutations severely impair formation of these complexes (Figure 6G). Thus, the T101L mutation appears to attenuate the severity of the combined Ntd1 and G15N mutations and stabilizes the PapD/(PapA)2 intermediate. This may be because partial filling of the P5 pocket by Leu alters the DSE reaction, resulting in increased PapD/(PapA)2 formation but abrogating further polymerization events.
The crystal structure of the PapD/(PapANtd1_G15N_T101L)2 suggests that the groove of PapA is longer than the groove of any other Pap subunits of known structure, and that this is why it can accommodate a longer Nte. This led us to suggest that both the DSE region and the region N-terminal to it are important for pilus formation. Site-directed and deletion mutagenesis confirm this view and thus validate earlier published observations by Mu et al. [22], which emphasized the role of the non-DSE region of the Nte. Thus, PapA uses an extended Nte and in that respect appears to be very similar to other major pilus subunits such as SafA of Salmonella or Caf1 of Yersinia [19,21]. The Nte of SafA, for example, is 17 residues long, and one residue outside its DSE region, F3, was shown to be important in capping the process of DSE and driving the reaction to completion (the DSE region of SafA consists of residues 11 to 17). Indeed, a mutation of F3 to Ala in SafA results in an equilibrium between reaction species because in this mutant, DSE is allowed to proceed in reverse. A possible equivalent of F3 in PapA is I4. Indeed, DSE is somewhat affected by the I4G mutation. Thus, some features common to the assembly of all major pilus subunits are emerging, which include the involvement of an extended protein–protein interface and a potential capping mechanism driving polymer formation to completion. The PapA polymer is however different from the SafA or Caf1 polymer in that it adopts a distinct tertiary superhelical structure. The structures presented here do not provide any clues as to how such a ternary structure could form. Indeed, packing interfaces observed in both the PapD/PapANtd1_G15N and PapD/(PapANtd1_G15N_T101L)2 crystals appear irrelevant (unpublished data). However, the structure of PapA elucidated here provides the basis for complementing the work by Mu et al. [24] and characterizing further the PapA/PapA interactions that preside over superhelix formation.
See Text S1.
The eight PapDHis/PapA constructs (PapDHis/PapAwt, PapDHis/PapANtd1, PapDHis/PapAΔ11–17, PapDHis/PapAG15N, PapDHis/PapAI4G, PapDHis/PapANtd1_G15N, PapDHis/PapAI4G_G15N, and PapDHis/PapANtd1_G15N_T101L) were transformed one at a time into E. coli C600 cells and grown in a 5-l fermentor vessel containing Terrific Broth (TB; Sigma, http://www.sigmaaldrich.com) kept at 37 °C and shaken at 600 rpm. The cells were induced with 1 mM IPTG, once the OD600 reached a value of 0.9, and kept growing for another 3.5 h. The complexes were purified after periplasmic extraction using Cobalt-affinity chromatography (Talon; Clontech, http://www.clontech.com), followed by hydrophobic interaction chromatography (phenyl source; GE Healthcare, http://www.gehealthcare.com), ending with a gel filtration step in 20 mM TrisHCl (pH 7.5) and 20 mM NaCl, using a Superdex75 120 ml column. This last step was crucial for separating the different polymer forms of PapA in complex with PapD (PapD/(PapA)1, PapD/(PapA)2, PapD/(PapA)3…).
The PapDHis/(PapANtd1_G15N_T101L)2 complex was purified as explained above. However after the gel filtration step, there was a second major peak that eluted at a volume of around 54 ml (the 1:1 complex eluted at around 60 ml) and was interpreted as a 1:2 complex (PapD/(PapA)2). The 1:2 complex was concentrated to 13 mg/ml for crystallization trials.
The PapD/PapANtd1_G15N complex was purified from periplasmic extracts using cation-exchange chromatography (SP HiTrap HP column; GE Healthcare) followed by hydrophobic interaction chromatography (phenyl source). The purification was completed by a gel filtration step in 20 mM MES (pH 6.0) and 20 mM NaCl on a Superdex75 120 ml column. The 1:1 complex eluted at a volume of around 60 ml and was concentrated to 8 mg/ml for crystallization trials.
See Text S1.
PapD/PapANtd1_G15N: two crystal forms of the complex were obtained; in both cases PapD/PapANtd1_G15N was crystallized at room temperature in a hanging drop. In the first crystal form (plates), the drop was equilibrated against a reservoir solution containing 25% PEG8K, 10% isopropanol, and 0.1 M MES (pH 6.6). In the second crystal form (rods), the drop was equilibrated against a reservoir solution containing 2 M ammonium sulfate and 0.1 M Na acetate (pH 5.6). The plates belonged to space group C2 and diffracted to 3.2 Å, whereas the rods belonged to space group P3221 with cell dimensions a = 167 Å, b = 167 Å, c = 178 Å, and diffracted to 2.6 Å The solvent content is 71%, with 4 PapD/PapANtd1_G15N complexes per asymmetric unit (Table S2). The data from one single rod-shaped crystal was processed to 2.6 Å, and the structure solved by molecular replacement using PapD/PapK as a search model with the program AMoRe [25]. The PapK molecule was first modeled to a poly-alanine chain prior to refinement. The first refinements were performed using simulated annealing and noncrystallographic symmetry restraints for the four complexes in the asymmetric unit, using CNS [26]. Then successive cycles of manual rebuilding with O [27] and conjugate gradient minimization with CNS were performed. B factors were refined individually. Toward the end of the refinement, the noncrystallographic symmetry restraints were only applied to some parts of the β-sheet core of the complexes. The refinement converged to the final R values of R = 23.0% and Rfree = 26.6% with good stereochemistry.
PapDHis/(PapANtd1_G15N_T101L)2: This complex was crystallized at room temperature in a hanging drop equilibrated against a reservoir solution containing 12% PEG8K, 5% isopropanol, and 0.1 M TrisHCl (pH 7.5). These plates belonged to space group C2, with one complex per asymmetric unit (62% solvent) and were improved by microseeding. The cell dimensions are a = 133 Å, b = 74 Å, c = 80 Å, and β= 109°. The data from one plate was processed to a resolution of 2.5 Å. The structure was solved by molecular replacement (AMoRe) using the previously solved PapD/PapANtd1_G15N structure as a search model. In the electron density map, there was extra density near the N-terminal extension of the PapA molecule. This density was good enough to enable manual fitting of another PapA molecule next to the first one. Then successive cycles of conjugate gradient minimization with CNS and manual rebuilding with O (http://xray.bmc.uu.se/alwyn) enabled rebuilding of some of the second PapA molecule loops that differed from the first one. B factors were refined individually, and no noncrystallographic symmetry restraint was applied between the two PapA molecules. The refinement converged to the final values of R = 22.3% and Rfree = 26.2% with good stereochemistry (Table S2).
PapD/PapA (chaperone/pilin) samples of wild-type and mutant PapA proteins were frozen and thawed five times in liquid nitrogen at 150 μg/ml in 10–20 mM TrisHCl (pH 7.6) and 20–50 mM NaCl. The frozen–thawed sample was used either on the same day, or allowed to sit at 4 °C for as long as 40 d to enhance oligomerization of PapA subunits. The sample (5 μl) was placed on carbon-coated, glow-discharged grids, washed with 10 mM TrisHCl (pH 7.6), negatively stained with 1% uranyl acetate, and imaged on a Philips CM12 electron microscope (no longer available).
The Protein Data Bank (http://www.rcsb.org/pdb) accession numbers for the coordinates for the structures of the complexes mentioned in this article are PapD/PapANtd1_G15N (2uy7), PapDHis/(PapANtd1_G15N_T101L)2 (2uy6), PapE/KNte (1N12), and PapD/PapK (1PDK). |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.